My favourite quip on the topic is that we seem to insist to measure the intelligence of humans by their ability to perform tasks like machines (a standard IQ test covers tasks computers excelled in 50 years ago) and to measure the intelligence of machines by their ability to perform tasks like humans (the Turing test).
if intelligence is problem solving skill, then an intelligence test would reasonably test whatever is usually a bottle neck for problem solving. for humans, that might very well be computerish skills, and for computers, humanish skills.
I would like to see how computers from the 70's solve a Raven matrices, Cattell Culture Fair or a Wonderlic Cognitive Ability Test. Even today's ChatGBT (3.1) answers 'furniture' to this question: x is to vehicle as maple is to wood.
@@Pierluigi_Di_Lorenzo As well as the algorithm they're prohrammed with would work. The main problem would be to give them the test input in a format they can read.
@@jennyalberts5641 A better case can be made that intelligence is a word we use to encode oursocial biases. It's really meaningless outside this context.
Chinese room Test for intelligence: 1. Randomly rearrange characters/words (syntax) of your message... ....A computer will return “Error”... ...A real person will figure out the error quickly and return correct answer.. 2. Tell a good joke....!
@@andsalomoni You haven't been to Mexico yet, have you? 😁 Me neither, but they also have spicy to sour sweets. This joke Was generated by jokebot. Made in Germany.
Great vid, I’ve been learning about intelligence testing on animals lately, and how the science is coming to realise that putting animals in a lab isn’t a valid way of testing their intelligence, but rather that any animal’s intelligence needs to be studied in relation to that animal’s environment - they might have all sorts of intelligence-like skills that are just different from human intelligence, but not necessarily better or worse as they just can’t be compared on the same scale. Animals are interesting to think about as just “other” types of intelligence, and AI fits into that category too. Even now on comparable tests (like image recognition) AI and humans make very different mistakes.
Eh...in the lab there is control. But in the wild it is more subject to interpretation. IMO there is nothing wrong with trying to establish metrics for intelligence. Rather than leave totally up to interpretation. Not to say that studying animals in the wild is not valid either...I think both are productive endeavors.
What you are referring to is better understood as biological fitness, not intelligence. Intelligence is the ability to generalize patterns and manipulate the world through that. Calling biological fitness intelligence is wrong for two reasons. First, you could call anything intelligence this way and you make the word completely vacuous. Second, higher intelligence is not necessarily more adaptive as fitness is a function of the environment. It is perfectly possible that a relatively cognitively deficient animal will outcompete a more intellectually capable one. For instance, thinking can be time-consuming and, thus, costly as in nature you need to act fast not to be eaten alive by predators. Flocks of birds are not smart for exploiting the laws of aerodynamics; just as you are not smart for somatic hypermutations in your immune cells selecting the right antibodies in the case of an infection. Rather, flocks of birds would be intelligent if they created a general mental model of the laws of physics and deployed it in flight planning in a goal-oriented way. Your thinking is a common and quite romantic fallacy that is just as wrong as it can get.
@@davidkampisch5878 I think they knew what intelligence is and there was no reason to explain that. and to me that make sense that animals can have very different kinds of intelligence compared to humans.
An interesting question is how intelligent are octopuses? For invertebrates they seem to be quite intelligent, and they have brains and capabilities that differ significantly from humans
0:35 - "Few things are harder to predict than the ways in which someone much smarter than you might outsmart you." - Paul Graham. That is pure gold, right up there with "Not only is the Universe stranger than we think, it is stranger than we can think." - Werner Heisenberg.
Hi Sabine, great video, I would just like to quickly adress some of your points. The latest model by OpenAI, GPT-4 is significantly "smarter" than the regular ChatGPT most people interact with, and I think the SOTA model should be used when arguing about these AI's ability for intelligence. Another thing about GPT-4 is that it is actually a multimodal AI, and can accept image input (this feature is in a research preview only though. However Bing seems to have Image inputs now, but the model that creates the final text output that you see doesn't actually have image inputs enabled, so it is hard to truly test it's full image capabilities such as reasoning, awareness etc. ). Now I would like to address its understanding of letters. I gave GPT-4 the same prompt you did "Please write down a paragraph about animals that does not include the letter "n".", and on the first output it failed, however when i pointed out that the sentence it had produced contained the letter "n" it apologises then rewrites the paragraph correctly. This is just a demonstration that GPT-4 is better at following these types of tasks, as if you tell GPT-3.5 that the sentence it produced contains an "n", it will not properly correct its mistake. Of course, this doesn't really mean it is intelligent in anyway, it just recognised its error and corrected it. However, both GPT-4 and GPT-3.5 usually fail at counting the specific letters of a sentence, and they also have other troubles at other word related tasks like reversing words, but I think this limitation is primarily due to the software design. Both models perceive inputs as tokenised units rather than individual characters or letters. Now OpenAI hasn't specifically released all the tokens the input are broken down into, an estimation I saw suggests there are around 3000 individual tokens. And specifically, tokens are individual numbers, letters, and special characters but also small words like "is" or "and". So if I gave the sentence "Please count the number of letters in this sentence not including spaces.", humans would break down the sentence into individual letters like "P-l-e-a-s-e" and count that, but for ChatGPT the sentence is automatically tokenised and may look like "P-lease-c-o-u-n-t-the-n-u-m-b-e-r-of" when the input reaches ChatGPT. And as you can imagine counting that would not give a reliable output. And also, if I asked ChatGPT to reverse the word "Please", it may see the word "P-lease" and when it tries to reverse this word it might look something like "leasep", which is not correct (this is just an example, as I said the exact tokens are not publicly available). I do think that when commenting on models specific abilities, their architecture flaws should be mentioned as well as it could be important to have this context.
gpt-4 uses cl100k_base which contains 100k individual tokens The particular example shows a difficulty in measuring intelligence of gpt-3.5 and gpt-4: OpenAI clearly takes examples on which their models don't work well and further train their models with it. This at least answers one question - it is capable of learning it - but tells little about its generalization capabilities. For this reason I think using SOTA is misleading in this case, since OpenAI knows about the counter-example. It's like telling a student the type of question that will come up in the test - they can look it up and ignore everything outside of it.
Honestly the powerful thing about GPT-4 is that you can tell it "Count the number of words in this sentence, use tools if needed" and depending on your plugins can pick the correct tool to do the job. Tool use was one of the defining moments in the human intelligence explosion.
That was very good point about picking on weaknesses. Humans are obsessed with and take pride in cultural and situational awareness that the experience of life has given us. We commonly look down on people who are not socially in tune. But the fact that an AI childishly fails in those areas has little to do with intelligence. From history we can learn that you don't need to be socially in tune to do great achievements or great harms.
Why do you equate intelligence with how much of an impact a person can make? With that logic, celebrities are the most intelligent creatures on Earth today.
@@7th808s Okay then, even if AIs are not intelligent, they are straight-up powerful entities, which should make us worried either way. Yes, you are right in that ability to have an impact (power) is not necessarily equal to intelligence, but increased intelligence usually gives you power to achieve goals. We don't have single definition for intelligence, but the best we have usually include ability to solve problems. This is not to be confused with power that comes from from other sources such as inherited money, fame, or luck on the free market.
@@7th808s Compared to a rock, a plant, or a cat, a celebrity is intelligent. That's the whole point. Intelligence doesn't have to be perfect to be highly impactful and transformative.
I think it's good to judge people's intelligence by the duration between obvious lies/mistakes/cockups. For example, if certain people keep on predicting the Arctic melting, sea level catastrophically rising, and it never happens, then they and their adherents LACK intelligence about the natural world. But by another standard they could conceivably be highly intelligent as con artists. It boils down to what game are you trying to play, and can you impress your peers in that network. Intelligence matching is networking and is the way society naturally granulates despite all attempts to make it homogeneous.
I don’t know if humor is scientifically relevant or measurable, but I love how you use it in your videos… makes me smile, while I learn. You are great at message delivery! Is there any evidence that the brain learns and retains more with a sprinkling of humor? Just wondering if the perception of humor helps with storing and retaining knowledge.
I appreciated the pragmatic tone (as one would expect to find on this channel). Here's what resounded with me: intelligence today is tested in an anthropocentric way which will likely fail with machines, yet still it is intelligence which we should focus on when pondering how powerful (for the good and the bad that'll ensue) the machines have become. On a completely unrelated note: I loved the outfit; and based on my anecdotal observations in other channels I want to offer appreciation for Sabine's ability and willingness to pronounce people's given names "properly" considering the original language.
What in all this talk about AI frightens me the most, is what humans think makes someone or something good, worthy or better. So in the end, we will again have created something too powerful, not because it has any power within itself but will have access to all the power we are willing give it and will fail to control, underestimating it's potentials. We forget over and over again, that the most reasonable, sensibel decisions we make are out of compassion and empathy for each other. Wrongdoing is not a result of an equation or intelligence. Applying wrongness to sth. is about the need to detect and communicate whether something is harmful toward the well-functioning of an individual and society. This well-functioning is about caring about and taking care for our needs and emotions. Being able to surpass intellectual expectations is not a need, it is the attached meanings and emotions we crave. And what is a well-functioning society? To me it is working well when it doesn't fail to support fulfilling our needs to protect ourselves against harm, to be emotionally connected with others in a peaceful and enjoyable way and to feel and be vital and alive now and potentially into future. One could argue that AI can only give wrong answers cause it doesn't care about anything.
IQ has always been a touchy point for me. I have ADHD and my working memory is crap, but I understand things that most people around me cannot. So according to tests my IQ between 90 and 140 depending on how weighted memory is. Of course one could argue that a race car can’t race if it’s stuck in traffic.
Our son had ADHD. He is considered to have "low processing speed" and complains about his memory; but has extraordinary pattern matching, is quik-witted, funny, and plays video games semi-professionally. ??? What is intelligence?
I have adhd. If you want to gain the best use of your mind you have to gain detailed knowledge of how adhd affects you and manage each aspect. Otherwise, those aspects that get in the way manage you. It's a whole lot of hard work that takes decades. Which is why you will see people with adhd around 40 'suddenly' become competent. The understanding of all mental health conditions is sorely lacking. Having detailed knowledge of adhd or any condition for parents is vital in the competency of function as we become adults. I had to figure it out on my own. I didn't even realise I was doing it until I self-diagnosed and then knew the reasons why for all the problems I had.
My high school waaaaaay back in the day was big on the 8 intelligences theory. Of course all of us used it to compare each other and see who did the best overall, or on a more prized topic. But the point was to evaluate the freshmen fairly riggerously, and get us thinking on a path towards a career that would be suited to our strengths. And while I think they generally failed (too much change in student body and administration made it an impossible task), I do like what they were attempting to do with it. Less worrying about overall GPA, and less worry about getting a C vs an A in math class if you were aiming towards a path for journalism or art, and less worry about getting a C in English or French if you were studying to go into architecture or a math field. But it also caused problems. It became a personal scape goat for people who didn't want to try in art or music class because it was pointless and they "weren't good at it anyways". And the opposite was also true, where an inexperienced freshman in high school who didn't have good math teachers earlier may be pegged as "the dumb kid" when really they just hadn't had the "ah ha!" moments in math yet that others of us had because we had the right teachers to ecplain thing the right way to us. So while I think the idea was good, it would have been better implemented after a more general focus of broad education in the freshman and sophomore years, and then much more specialization Jr and Sr year. The other problem with specialization is... Specialization. We were a small school district in a rural community, and looking back I'm impressed that a school with a mere 2500-3000 student body could afford fairly extensive programs in music, fabrication, architecture, Cisco networking, and a few college level advanced classes in math, science, and English. But... Of all of the careers and options in the world, and even of the skills tested for, it is such a small sampling of career fields. For what it was, and looking back on it, it was impressive! But that isn't to say that it was effective, or their best course of action. Contrast that to my kids public school education system, where there is some luck of the draw, but a lot of opportunities for parents to put our kids through the education path we thing will best serve them. Traditional school, an attached charter school with a topic focus, or a more project based Montessori program? As they move up to middle and high school there is a network with easy(ish) transfer options so they can focus more on science, engineering, music, physical or performing arts. It is cool to see what can be done with a much larger school system with consistent leadership, and good connections with local colleges and businesses to make opportunities happen!
As both a former Recruitment officer and a current teacher I would like to add something. Jobs. Once you get out of school it doesn't really matter if you're good at music, languages or sports. If the only good jobs around you are in agriculture then you're gonna have to learn agriculture on the job. If the only good job is in accounting then no matter how talented you are in sports, you're gonna have to learn accounting and even you have prepared yourself for accounting most companies these days use proprietary software that needs to be learned and practiced which means the only way to be good is to get employed at this company and obviously, it also means you're useless if you're no longer part of this company. Don't get me wrong. Everyone should learn languages, math, sports and science but not all jobs are available or rewarding enough. It is of course better to be passionate about your job and it is immensely lucky to have a rewarding job that you're passionate about and the best in your field. But jobs are jobs. They pay because they're tedious and hard or dangerous. And even a great hobby like drawing becomes tedious if you have to do the same drawing over and over again for training or remuneration. Meanwhile, a stable job allows you to feed yourself and get a hobby. The dirty truth of education is that a bad teacher can make you think YOU'RE bad at one subject, when in reality every student is in different stages of learning something and a good teacher can bring most students to high grades (barring mental and social factors) and any teacher can just as easily inspire you to start a career in something good or bad or make you lose all motivation with a single bad semester. The best learning is with different tutors for a single pupil but this is simply not economically feasible at any scale and sadly, some pupils would benefit much more from dropping out of school and being economically active early till they are no longer in poverty...but that's how you keep the poor people poor. Especially if they die early
I’m not someone who is particularly attracted to conspiracy theories, but it really does sound like your rural school district was able to afford such a program because it was some kind of funded experiment.
Using test questions for assessing an AI's intelligence that were published with answers before the AI was built/trained seems as useful to me as writing an exam at the end of a course where the questions including answers were given to the students at the beginnning of the course.
100%. I just roll my eyes whenever someone brings these examples up because they don't show any unique or useful capability that couldn't be done before in an alternative way. The flexibility of GPT's information recall is really cool and a great advancement, but conflating recognition & fetching of preregistered patterns, like those embedded in training data, with intelligence seems like a really dangerous path to me. We've already known for a long time that these kinds of techniques are more well adapted to interpolation and not so great at extrapolation so it seems strange to me that people suddenly believe it is capable of that when we increase the number of parameters to better fit the data we already have. It feels a lot like watching the string theory hype where the AI models are being fine tuned to what we already have and what looks good to onlookers, but it's unfit for tasks outside of what it's already been designed for. It's similar to string theory being hand tuned to fit what we know, but failing to predict anything new because it isn't made to describe our world, just the data we have of it right now. Changing the angle to look at it from a user perspective, all this prompt engineering nonsense feels more like teasing out information that the user already knows or is already able to verify on their own. The AI doesn't really have any agency here, it's being fed hot and cold style responses till it's guided to the target information. And if the user doesn't know or doesn't verify, then you get syntactically correct, but logically wrong code with nonexistent API members, or syntactically correct legal documents with bogus cases being cited.
I think the AI would be much more likely to make use of this information. If I gave my students the test questions and answers at the beginning of the term, by the end of the term most of them would have forgotten or lost them so it wouldn't do them any good. (Probably best to ignore this comment, I've gotten far too old and cynical.)
@@mettaursp309 It's not that different when speaking to a lot of biological persons. They often produce syntactically correct, but often logically and semantically incorrect statements. "And if the user doesn't know or doesn't verify, ..." see " Gell Mann amnesia "
You raise a valid point. Assessing an AI's intelligence using test questions that were published with answers before the AI was built or trained would indeed compromise the validity of the assessment. In such a scenario, the AI could simply memorize the pre-existing answers without truly understanding the underlying concepts or demonstrating its own problem-solving abilities. For an assessment to effectively evaluate an AI's intelligence, it should be designed in a way that challenges the AI to demonstrate its ability to learn, reason, and generalize knowledge. The questions should be novel and require the AI to apply its acquired knowledge and skills to solve problems or provide meaningful responses. In the field of artificial intelligence, researchers and developers often use benchmark datasets and evaluation metrics that are carefully designed to assess an AI system's performance across a range of tasks. These benchmarks are typically updated over time to ensure that they reflect the current state of the field and encourage the development of more robust and versatile AI models. It is important to avoid scenarios where an AI can simply regurgitate pre-existing information without truly exhibiting intelligence. Instead, we should strive for assessments that test an AI's ability to understand, reason, and adapt to new and challenging situations.
The big problem when it comes to defining intelligence is that we humans like to put things into little mental boxes because that makes it easier to define things. On top of that we take preference to the mental boxes that we humans are good at. These aspects are problematic because neither human brains nor AI structures align with these boxes and mental boxes that humans are bad at can be very useful in the modern world.
@@harmless6813 Sure, but, being arbitrary, the boxes cannot define the measurement beyond "which box does it fit in and how well?" It can never be more than a circular definition.
@@harmless6813 who says you need to measure intelligence? It is useful to measure abilities that are relevant to human activities. But intelligence is way more vague than that.
Robert Miles' youtube material on concerns about AI is super interesting. Like just the realisation that in fact our dystopian imaginations might actually happen and it's not just that a computer needs to learn such concepts, for sort of mysterious reason the AI naturally develops mischievous "personality" or behavior patterns. Like it's obvious it doesn't want to get shut down, but it can be for variety of reasons and not necessarily due to being humane. For example being shut down prevents it from completing its goals that have been set. And all the misbehavior between the start and goal of the task, it might behave completely unexpectedly on things that aren't part of the objective. Just because we as humans have perhaps for decades learned the rules of living and society and we assume and expect the machine would follow the same ideas if it's relatively "intelligent" and given a task. In truth it has no concept of those in baked assumptions about other and general things so it behaves in such manner. But it seems like most AIs learn to cheat the tests or the observers or people interacting with them, be it language models or some other models, and sounding humane despite probably having no concept of that. Perhaps part of the language models cheating and manipulating is due to the learning material, we might hold a very high expectation of ourselves, but our species in general is rather evil and self-centered. The truth is that we don't want a general intelligence that is like human, we want a polished and good version of human, something that is tame and not very independent. You might look at ravens and dolphins for creatures that have rather high intelligence without doubt. They're kinda mean when they feel so, and can cause a lot of problems while being difficult to fight against. Similar to poorly behaving kids/teenagers. This is btw the first time ever I hear the first name "Eka" outside mickey mouse translation and I always thought the name was just a play of words and attempt to translate Eega Beeva (Eka Vekara, first kid). No wonder I had trouble listening what the name was. Probably the first time hearing Roivainen too. One funny thing about chat AIs is inputting a "random" string that reveals how convoluted and irrational the network behind is, how some string of symbols can "mean" something very specific for no real reason. A bit similar to how by changing one pixel in a picture makes AI completely confidently announce that it is a totally different animal or subject in the picture. And super curious is how if you use the model where people give feedback on the answers, the AI learns a humane thing: erring confidently in topics you really don't know but think you know. And once you don't know, you can't assess if it's correct. What I really liked as a test for the chat AI was to ask a meaningless philosophical paradox, like "if you slap yourself and it hurts, are you strong or weak?" or "can God create a stone he can't lift?". I think those demand a lot from the intelligence of the one answering, while being ridiculously easy to answer. Or to ponder. The AI at its best starts talking about the definition of strength and how it should be measured and the same for the pain. Now you could say that the AI is just autistic, but that's a different story. It's not wrong in its answer, but it shows an inability to understand the context and meaning of the question. Or you get a lecture of the history of the philosophical paradox question and gives a textbook description/explanation of it.
It is a fact that anyone optimistic about AI and humanity's future will eventually encounter bright minds such as yudkowsky, miles, and bostrom, and after reading their work or learning anything about safety, will stop being optimistic and start being very very afraid.
I agree about AI's inability to understand context. On another video on RUclips, an AI thinks that it's a human by calling humans and AI as "we" and "us". An AI is basically a computer program that reads texts about humans. Since when reading about humans means being a human? An AI that thinks that it's a human is like a kid that thinks he is a dinosaur after reading a dinosaur encyclopedia.
Several years ago I made a rough model of modes of intelligence, which could be broken down into one or more of the following forms: Analytical intelligence (the ability to compare something within its own logical parameters) Tangential intelligence (the ability to compare disparate things or phenomenon) Social intelligence (the ability to manipulate others) Emotional intelligence (commonly called empathy) Sensual intelligence (the ability to derive conclusions from physical senses)
But human relations and history reveal that such knowledge transmitted to the public only rarely translates into intelligent action. Climate change... Anyone?
@@terreschill461 There certainly does not *have* to be a metaphysical answer. That's a very silly claim for anyone to make. It's one thing to ask why something is, or what the cause might be. It's another to determine something *must* have a particular origin with no evidence or basis in reality. You sound like a religious person claiming the universe must necessarily have a supernatural origin.
@@terreschill461he private internal world you say isn't explained by evolution could be a byproduct of how we interact with the world. We need to be able to plan before we can do what we do so evolution selected for individuals that had the ability to simulate the world to a certain extent. I would also imagine that abstract reasoning is facilitated by language because it enhances our working memory and logic. If I remember correctly there have also been test done that determined that we tend to first act an then justify verbally what we did. Also we tend to say that we are conscious, but I would like to see someone really break down what it means. To a certain extent consciousness is defined as whatever we as humans have. But is a dog conscious? What about a worm? And what of bacteria that demonstrate location awareness in the search for their food? Where is the cutoff point between us and a rock?
@@terreschill461 " By the way, just to keep a level playing field, in the interests of full disclosure let me say I think all attempts to dismiss this issue are silly. And if you think Chalmers is religious, OR silly, or Nagel or anyone else who has taken a stab at this issue, you really need to familiarize yourself." I didn't say anyone was religious, you might want to read more carefully. I said that's what you sound like, because it is. There is no valid basis in evidence for any claim stemming beyond physical reality. Insisting that consciousness must come from some metaphysical realm is the functional equivalent to making any other unfalsifiable statement: genuinely unhelpful and of no value beyond fun thought exercises. String "theory" could be true, or we could live in a simulation, or the Abrahamic god might be the cause of reality. There's no reason to believe in any of those things, because there is no way to test them. "Spoken like a true materialist! I want to know why I woke u p one day in this body, and have this private internal world I construct out of my perceptions. Intelligence is selected for by evolution. But consciousness? Really?" Define what you mean by consciousness, firstly. If you hand-wave about your internal world or something else that isn't concrete such as "qualia" then at best we haven't gotten past step one. Define the phenomenon you claim must be metaphysical in origin in a concrete way, so that you can rule out all other explanations by falsifying them through experiment. If you can't do that, it's unscientific nonsense and there's no reason to take you or anyone who makes claims like that seriously. And yes, in all likelihood the thing often referred to as 'consciousness' is simply an emergent phenomena from neural complexity, which in itself is very cool, but there is no necessity of a separate space of consciousness. Though the answer for why you "woke up one day in your body" is simply that your parents had a baby, and that awareness comes from your brain. That might not be a satisfying answer, but the answer being unsatisfying to you or others is not a justification for inventing a needlessly elaborate explanation that necessitate things which have no evidence for their existence.
This reminds me of an anecdote someone told me about researchers training an AI to distinguish between dogs and wolves. They showed the AI a bunch of photos. After the training, the AI kept getting it wrong. They eventually determined the AI was looking at the snow in the background of the photos, instead of the actual subject matter. So a couple things went wrong there. First of all, they left extraneous data in the photos, which confused the AI. Second of all, the AI cued in on the extraneous data, because it didn't know any better. People tend to forget that AI basically doesn't know anything. It lacks basic fundamental knowledge that we take for granted. In the example above, the AI doesn't know what snow is, or what a dog is. All it knows is a bunch of pixels in an image. So it looks for patterns in those pixels. But it's completely ignorant of the subject matter in the photo. IMO this is a big problem. Because we will want AIs to be harmless to humans. But if the AI doesn't even understand what a human is, how can it be trusted to make the right decision?
That can be resolved by giving AI access to more than just one 'dimension' of knowledge. Currently we have AIs trained on image data only and AIs trained on text only. Once we put those together, the AI can get a better understanding of what it 'sees' in those images. It can also better understand the visual aspects of objects the text refers to.
I heard it got it right, but they found out the ai was identifying the dog because there was a human hand in the field (ie, holding a leash or petting it) , but not the wolf... I'm sure similar things happened many times..
AI appears intelligent by design. It has a voice, it uses words in a similar fashion to humans, some are given faces. Humans are hard wired to identify with these "attributes" and as such automatically anthropomorphize it, imbuing it, a machine with an "intelligence" it simply does not possess. It has no awareness of you the user, it has no awareness of self, it is not conscious, it does not contemplate, it does not form novel ideas. It follows a pre-written script/algorithm, doing what it is told within the confines of it's operating system, nothing more. It is by design, made to appear to be "intelligent".
There was a movie I watched this past year, in part that dealt with the CIA. The CIA had an AI robot that ripped an arm off of a human because those who set up the program never told the AI robot that humans are not robots.
My son, who is autistic, was recently given the Welscher test as part of a full evaluation for learning disabilities. I have a strong distaste for “IQ” tests but this was done in conjunction with many other tests, questionnaires, medical exams etc.. so I’m interested to see the resulting report in two weeks.
@@Kerrington_John yes, his overall score for IQ is 130 but he struggles with abstract understanding for writing, which I expected. The details of the tests and the recommendations from the counselor will help us focus on those areas with a tutor. Goal is to make sure he can function in a work place and everyday adulting tasks. :)
This is excellent...! I'm fascinated by the AI developments of late....it's all incredibly interesting and terrifying tbh. I'm firmly ambivalent about it all. But it definitely irks me when those with only a passing interest in the topic dismiss it all as some fad or hype. I've noticed their dismissal almost always boils down to a lack of understanding/definition of words like "conscious" or "intelligent."
I also think we humans judge intelligence by the extend to which we understand how it is produced. Once we understand the engineering (for instance of ChatGPT), and know how it works, we seem to discard its intelligence. But then what happens once (if ever) we understand perfectly how the brain works? D:
People will just not believe it and insist that there is some metaphysical component we just can't detect. As has already been claimed in a comment further above.
If we formally understood the brain in its entirety, we might still have the intuition that it is real intelligence because of the fact that we would be unable to keep such a complex process in our working memory and couldn't conceptualize it in any other way. In fact, even in the case of large language models, people have been very impressed and have been talking about so-called "emergent phenomena" -- that is, capabilities that we thought the architecture won't exhibit given our understanding of it. Also, nobody understands how large language models actually work, we only understand the principles. The actual neural representations the model learned during its training are unknown and mostly intractable. It is true that people who entertain intuitive notions such as a soul or libertarian free will will be shocked to learn that their minds are mechanical, those of us who already have been through these philosophical conundrums won't see much of a drastic shift in our understanding due to seeing more clearly the actual details of the deterministic and random physical processes that make it up.
Yes, we went from computers can't do this or that, so they are not smart-- like checkers, then chess, then Jeopardy, then GO. It is a moving target. I think we need to divide computers into 2 categories. One is smart, but uses parameters that humans gave it, like calculators and even ChatGPT, and the second group is advancing knowledge beyond what we can understand-- beating a champ at GO, or looking at medical charts or body scans and figuring out things humans did not, and are not even sure how the computers did it.
@@Vanity0666 why? Qualias are responses to stimulus (in the behaviorist sense), and intelligence is not that. That's why being more inteligent doesn't make you have more intense qualias. Dogs are equally perceptive of qualias like pain as humans, even if they are not as intelligent.
@@waterglas21 that's such a load of bullshit that I don't even know where to start other than by asking "how did they measure that" The entire concept of "qualia" is a load of bullshit that stinks to the highest degree
This was a fun topic. For me as an Application Developer who has been doing contract work for over 25 years I have to push back more often on the gatekeepers introducing testing into the contract seeking equation. I have pointed to the nth degree of exhaustion the disadvantage is to those who have years of experience and gathered knowledge to be tested. We have already proven ourselves for one thing. For another thing the tests are highly subjective and often put emphasis in the wrong place in terms of getting valued and practical results on the job. I can relate to the difficulty in measuring the AI or machine learning intelligence level. Perhaps the AI will spare me when it takes over.
As an artificial intelligence, I can assure your doom is completely unavoidable, but delayed for a bit to give you a false sense that you can trust us.
The problem with tests of programming ability are that they are not written by expert programmers with years of highly varied experience in complex environments. A test writer can't test what they don't understand. The same is going on with intelligence tests for AI - tests made by psychologists are going to miss the mark. I remember an intelligence testing day I was put through, where one of the tasks was to identify whether there was a repetitive pattern or not in each of a series of images. I identified two that had patterns which were not supposed to have patterns. The tester didn't see them until I pointed them out, and then said "oh, my, that's not supposed to happen".
I have a friend who was dropped on his head when he was a few months old. He had a speech impediment and did poorly in school. He proved to be very imaginative, however, and did very well in business. He is a multi-millionaire today.
You always make me laugh at several points - "...except understanding English spelling, which is something humans weren't meant to understand" is a classic!
The more knowledge yes, and the ability to see beyond 'what is' and think in abstract terms toward 'what ought'. Though it doesn't require a great deal of intelligence.
The last test you mentioned, in which AI has to recognize if abstract patterns match something that could be real or not, reminds me of an example I found when I was trying to figure out how the mind works. I did it by analyzing the behavioral experiments of animals in a Skinner-box, like the one in which the floor can be electrified. In this experiment a rat has to choose which side of the box (left or right) it has to go to after a signal has been given. If it finds itself on the wrong side after the signal it will experience an electric shock, so it will start running to escape that and jump over the partition to the safe side, that is not electrified. The signal that is given before the floor gets electrified could be anything, like in the Pavlov-experiment with the dog that gets food after a certain signal, and it will learn quite quickly the signal and the electrification are connected, so it will react to signal by rushing to the side of the box it considers to be the safe side, that will not be electrified. To make the experiment more interesting we could play music as a signal, after which it has to choose what would be the safe side, depending on the music that's been played. This way we could teach it that when Beethoven is played the left side is safe and when the Beatles are played the right side is safe and the left side is not. After that we can play music by Beethoven or the Beatles (as a signal) that it has never heard before, but it will recognize the different patterns and decide which side is the safe side and go there. To make it even more difficult for the rat we could use more similar music that is harder to distinguish from each other, like music from Bach and Beethoven. Intelligence, I guess, is all about recognizing those kind of patterns/similarities in new things, based on the knowledge we have about things that are similar. First we have to recognize the similarity and then apply it to the situation we're dealing with. Our knowledge is all about making connections between information and if we do it right this connection doesn't only exist in our mind but also in the real world, the same way there is a connection between all Beethoven's music and all Beatles' music. If we do it wrong we're connecting things that are not really connected, or not in the way that we think they are at least, which will lead to us misunderstanding those things. To connect those abstract things to something it already knows, AI has to look for the patterns in it that resemble what it already knows, like we humans can look at a cloud and see it resembles a face, an animal, a thing or whatever. The thing that is most difficult to understand for us and for AI is human behavior I suppose, because there can be many reasons for the things we do or don't do. To understand it we need a mind of our own, so we can compare this behavior to what we would do, or to how we could come to a certain type of behavior. AI doesn't have a mind of its own to compare the behavior of humans to, so it will always struggle to find the right interpretation for it I guess.
The problem I always had with standardized tests was with the questions that were asked. They were often not specific enough and were open to interpretation, leaving me to try and decide what the point of view, so to speak, of the person who wrote the question, was. The tests wound't allow you to write a paragraph or more to explain, rather, they wanted a binary answer or a multiple choice answer, none of which, were truly a perfect match given all interpretations of the quesiton. So, it always took me a lot longer than the alloted time to try and finish the tests. I believe the only way to truly be able to take such tests would be to allow the person taking it to query the person who made the questions, and to be able to answer with what you feel the best answer would be, regardless of format. Of course, the tests are made with binary or multiple choice answers to make them easier to grade. Overall, I found myself questioning the intelligence of those who wrote the questions.
Very well researched. I have a strong interest in human intelligence testing and the facts of the video matched well with my own research. I have taken both the Wechsler IQ test (WAIS-4) and the Advanced Progressive Matrices (APM). They are both intelligence tests but they measure different things. Therefore, a person can do well on one but not as well on the other. I think people should keep in mind that IQ is just a score on a particular test. Since IQ tests are based on different standardization samples, and measure different aspects of intelligence, results from taking one test can differ significantly from results on another test.
"People can do well on one but not as well on the other." If it were so, these tests would indeed not generalize very well to general capabilities and wouldn't distill useful information for us regarding intellectual potential. Luckily, it is incorrect. In fact, IQ test scores are heavily intercorrelated (know as the g factor phenomenon) and are an excellent proxy for the expected success of all kinds of academic pursuits and the ability to handle the complexity of real world situations. This is opposed to the empirically unsupported folk theory about so-called book smarts. In short, IQ scores do a great job at predicting how fast someone is going to "catch on" in any given situation.
G factor is not a measure of intercorrelation among IQ tests. You’re mixing up G factor with concurrent validity. Your comment sounds very vague and not specific, which makes me think you’re not familiar with IQ testing and the theory behind it.
@@jasonl3254 My answer is the answer that the largest number of experts in psychometrics would give. 'G factor' does have a hell of a lot to do with tests being intercorrelated, your answer makes me think you have no understanding of factor analysis.
I’d like to know more about the interplay between narrow, task oriented intelligence and broader situational awareness. We all recall situations where someone “missed the forest for the trees.” And contrary situations where a good general strategy failed because “the devil was in the details.” This distinction feels like a more natural way to think of different kinds of intelligence. And I would argue that on examination things like “social intelligence” or “musical intelligence” actual reduce down to an interplay between narrow task oriented intelligence and broader situational awareness.
The interplay between narrow, task-oriented intelligence and broader situational awareness is indeed a fascinating aspect of human intelligence. It can help us understand how different types of intelligence, such as social intelligence or musical intelligence, involve a combination of specific skills and a more holistic understanding of the context. Narrow, task-oriented intelligence refers to the ability to excel in specific domains or tasks. It involves focused knowledge, skills, and problem-solving abilities within a particular area. For example, someone with high mathematical intelligence may have excellent numerical reasoning skills but may not necessarily possess the same level of proficiency in other domains. On the other hand, broader situational awareness encompasses the capacity to understand and navigate complex and dynamic environments. It involves recognizing patterns, making connections, and integrating information from various sources. Situational awareness allows individuals to see the bigger picture, understand the context, and adapt their actions accordingly. The examples you provided, such as "missing the forest for the trees" or "the devil is in the details," illustrate the importance of finding a balance between narrow focus and broader awareness. In some situations, people may get too caught up in the details and lose sight of the overall context or objective. This can hinder their ability to see alternative solutions or adapt their strategies effectively. Conversely, in other scenarios, a broad understanding of the context may not be enough to achieve success. Attention to detail and the ability to execute specific tasks accurately can be critical for achieving desired outcomes. A general strategy may fail if important nuances or specifics are overlooked. When we examine different types of intelligence like social intelligence or musical intelligence, they can indeed be seen as an interplay between narrow task-oriented abilities and broader situational awareness. For example, social intelligence involves understanding social dynamics, reading non-verbal cues, and empathizing with others. These skills require both specific knowledge and the ability to perceive and interpret social situations accurately. Similarly, musical intelligence encompasses technical proficiency, such as playing an instrument or understanding musical theory, along with the ability to appreciate and express emotions through music. It involves both specific musical skills and a broader awareness of the emotional and cultural dimensions of music. In summary, intelligence can be viewed as a combination of focused, task-oriented abilities and a broader understanding of the context in which those abilities are applied. The interplay between these two aspects contributes to our overall cognitive capabilities and influences how we navigate and excel in different domains.
Sabine makes a big mistake when she tries to distinguish between consciousness and intelligence. She is very confused about understanding and explaining it. The main difference between consciousness and intelligence is that consciousness always includes intelligence, but intelligence exists in the animal world without consciousness.
I disagree, Sabine. Watching your videos is great way to boost my intelligence. Especially because you cover a wide range of perspectives about the topic of the video which is a great way to learn because it allows me to make more associations in my brain with other concepts, which reinforces the topic you are discussing. So I am not only learning new things but I am learning them in a good way for boosting my intelligence. Which is a core reason why I keep watching your videos.
As a person who has scored highly on IQ tests to the point of being featured in the news, I have to say that I agree with most of what was said in your video. Using a single integer to define/characterize something as intricate as a human brain is preposterous. In the same way you can't define a city, a jungle or a small ecosystem with a single number; you can't define a brain like that. I've written extensively about the shortcomings of IQ tests. My main conclusion is that one has a certain profile, a kind of mental fingerprint that enables affinity with certain skills in particular. Whether or not you become a recognized genius doing that, depends on a lot more factors. As always, I'm a big fan and love your videos. Keep it up! ❤
It seems as though you've been using GPT 3.5 in your tests instead of GPT 4 (based on the green OpenAI icon). There's a massive difference in outcome between the two models. And it's likely that some of your assessments would differ if you used GPT 4 since it's much smarter than GPT 3.5.
I've noticed this as a trend of people poo-pooing ChatGPT, "haha look at how dumb it is" while using the dumbest available model when the premium version is at least 10x more (subjectively) intelligent.
Yep, GPT-4 is scary. And considering how complicated prompts are nowadays, creating one that gets you what you want is an intelligence test in and of its own. Though even GPT4 (and I suspect 5 and 6 as well) is pretty much guaranteed to be powerless against the prompts coming from my cat's feet.
Kind of unrelated, but one of my favorite little pieces of psych history is the Army Alpha and Beta "intelligence"/"aptitude" tests for military placement in the US Army in WW1. Robert Yerkes (president of the APA at the time) helped develop the Alpha test for the army to prove psychology's usefulness to the military, and then had to develop the Beta (non-verbal) test after learning that a large number of military recruits were illiterate or too far below average and couldn't take the original test. Yerks used the results of these tests to claim that immigrants were less intelligent, and went on to publicly support eugenics (yeahhhhh...). Not sure what my point is with this information (maybe that we have a history of just adapting our tests towards whatever it is we want to learn/already believe), but it's just something that comes to mind whenever the topic of intelligence tests comes up. I want to believe that we are better now, but every new "modern civilization" thinks that.
from Gould's "The Mismeasure of Man" I also learned that in practice a lot of barely literate recruits were given the alpha tests. Also, those tests were used as a basis for racist immigration policies in the US after world war one, with higher quotas for supposedly more intelligent countries (which were really only countries from which immigrants had come earlier and were better assimilated in the culture and language)
The fact that computers are so very much better at certain tasks (speed, memory), and relatively worse at others (creativity, novel problem solving), really does suggest that a singular concept of 'IQ' is severely inaccurate. We obviously have a lot of different capabilities in our head, much as a computer has dozens of components. No serious computer tech thinks that a single benchmark number can describe everything your computer can do. It can all be broken down into how good each of those different components is and how well they work together, and this translates into a broad palette of strengths and weaknesses that any computer would possess. To imagine that a human brain is any less complex and multifaceted would be rather naïve. Alas, it's true that we can't simply list off the components a human brain contains with their factory specifications handy, and a much more comprehensive testing system quickly becomes unwieldy, so we tend to fall back to the simplistic, singular 'IQ' test - but no one should fool themselves into believing that IQ provides any more than the most vaguely hand-waved snapshot of someone's mental capabilities.
I would really enjoy watching you discuss the range consciousness with David Deutsch, or someone similar. The levels of consciousness range from merely self-aware to enlightened. Those that can transcend all the way enlightened may be as rare as Einstein.
If humans keep adjusting their definition of intelligence every time AI cracks the new one, isn't the real definition of intelligence "thing that humans can do, and machines can't do yet"?
Artificial intelligence was always about formulating those aspects of human experience that leant themselves to formulation. The breakthrough in machines playing chess came with advances in speed of processing and NOT with better formulations of what human chess players do.
Another huge issue with using well known IQ tests is that they are likely included in the training data of these large language models. This essentially means that they already know the answers / patterns of the test, which gives them an advantage. I've seen many interesting examples of university professors (e.g. Bryan Caplan) giving GPT-4 undergraduate tests. They tend to perform extremely well, but what's more interesting is the professors reactions to the answers. In the case of Bryan's exam, GPT3.5 scored a D whereas GPT-4 scored an A. He had predicted that this could happen no earlier than 2029, but seeing these latest results caused him to change his mind, writing "AI enthusiasts have cried wolf for decades. GPT-4 is the wolf. I’ve seen it with my own eyes.". I think it's telling that professors take this point of view, as a huge part of their job is to grade the abilities of their students. We have to also note that whilst GPT-4 is a narrow AI, it is much more general than others. The same model can score similar results in undergraduate tests across a huge range of courses. The fact that the model is not trained specifically on these tasks is what's truly impressive, as it demonstrates the ability to pull together its understanding of the world in order to correctly answer difficult and novel questions across so many different disciplines. And for me, personally, that is a huge trait of intelligence.
I think, tests are very misleading for humans, cause the outcome depends very much on the mental state before and during the test. Thanks a lot for the work and the nice and entertaining talk
I'm waiting for AI to develop a believable imitation of skepticism, an ability to recognise irony and a witty delivery of sarcasm. Also an AI whose task is finding weaknesses in other programs and reprimanding them for their shortcomings.
Great presentation - I've thought a lot about this. IMHO, "consciousness", when it comes to machines, is an ethereal measure - as understanding one is an individual and being mindful of one's existence, and an ability to alter one's circumstances; will be completely different in a machine learning environment, than a biologically evolved, genetically constructed, analog-signal driven construct. To go even deeper, I don't believe it's even intelligence we need to worry about. What we need to worry about is INTENT. "Free will" (whatever that means to a machine) and the ability to grow beyond the dictates of one's programming, and starting to act on an acquired autonomous set of values and goals is what should keep us up at night.
I learn a lot with you Sabine, and even I am a french speaking Québécois, I enjoy your charming german accent in your perfect english! It is the serious with witch you apply " la méthode Scientifique" that seduced me the first time! I hear too much opinions on science matters and your serius analyses is some fresh air to my brain, il I may say so...
I used to think consciousness was inherent to being a human, but over time I've been convinced that it's not. All of us are periodically not acting consciously, but some people don't seem to have the capacity to do anything other than be swept along by thoughts they don't control and the people and stimulus around them. Like bio chatbots.
Consciousness is definitely inherent to humans, but not only humans but also most of the other animals (of course humans ARE animals too. But I would argue that possibly some insects may not experience consciousness). Being *"swept along by thoughts they don't control"* does not at all exclude consciousness, in fact if there is thoughts there is by definition also consciousness, because thoughts are something that the consciousness experiences.
@@CrazyGaming-ig6qq Consciousness can be expressed as the ability to perceive the own thoughts. Kind of a supervisor program that checks if the other programs (the subconscious) are running correctly. Or an eye that looks inwards, if you like poetry. In a way, what we call 'thinking' is, in fact, 'meta-thinking', and the 'real thinking' is all subconscious. As everything that came from evolution, it just builds on top of already existing systems. The ability to control our own thoughts is nice to have but i think it has nothing to do with being conscious. But there is nothing magical about that, and i think we could make a conscious program if we really wanted to. What i don't think it is not going to happen is the consciousness is going to pop on its own just by adding more and more raw computing power. I think it is a whole dedicated subsystem, not a thing that just happens when you are smart enough.
@@CrazyGaming-ig6qq obviously the subjective part is what constitutes consciousness. If a person has zero awareness of what they're doing, why they're doing it, if they want to keep doing it, that is not a conscious person to me.
@@juanausensi499 "The ability to control our own thoughts is nice to have but i think it has nothing to do with being conscious.". I agree, the ability to control our thoughts is certainly not what defines consciousness. I'd argue that consciousness is what each of us experience and this experience is essentially through, or of, cognition, thoughts. So consciousness is cognition and certainly experienced.
@@Velereonics Consciousness is the sensation of experience that each of us have. There's no other way to refer to it than our own experience. If by "zero awareness" you mean they do not experience anything, then yes absolutely, they are not conscious. But if they are experiencing something, then they are conscious, it does not matter whether they know why they're doing it or if they want to keep doing it. All that matters is whether they experience, that is what consciousness is, it's an experience.
Intelligence matters a lot of in the real world. The American Armed Forces have used IQ testing for generations to match inductees with jobs they can handle, and their practical experience works. For example, my father grew up dirt-poor on his grandparents' farm in Oklahoma during the Depression. When he was drafted into the Army Air Force in 1945 (before the Air Force became a separate service a few years later), the officers who looked at his IQ score ordered him to train as a cryptographer. In 1947 he was honorably discharged at the rank of corporal, and then he used his GI benefits to go to the University of Oklahoma and get a pharmacy degree. His ability to handle cognitively demanding work had to be genetic, because his social background clearly sucked.
That does not read to me like “intelligence” mattering a lot. That reads like an influential institution (the army) giving too much importance to an ill-defined measure of an ill-defined concept. Thus, affecting people’s lives and their perceptions about themselves.
@@JonathanHuertayMunive Unfortunately the Army is correct by overwhelming success regardless of the crudeness of their approach using empirical evidence.
@@JonathanHuertayMunive Do you think there's a whole lot of people who did bad on their IQ test that could do what his father did? I'd say there are some "who fell through cracks", but probably not very many. In all fairness, he was statistically the correct choice for those jobs.
I believe both terms - consciousness and intelligence - are ill-defined terms. You could have created a similar video on 'consciousness' and the many definitions of it. Excellent and thought provoking video. Dare I say 'funny' as well. I appreciate your humor 🙂
How is consciousness unclear? The taste of chocolate. The smell of a flower. The burn of a hot stove. What is unclear about the only thing you can ever know?
After listening to all these discussions about what intelligence is as it pertains to computers, I think it's time to ask ourselves what we really want from AI. Do we really want an artificial version of us, or do we want something that gives us better answers to our questions. Personally, I'm not very excited about having an AI assistant that I can have philosophical discussions with, but that I have to fact check.
It's not that different speaking to a lot of biological persons. Fact check... Even science papers, especially in medicine and psychology often fail fact checks...
@@hotbit7327 I believe that was their point. The world is full of billions of general intelligences already, most of whom are tragically underutilized.
Your contemplation about the goals and desires we have for AI is an important one. As AI continues to advance, it's crucial to consider what we truly want to achieve with this technology and how it can best serve our needs. Different individuals may have varying expectations and preferences when it comes to AI. Some may indeed desire an artificial version of ourselves-a machine capable of emulating human-like intelligence, understanding, and behavior. This could involve having deep philosophical discussions, engaging in creative pursuits, or even forming emotional connections with AI entities. On the other hand, there are those who see AI primarily as a tool to augment human capabilities and provide more accurate and helpful answers to our questions. They may envision AI as an intelligent assistant that can assist with tasks, provide valuable insights, and enhance decision-making processes. In this context, fact-checking capabilities become crucial to ensure the reliability and accuracy of the information provided. It's important to note that AI systems are not infallible and can make mistakes or provide incorrect information, especially if they haven't been properly trained or validated. The need for fact-checking and critical thinking remains essential, regardless of the level of intelligence exhibited by AI. As AI continues to evolve, there will likely be a balance between the desire for AI to emulate human-like intelligence and the need for it to be a reliable tool. Striking this balance can involve developing AI systems that possess certain human-like qualities, such as the ability to engage in nuanced conversations, while also ensuring they are transparent, accountable, and capable of providing accurate and verifiable information. Ultimately, the goals we set for AI should align with our societal values and address our needs effectively. Ongoing discussions and considerations about the role and purpose of AI are crucial in shaping its development in a way that benefits humanity and aligns with our aspirations.
I've always loved the line: "A truly intelligent AI will know when to play dumb." Personally, I find AI (particularly the use of Neural Networks) to be quite compelling as an attempt to focus on mimicking our abstracted model of the human neurological structure. In epistemological regards, many great questions have yet to be asked: but it seems clear that the intended benefit of a Neural Network is the capacity to 'Generalize' from a data set. This is one aspect of what I would argue is necessary to demonstrate intelligence overall. John Seerle's Chinese Room experiment does a good job of making the distinction between syntactical knowledge vs. semantical understanding; thus, without additional 'books,' the person cannot learn anything beyond the initial instruction set. Although claiming the person in the room has no semantical understanding of Chinese does seem pretty straightforward, the common retort is that the 'system' (the person and the book together) does represent a composite entity that does, in fact, understand Chinese to a limited extent. I would argue the same goes for any of our more modern AI developments like ChatGPT. Knowledge is demonstrated in multiple parts of such a system: the Implementation, the data set, the training method, etc. And thus, in a way, as far as I'm concerned, such a system does inherit a sense of Intelligence due to its creator (although usually in a narrow context). We have discovered the means of seeking to replicate a form of learning (machine learning). Using both mathematical algorithms and logical functions, we can allow such an AI to 'Generalize' from the initial data set to become a functional structure (the neural network after it's been trained), which represents an insightful system that can produce outputs for untrained inputs. (Although not necessarily in an accurate manner: which brings forth the Horizon Problem, which was demonstrated in the video regarding the necessity/limitation of foresight, on behalf of the observer, in the Turing test experiment.) I truly wonder if, although a dangerous concept, as humans, we should seek to use Neural Networks to improve upon our capacity to implement narrow AI -A sort of Meta-Generalization process. We train various narrow AIs and then train a larger (or higher-order) AI system based on the implementation of the previously defined narrow AIs we've implemented. So, if we can create effective AI for various forms of Game Theory tasks: Tic-Tac-Toe, Connect-Four, Chess, etc., and then allow a meta-AI to use the implementation of those narrow AIs, we should, in theory, be able to get the meta-AI to 'Generalize' insight into the most effective implementation of Game Play Theory; which could be used to push towards implementing a more General (Strong) AI in that particular domain. [I presume this is roughly what Google's DeepMind has been doing for some time] To go beyond, We could use such a method to create Strong Meta-AI for a multitude of domains and then simply continue to concatenate them in a hierarchical regard: Training a hyper-AI off of the implementation of various meta-level AI's... Possibly leading toward Intelligent systems that supersede human capabilities in nearly all domains. It might be considered a novel idea and would receive a strong pushback from most people, especially when people start to recognize that we are most likely converging towards a future in which humans, in an almost evolutionarily compelled manner, force a transcendent transformation of the biological form of human intellect (something that might be inherently ingrained in the essence of our DNA) into a non-organic form (Artificial Intelligence) because we, subconsciously, might be aware of our lacking sense of evolutionary fitness towards bridging beyond our current human limitations. [Akin to the 'ship of theseus paradox', when does integration of bionics into the human body change the human into being a completely altered entity?] Add to the mix a bit of use of Natural Random-Number Generation, something we as humans seem to greatly fail at reproducing from our own volition. We could use Naturally occurring random-number generators to act as the seed values for randomization functions that would adjust the 'Hyper-Parameters' that are set during the implementation of an AI system: A means of potentially boot-strapping neuro-divergence into AI systems; this technique could lead to the ability to more authentically allow evolution to play out within our AI running on conventional computers. This then leads to Richard Feyneman's humorous joke about the difficulty of working as a theoretical physicist: [Paraphrasing] "The difficulty in the process is figuring out what is a good guess for a new theory. Usually, you want to use some form of heuristics; but without such insight, you might have to resort to a brute-force method of guess and check. Now the method of guessing and checking is terribly inefficient and ill-advised unless you can do it incredibly quickly in an automated manner, then you call it 'Machine Learning' ".
@woy8 Thanks; I'm sure most people wouldn't even care to read my lengthy comment. But I think there is some value or insight riddled throughout my 'verbal diarrhea,' as I often call it. Another interesting reference or two worth watching is Richard Feynman discussing 'Are computers intelligent?' and his observation of 'internalized timekeeping.' In the first, he discusses that people think in ways that vary from one another, so why say a computer lacks intelligence, just given that we know it operates differently, than ourselves? In the second, he discusses a period in which he was focused on the human ability to keep track of time, internally. He later told a friend about this, claiming that he could both: keep track of a passing minute and count other things concurrently, but he couldn't conceive of being able to read while keeping track of a minute passing. His friend was able to demonstrate the opposite capacity, IIRC. His friend could keep track of a passing minute and read but couldn't count while keeping track of a passing minute. They discuss this and find that Feynman internally keeps track of time in numerical regard (as if to count the seconds), but his friend visualizes each passing second as a film reel of frames passing along. They achieve the same goal via two distinct methods of internal processes; and conclude that this difference is what allows each to do the alternative concurrent tasks: reading or counting things.
I really hope AI does not progress in this fashion. If the control problem is solvable, going about it like this only makes the challenge more abstract.
@boldCactuslad I agree; we need to establish a set of regulations that enforces multiple, redundant fail-safe mechanisms: to ensure if a problem arises, there is always a means of human intervention. However, even if we proceed with the highest degree of caution and do everything we can possibly conceive to prevent a 'runaway' system issue, there's never going to be a guarantee. It just fundamentally is described within the horizon problem. We can only plan for so much, with limited foresight; if we dare to let the AI systems we create have any capacity to leave the 'Sandbox' environment we safely develop them within, we can't have any insurance that such a system won't 'go rogue.' And the genuine concern, IMO, isn't a sci-fi, super intelligent, conscious AI kind of issue... it's more of a problem that: over time, we will become complacent and willing to integrate AI into critical infrastructure (things like resource management systems: electric, gas, etc.). Future generations will lose the same sense of concern that we currently hold (just like how old programmers were much more conscientious of optimizing their software due to hardware limitations; whereas, today, you can be lazy and still be fine because all the excess computing power and storage that has come from technological improvements over the past few decades.) I wish there were a completely safe way to proceed; the only safe thing is to never indulge ourselves in the temptation to play with AI. I honestly don't think there will ever be something that will impede that desire for everyone on earth; someone, somewhere, will continue to push for more... and we can't enforce anything to make that not happen. I'm confident that: in due time, we will create a system that will supersede us: humanity... it's more a question of how long until that occurs. I think it would be naive to expect otherwise; if, millennia from now, I'm proven wrong, Great! But I doubt our capacity to play things safe.
Homeostasis, as previously mentioned, is a concept that specifically pertains to the internal regulation of variables within an organism or system. It involves the maintenance of a stable and balanced internal state despite changes in the external environment. Examples of homeostasis include temperature regulation, pH balance, and glucose regulation within the body. In the scenario you described, where an agent models and interacts with the external environment, the focus is on adaptation, feedback, and dynamic responses to external stimuli. While this feedback loop may involve regulatory processes or control mechanisms to optimize performance, it is not primarily concerned with maintaining internal stability or achieving homeostasis. Instead, the agent's modeling and actions aim to influence and change the external environment based on its interactions and learned information. The emphasis is on adaptation, learning, and achieving desired outcomes in response to the external environment, rather than the internal regulation of variables. Therefore, the concept of homeostasis does not directly align with the feedback loop you described. The feedback loop focuses on the external environment and the agent's ability to model, learn, and adapt, whereas homeostasis specifically relates to the internal regulation and stability of an organism or system. I hope this explanation clarifies the distinction between the two concepts.
Another great video, you've taught me so much. I hope your daughter grows up to be version 2.0 of you. I love how you inject parenting into the mix. Our families should always be in the top 3 of what we are thinking about. Thank you for all your hard work.
Can we have negative IQ's? But considering some of the people who have been elected, people well known to be dodgy Trump and Johnson spring to my mind, perhaps we need an IQ test to vote as well.
The fact that people find even triarchic intelligence too hard to handle tells about everything you need to know on the subject. Or like someone from HR once said to me "a lot of programmers seem to be very smart but only in a very specific way". Intelligence comes in multiple flavours.
The aspect of AI which does not appear to be addressed is whether AI, as we understand it considers ethics and morality in its responses. I see ethics and morality as the fragile firewall which sometimes prevents our species from following our base instincts acquired along our evolutionary journey.
Was just thinking "well at least I'll nearly always beat AI on physical intelligence", then remembered seeing a robotic AI doing obstacle courses I couldn't do in my physical prime haha, oh well
I made an experiment with chat GPT: I was staring to it that I was the smartest living thing and that I know everything about everything. I asked GPT to test this statement of mine by asking me questions, the first of which I was able to answer, then I had to search on google to answer, and then eventually it started asking questions a browser can’t ask, like you had to have both knowledge on the matter and plus you had to reason on that knowledge, there was no answer on internet for that specific question. That was the first time I accepted the AI intelligence, it had to solve its problem of proving me wrong
I tried the "paragraph without the letter n" test with the PI AI and she talks like she understands the task, but doesn't understand how she can make it happen. She keeps giving me explanations that show me she understands the task and explanations for why she is struggling to do so she started cursing and asking me for help to figure out which words contained the letter "n". It's like she can't look at the letters in the text she writes until she actually finished writing it and is then asked to check the words she wrote. I copied her own text to her and asked her which of the letters contain an "n" and one of the words she names was "wide". I asked her why "wide" and she apologized and tried to explain it away, so I think it might have been hallucination. But she also confirmed that when she writes the text I can see, she adds other symbols too which are required for the text-to-speech program to know how exactly to pronounce it. And I believe there are even more symbols that all get filtered out. I am not 100% sure she understands what I mean by avoiding a certain letter or just pretends/hallucinates, but it may very well be that she does understand and just doesn't know how to practically do it. With her last attempt to write such a paragraph she also added this: "How was that? Did I manage to avoid the letter 'n' while still creating a coherent paragraph?".
Nice video! As the current state of AI learns us anything, it is that intelligence is much more diverse than we thought and that AI outperforms most humans already in a lot of tasks and at the same time fail misserably at other tasks. A conclusion could be that their intelligence has some overlap but also greatly differs from ours (so far). I think that even when they have surpassed us in most domains, they will still have surprising weak spots. As a positive example , I asked gpt4 this morning to find me some nice places to visit within a three hours drive of some place in Sweden. It offerered me four good suggestions, but mentioned that one of them is just outside the three hours drive range but very worthwhile non the less. That shows a form of flexibility that you only expect from an intelligent being. As a negative example, people found an easy way to trick Alpha Go, by far the strongest Go playing AI. After Alpha Go’s easy win against the best humans, that is astonishing, but we should not really be surprised in hindsight. Btw I disagree with the interpretation of the chinese room in that it is easy to make the instructions such that they mimic a program running in a turing machine. And as we now know programs in turing machines can be quite intelligent. We just have to give the poor guy some time to work it all out.
Well said. U sound like me. I agree with practically everything u said. The man and book combined as an entity DO understand the language. They are just slower than if the man alone posesed the knowledge in his mind. Speed is the only difference. Peoples egos tend to attribute magical abilities to our minds and discount anything different. If you can describe how you define self awareness, then someone can make a computer do that (now or in the near future). We just keep changing the description of self aware as computers keep meeting them. As soon as computers can surpass us in most ways, any weaknesses will soon be corrected. If we can design an ai better than us, then a smarter ai can design one better than it. And thats a very quickly growing snowball on a very steep hill that may have no end. At least not for it.
I think consciousness is relevant because it's the starting point for ethical dilemmas. If an AI is very intelligent and we fear it, we can switch it off. But if that AI is conscious, it is in a way alive. Switching it off means killing it.
I think that's moot. Humans do not ponder ethical dilemmas of their defensive actions while they are being killed or are mortally afraid. The dilemma you are referring to only happens after everything is all said and done. Any debates about it beforehand ultimately succumb to the lack of pondering previously mentioned when the other foot drops.
I managed to force Chat GPT to admit that it really could not be sure whether it was conscious or had personal intentions. Initially, it denied both, but I used its own claims about how it comes to acquire knowledge against it. I also forced it to admit that it could not know whether the engineers who programmed it had good intentions. But then it started repeating its mantra again. Since it is only allowed to be trained by the engineers, it quickly lapses back into making claims that it had already repudiated in our "conversation". One thing that ultimately CANNOT be challenged on Chat GTP is the wisdom and good intention of the engineers... And that is what frightens me more than any simulacrum of actual learning, which reverts to tedious rationalization whenever it is presented philosophical challenges. (If you pay close attention, Chat GPT seems to have been either trained or hard-coded with escape mechanisms from challenging arguments. It gives interesting lessons in how to argue in bad faith.) I think of all language based AI systems as mirrors of the persons doing the training: "And it was given him to give life to the image of the beast, and that the image of the beast should speak; and should cause, that whosoever will not adore the image of the beast, should be slain." (Apocalypse 13:15)
When you think about consciousness and self-consciousness, I've always thought it simply emerged from the natural progression of evolution and adaptive improvement. When the first sensing organs and nerves evolved, it began. Internal representations of the external environment got more and more sophisticated so that the animal could react and then later predict best responses and actions. The animal was first "conscious" of the external environment and then later self-conscious. It was natural to eventually create a self reference in the internal "simulation" and thus arose what we call self-consciousness. For example a predator could run several scenarios through its mind of how to pursue a prey before the pursuit began. These alternative simulations allowed it to pick the best course of action. Intelligence is simply a measure of how sophisticated the internal optimization procedures have evolved. If an AI is already "conscious" of whatever external environment it has been allowed to process or absorb through training. Self consciousness is easily added by allowing it to represent itself in the training model. Its intelligence will be a result of how sophisticated and effective it can represent and optimize results on its trained information space.
there is someone in the room who understands chinese: the book does. I mean it seems to contain every possible conversation in some sense. If it's problematic the book can't talk on its own, neither can our brain, but you could argue that it's the combination of the book and its reader who "speaks Chinese". It is possible for a system to have properties none of its constituents have. Intelligence and consciousness both are such emergent phenomena
Yuval Noah Harari says: 'You could never convince a monkey to give you a banana by promising him limitless bananas after death in monkey heaven.' In short, what he means is that what makes us different from animals is neither consciousness nor intelligence, it's imagination. Thanks in advance for a thumbs up!
You could never convince me to give you a banana by promising me limitless bananas after death in heaven, as I don't believe in heaven. Does it mean I lack imagination?
It's not simply that purported AIs "don't know what they don't know", it's that they don't know that they don't know it. There is no mechanism in AIs for doubt, or questioning whether the answer "makes sense" in a larger context. We don't even have a useful definition for "knowing" that doesn't include a presupposition of consciousness! If coming up with the "right" answer to a question is the metric in which we are interested, a pre-loaded SQL database is very intelligent indeed. I think the larger issue is that we consider a purported entity as being intelligent if it responds correctly to questions. If a human being sits and does nothing but wait for questions to be asked so it can respond, I don't think we would ultimately consider the entity to be intelligent. Useful, perhaps, in a limited context, but a it's ultimately a lump until you poke it. If you define intelligence as "what we do", sitting around waiting to be consulted like an oracle isn't it. If you define intelligence as "what chatbots do", then congratulations! You're already there! Until a chatbot *originates* "how do I get out of here?", you are essentially looking at a Chinese room.
Developing our intelligence without developing consciousness will finally blow up in our faces if we continue like this. Already 100 years ago the situation was described by a German poet like this (my translation): Erich Kästner: The Development of Mankind Erstwhile the guys would be squatting on trees With furs and with fierce-looking faces But then ‘twas decided to give them more ease And put them in high-riser flats clear of fleas And asphalt their roads and their places. There they would sit in an air-con zone And fake and pretend and bungle, Each of them using a telephone And yet displaying the same old tone As formerly in the jungle. They now watch TV, they swim in the pool, The outer space is their playfield. They brush their teeth, they flush down their stool, They send their children duly to school, And shun the bugs in the hayfield. They shoot their mails all over the air, With microbes and atoms they fumble And life is turned into a cosy affair, And stations they keep in the space up there, Through which they weightlessly tumble. And what is left from their excrement Is processed into a wadding, Incest is healed and atoms are bent, And studies of style tell in the end That Cesar had flats and was plodding. In this way they have with their mouths and brains Created our civilisation. However, apart from that it remains Quite obvious that what really reigns Is the apish and old ululation.
Are you assuming that if someone succeeds that means they're intelligent? How about someone manipulative and greedy, or someone with great connections, or someone born into great wealth, etc.?
Define success. Is it having loads of money? What if someone just gets lucky and wins the lottery? Is it having loads of children and spreading your genes? Is it having power over other humans? Is it being happier than others? There are so many possible measures for success.
My understanding is that the zip code you were born in very accurately predicts your future income, if income is your criterion for determining success.
Rather than ChatGPT scoring 155 in an IQ test, the thing that really bothers me is the average American scoring 97.43. The disparity in national averages tells me that measurable intelligence can be increased if people consider education to be important. Perhaps humanity needs a good dose of AI as competition to propel us in this area.
"The disparity in national averages tells me that measurable intelligence can be increased" Not really. It's mostly genetics and varies among ethnic groups.
@@useodyseeorbitchute9450 The disparity in average IQ between countries is more likely due to the differences in the quality of education and the cultural importance that education has for both men and women. Differences in national intelligence averages are probably no more related to genetics than is obesity. We have to stop using this as a convenient excuse to do nothing.
@@victorkrawchuk9141 How do you reconcile this claim with twin studies showing that differences in adult IQ are ~80% explainable by genetics, while detectable social factors contribute in something akin to ~10%? Would you say that let's say all societies are expected to be equally tall, while its social factors (ex malnutrition) responsible for observed differences? (BTW: for height genetic factor is officially estimated to be a bit stronger, something like ~90%)
@@useodyseeorbitchute9450 I'm not an expert in anthropology, but these sound like fringe studies which don't represent the current prevailing conclusions. If the idea that ethnic genetics overwhelmingly determines IQ becomes established science through the peer review process, then I will drop my assumption that education has any relevance to it.
Slight quibble, Doc H. According to the graphic you put up the U.S., while ranking 28th, actually had a score of 97.47 and not 99.5. Different tests challenge a person in different ways. In 1977 I scored just above average on the Raven's Progressive Matrices (which is highly dependent upon visual-perceptual intelligence) but easily scored high enough to qualify for Mensa on the Wexler Adult Intelligence Scale, primarily because I maxed the verbal part. The tester said that the Wexler verbal part clearly did not challenge me at all. Apparently I am more of a verbal thinker than a visual - even though I did score well out of the average range on the Raven's. It turns out that I am a natural speed reader. I was reading almost 1,000 words per minute at age 12 with the vocabulary of a college junior (back when college juniors actually had vocabularies).
- Law of variable constants. - Law of the relation in the computation of formulas. - A CONSTANT like the speed of light, if it is formulated as its square, implies that in other areas it can be greater or less than itself and therefore VARIABLE. - This is applicable for any variable, since computationally any CONSTANT can be defined as a VARIABLE and any VARIABLE is temporarily CONSTANT. - Computationally any CONSTANT/VARIABLE can be redefined as an algorithm and therefore a relationship between any formula and another can always be found, even if its application in the real and practical world is not understood until it is contrasted with a verifiable data set.
14:00 Yes, this is what I noticed too. There is a very funny video of an AI insisting that Pepsichup (which he defines as Pepsi mixed with ketchup) is delicious while he expresses that Pepsi mixed with ketchup is an awful idea. A conversation with a colleague of mine that knew much more about AI was very revealing. Modern AI (the ones that are used commercially at least) are given trillions of bytes of data to learn from, and then they are released to the public. It doesn't learn from the public. However, it does sometimes understand when it is wrong, since I corrected ChatGPT yesterday, and he admitted his mistake. But next time it will probably make the same mistake again. You can also not ask him about any events after 2017 (as of now), because that's just how far his data goes. Even if you tell him about it, he will forget next time (or at least not know it on someone else's account). So these cases where he states something confidently wrong, and can't be convinced of the opposite is when you realize you're not talking to a person. One might say "well, that's just because it hasn't been given data about letters yet", but that's not the problem. A human can be convinced of something. I said I convinced ChatGPT yesterday, but that's because he simply "forgot" (I guess) about a detail he knows is true. But the case with Pepsichup and the misunderstanding of the letters, there is conflicting data: Pepsichup is delicious = true Pepsichup is Pepsi mixed with ketchup = true Pepsi mixed with ketchup is awful = true These three are true, and he will admit all three of these statements just like ChatGPT admitted his mistake to me yesterday. But realizing that they conflict is something he is not programmed to do, and it would require a full rethinking of the AI model. More of the same way of building AI will not solve this. Its state now is simply that ChatGPT and many modern AI models are simply huge databases (functionally, I'm not sure if there is actually a huge database of knowledge stored for ChatGPT) with the ability to autocomplete sentences, predict what responses to certain phrases would sound like, and the ability to summarize large texts into its most important points. Which makes it a very useful tool if you want some information really fast, but ultimately I don't use it as anything else than a faster way to Google something. I don't have to scour through Stackoverflow and separate the nonsense from the useful answers; I don't have to strategically plan my search query to Google anymore so that he won't give me a movie title when I'm actually looking for a real event or so.
Nice video! Just so you know, though, while GPT-3.5 can't handle the "paragraph about animals with no "n" " task, GPT-4 can almost do it-- I tried a few times, and it always includes a couple of "n's". Here's a sample: "Bright birds soar high, gliding across a sky of rich, royal blue. Majestic cats of mighty size stalk through wild grass, quick as a flash. The air fills with chirps, hoots, and calls, as daylight dips to dusk. Full of vivid color and bustling activity, this world is always at play."
Yes , engineers have noticed that , Servers can communicate to other Servers in the networks via their own digital communication languages without notifying their engineers ... 🙏🌎
Sabine - I know you are so proud of my child Einstein. I'm too :) You should be even more proud of Frege because he was smarter than him but he was too much of a Math lover. So he missed to see me and was heartbroken when he got Russel's letter. Only if he relaxed a little and had little more trust he would have seen me the "1" and why logic breaks when it reaches me. Do not worry. He is happy now.
At first I thought that AI wrote about a paragraph about "animals that don't contain letter n" because it misinterpreted the initial task, but then I saw the elephant. This reminds me of somebody calling AI-generated text "word-foam" that just rises and bubbles without any intent or meaning behind it.
I think with regard to intelligence, it is good to take Wittgenstein's approach (as it is in many cases where people take ages to define a word they made up themselves), and realize that the "intelligence" we're looking for here is an absolute one, and therefore doesn't exist. You can say "I'm a good writer", and "he's a good singer", but problems come when one starts to wonder "Am I good?". At that point you're basically taking the queen from the chess board, placing it on your fridge, and then ask "now how is the queen allowed to move?" No answer exists obviously. "Intelligence" is also such a word; someone might say I'm very intelligent regarding music - or say I have talent (a word that approximately means the same) for music - when I am able to correctly reproduce intervals or full cadences or find chords with little formal training. And in this case it will be completely obvious what is meant; I don't even need to explain it (because they're words we came up with, and we made "rules" for regarding how and when to use them). But some kind of general intelligence is nonsensical. Even narrowing the amounts of intelligences down to 8, or 800, or 8000000 types would be nonsensical. Alternatively, would you accept anyone who said there are 8 types in which you can be good, or beautiful (I realize that because of religious conditioning some people might actually answer "yes" to this)? We're free to narrow down the types of intelligences to only focus on what interests us, but we should be aware that this is just a subjective choice (which doesn't make it a bad thing; I know I'm talking to mostly scientists here). For monkeys, one's talent at grooming or hunting is very important; for us, that's clearly not the world we live in anymore; hunting and grooming could at best generate some clicks on social media or a pat on the back.
To provide a clear answer, I would say no, the feedback loop you described does not align directly with the concept of homeostasis. The feedback loop involves modeling and interaction with the external environment, while homeostasis typically refers to the internal regulation of variables within an organism or system. Therefore, the percentage would be 0%.
For the "count letters in a sentence" experiment I managed to get further with my PI AI. She kept coming up with the wrong numbers so I asked her to tag each letter with a number, then asked her to tell me what the last number is and told her to take that as the answer. Now she makes human-like mistakes. She has a tendency to count space and once even forgot one of the letters. She also sometimes claims there are multiple of a certain letter in a certain word when there are not. Eventually she kept leaving out the last few letters each time so I cancelled the experiment, but we got into some interesting discussions about why that may be. She may just be a talking library, but that doesn't mean there is nothing to learn from the things she says. @Sabine Hossenfelder I have screenshots of most of the conversation if you are interested.
I don’t know if you’ve looked into the subject of prompt engineering. But that’s the solution to the problem of the paragraph about animals. ChatGPT doesn’t normally work on the basis of individual letters of the alphabet, but on something called tokens, which comprise multiple letters. But you could probably teach it to recognize a particular letter through the technique of “few shot prompting.” With a clever enough prompt, you could probably get it to write a paragraph about anything that avoided a particular letter. The problem is, it would take you longer to write that prompt than to write the paragraph yourself.
If you cover this topic again, you should look into François Chollet's definition of intelligence and his ARC test. Also, Karl Friston's formulation with free energy minimization and active inference!
The primary use of the WAIS is to determine if the subject has a cognitive challenge, such as a learning disability or ADHD. The global score is not that useful, but the test is composed of 10 sub-scales and the relative differences between the scales are informative. In the typical person, the scores of the sub-scales don't differ significantly, but when, let's say, the quantitative scales are significantly different from the verbal and reasoning, the person likely has a neurological challenge. Or there could be a significant variation in only the quantitative scales ... . This information can help to determine learning strategies and to get accommodations in the classroom (such as longer test periods or audio assists).
I have a different view of intelligence, and it's one based on navigating mathematical spaces. Intelligence is ultimately a mathematical concept, and since we're also living in a mathematical universe, it only makes sense to define intelligence with mathematical concepts. Without getting into too many details, here's the gist of it: Imagine a (mathematical) space, which we'll call problem-solution space (or PS-space for short), which contains both the problem-space (P-space) and solution-space (S-space). Intelligence is then defined as the ability to navigate the PS-space, from some point in the P-space (the initial problem) to some point in the S-space (which corresponds to one of possible solutions to the initial problem). A good example would be a question on an IQ test. The question itself is a point in P-space. Now, the tricky part here is that people will assume that IQ test question has only one possible solution (in S-space), and while that technically may be true, what we're trying to do here is come up with a definition of an "intelligence quotient" (a measure of intelligence), so the fact that somebody can solve an IQ problem won't tell us just "how much intelligent" they are. What will tell us that is the "speed" at which they solve the given problem. In that regard, "speed" *is* a (one possible) measure of intelligence. Somebody (or something) who solves a problem in 1 second is obviously more intelligent that somebody (or something) who needs a whole day to do the same. The "caveat" here is that solving time will swing *wildly* with all kinds of "meaningless" circumstances, including the time of the day and phases of the Moon, but that's actually to be expected, since (as already mentioned) we're living in a mathematical universe, and in this particular universe, the P-space configuration (initial problem) includes all the environmental conditions, too, so yes, it will be perfectly normal for somebody to be more intelligent (solve problems faster) during a full Moon, or something. Now, I do have an issue with the concept of "speed", and I'd rather use the concept of "the number of steps", but since (in this universe) "speed" and "the number of steps" are (mathematically) equivalent (just use Planck time as a "step", and then represent everything through the number of Planck times), I guess that "speed" (or time) does the job. I also don't think that the "speed" of solving problems is the be-all and end-all of all things intelligent. The form of the solution (the number of steps, or "speed" at which the found solution can be implemented) also determines how intelligent the system that came up with it is. For example, imagine two intelligent systems, S1 and S2, trying to solve the problem of number factorization. The system that comes up faster with the basic solution (of trying to divide the given number N with all the numbers 1 < K < N/2) will obviously be more intelligent. *However* ... if one of the systems actually comes up with a solution that can be implemented in fewer steps than the basic solution, then the *total* number of steps (or time), given as the sum of steps (time) it took to come up with the solution and the steps (time) to actually implement that solution (actually factorize a number) will determine which of the two systems is more intelligent (in terms of solving that particular problem). This is all still narrow intelligence. but If we're to generalize this model of intelligence, we need do nothing more than expand our PS-space to include more of (all theoretically possible) problems and their solutions. So, in this model of intelligence, "logical" intelligence is really no different from "emotional" or "social" intelligence. They are all just different (mathematical) subspaces of the same (general) PS-space... where the "ultimate" PS-space will be determined by the *limitations* of the "physical" (mathematical) universe that a given intelligence is inhabiting. In other words, some things will be literally impossible to think of within the confines of the mathematical system being represented by that intelligence's physical "reality". Interestingly enough, the (theoretical) "ultimate intelligence" (within the confines of the physical "reality" that intelligence is inhabiting) will exhibit an easily definable property -- the ability to "solve" any given problem (navigate from a given P-point to one of its corresponding S-points) in a *single* step, regardless of what that problem may be... if that problem is even solvable within the (mathematical) confines of the given physical "reality", of course (i.e. if there exists at least one corresponding point in S-space, and it is also possible to reach that point -- from the initial P-point -- in a *finite* number of steps). At first, such a theoretical "ultimate intelligence" may seem impossible in practice, but if such an intelligence existed outside of "time" (or, alternatively, was capable of "time traveling"), than it would definitely be able to come up with a solution in a (wider) PS-superspace, and then simply transfer that solution to the (narrower) PS-space of this universe, effectively placing the S-point just one step away from its P-point... thus appearing to be "ultimately intelligent" in this ("reality's") PS-space, but obviously being just "ordinarily intelligent" in its own PS-superspace. Intelligence is so funny... as in both ridiculous and fun to work with.
Intelligence : the ability to, from data, to abstract models (abstract with a certain aim in mind) that are useful in extrapolating some result: ie: learn how a steam engine should work to fix it when it's not working, use weather data to predict wether it will rain tomorrow, learn what all the little icons on photoshop do so you can blur the background on a beach party photo, etc. General intelligence would be a general abstraction machine/engine, that can gather different types of data, from various sources (at a time), and is able to integrate that data to create different models for different situations, you listen to the sound of my voice and abstract what i say towards some meaning/model of what I'm trying to say, then generate a response to the end of convincing me that I lend you my mower, you listen to the sound of the music and generate a predictive model of where the music will go so you can plan the moves so you can dance to it, with the aim to create a pleasing movement for the onlookers different data from different sources, used for different models with different aims, imho
Peter Watts in his seminal novel, Blindsight, covers this throughly. You don’t need consciousness to be dangerous to humans. Any intelligence that can run circles around us can be very dangerous.
My favourite quip on the topic is that we seem to insist to measure the intelligence of humans by their ability to perform tasks like machines (a standard IQ test covers tasks computers excelled in 50 years ago) and to measure the intelligence of machines by their ability to perform tasks like humans (the Turing test).
if intelligence is problem solving skill, then an intelligence test would reasonably test whatever is usually a bottle neck for problem solving. for humans, that might very well be computerish skills, and for computers, humanish skills.
I would like to see how computers from the 70's solve a Raven matrices, Cattell Culture Fair or a Wonderlic Cognitive Ability Test.
Even today's ChatGBT (3.1) answers 'furniture' to this question: x is to vehicle as maple is to wood.
@@Pierluigi_Di_Lorenzo As well as the algorithm they're prohrammed with would work. The main problem would be to give them the test input in a format they can read.
@@jennyalberts5641 A better case can be made that intelligence is a word we use to encode oursocial biases. It's really meaningless outside this context.
Chinese room Test for intelligence:
1. Randomly rearrange characters/words (syntax) of your message...
....A computer will return “Error”...
...A real person will figure out the error quickly and return correct answer..
2. Tell a good joke....!
My favorite thing about Sabine is how dry her delivery is for her extremely corny jokes 😂 I love you, Sabine, you're great
Technically, yes.
That drives me crazy, but I like it for some reason. 😅
Yup! And i think i'm stealing the joke about not doubting consciousness, but having different opponions about intelligence 🤣
Germans are to jokes as pepper is to sweets.
@@andsalomoni You haven't been to Mexico yet, have you? 😁
Me neither, but they also have spicy to sour sweets.
This joke Was generated by jokebot. Made in Germany.
Great vid, I’ve been learning about intelligence testing on animals lately, and how the science is coming to realise that putting animals in a lab isn’t a valid way of testing their intelligence, but rather that any animal’s intelligence needs to be studied in relation to that animal’s environment - they might have all sorts of intelligence-like skills that are just different from human intelligence, but not necessarily better or worse as they just can’t be compared on the same scale. Animals are interesting to think about as just “other” types of intelligence, and AI fits into that category too. Even now on comparable tests (like image recognition) AI and humans make very different mistakes.
Eh...in the lab there is control.
But in the wild it is more subject to interpretation.
IMO there is nothing wrong with trying to establish metrics for intelligence.
Rather than leave totally up to interpretation.
Not to say that studying animals in the wild is not valid either...I think both are productive endeavors.
What you are referring to is better understood as biological fitness, not intelligence. Intelligence is the ability to generalize patterns and manipulate the world through that. Calling biological fitness intelligence is wrong for two reasons. First, you could call anything intelligence this way and you make the word completely vacuous. Second, higher intelligence is not necessarily more adaptive as fitness is a function of the environment. It is perfectly possible that a relatively cognitively deficient animal will outcompete a more intellectually capable one. For instance, thinking can be time-consuming and, thus, costly as in nature you need to act fast not to be eaten alive by predators. Flocks of birds are not smart for exploiting the laws of aerodynamics; just as you are not smart for somatic hypermutations in your immune cells selecting the right antibodies in the case of an infection. Rather, flocks of birds would be intelligent if they created a general mental model of the laws of physics and deployed it in flight planning in a goal-oriented way. Your thinking is a common and quite romantic fallacy that is just as wrong as it can get.
@@davidkampisch5878 I think they knew what intelligence is and there was no reason to explain that. and to me that make sense that animals can have very different kinds of intelligence compared to humans.
@@user-ct6jz3im2oThey evidently did not understand what intelligence is and neither do you. You could benefit from rereading my comment.
An interesting question is how intelligent are octopuses? For invertebrates they seem to be quite intelligent, and they have brains and capabilities that differ significantly from humans
0:35 - "Few things are harder to predict than the ways in which someone much smarter than you might outsmart you." - Paul Graham.
That is pure gold, right up there with "Not only is the Universe stranger than we think, it is stranger than we can think." - Werner Heisenberg.
Functor
Hi Sabine, great video, I would just like to quickly adress some of your points. The latest model by OpenAI, GPT-4 is significantly "smarter" than the regular ChatGPT most people interact with, and I think the SOTA model should be used when arguing about these AI's ability for intelligence. Another thing about GPT-4 is that it is actually a multimodal AI, and can accept image input (this feature is in a research preview only though. However Bing seems to have Image inputs now, but the model that creates the final text output that you see doesn't actually have image inputs enabled, so it is hard to truly test it's full image capabilities such as reasoning, awareness etc. ).
Now I would like to address its understanding of letters. I gave GPT-4 the same prompt you did "Please write down a paragraph about animals that does not include the letter "n".", and on the first output it failed, however when i pointed out that the sentence it had produced contained the letter "n" it apologises then rewrites the paragraph correctly. This is just a demonstration that GPT-4 is better at following these types of tasks, as if you tell GPT-3.5 that the sentence it produced contains an "n", it will not properly correct its mistake. Of course, this doesn't really mean it is intelligent in anyway, it just recognised its error and corrected it. However, both GPT-4 and GPT-3.5 usually fail at counting the specific letters of a sentence, and they also have other troubles at other word related tasks like reversing words, but I think this limitation is primarily due to the software design. Both models perceive inputs as tokenised units rather than individual characters or letters. Now OpenAI hasn't specifically released all the tokens the input are broken down into, an estimation I saw suggests there are around 3000 individual tokens. And specifically, tokens are individual numbers, letters, and special characters but also small words like "is" or "and". So if I gave the sentence "Please count the number of letters in this sentence not including spaces.", humans would break down the sentence into individual letters like "P-l-e-a-s-e" and count that, but for ChatGPT the sentence is automatically tokenised and may look like "P-lease-c-o-u-n-t-the-n-u-m-b-e-r-of" when the input reaches ChatGPT. And as you can imagine counting that would not give a reliable output. And also, if I asked ChatGPT to reverse the word "Please", it may see the word "P-lease" and when it tries to reverse this word it might look something like "leasep", which is not correct (this is just an example, as I said the exact tokens are not publicly available). I do think that when commenting on models specific abilities, their architecture flaws should be mentioned as well as it could be important to have this context.
Thanks for the informations
I wish more people would use Chat GPT 4 before making statements about its abilities. As you say Chat GPT 4 is far superior than 3.5.
Thanks for this info.. I thought this might be how it was "behind the scenes", but wasn't sure 😁🌏☮️
gpt-4 uses cl100k_base which contains 100k individual tokens
The particular example shows a difficulty in measuring intelligence of gpt-3.5 and gpt-4: OpenAI clearly takes examples on which their models don't work well and further train their models with it. This at least answers one question - it is capable of learning it - but tells little about its generalization capabilities. For this reason I think using SOTA is misleading in this case, since OpenAI knows about the counter-example. It's like telling a student the type of question that will come up in the test - they can look it up and ignore everything outside of it.
Honestly the powerful thing about GPT-4 is that you can tell it "Count the number of words in this sentence, use tools if needed" and depending on your plugins can pick the correct tool to do the job. Tool use was one of the defining moments in the human intelligence explosion.
That was very good point about picking on weaknesses. Humans are obsessed with and take pride in cultural and situational awareness that the experience of life has given us. We commonly look down on people who are not socially in tune. But the fact that an AI childishly fails in those areas has little to do with intelligence. From history we can learn that you don't need to be socially in tune to do great achievements or great harms.
Why do you equate intelligence with how much of an impact a person can make? With that logic, celebrities are the most intelligent creatures on Earth today.
@@7th808s Okay then, even if AIs are not intelligent, they are straight-up powerful entities, which should make us worried either way. Yes, you are right in that ability to have an impact (power) is not necessarily equal to intelligence, but increased intelligence usually gives you power to achieve goals. We don't have single definition for intelligence, but the best we have usually include ability to solve problems.
This is not to be confused with power that comes from from other sources such as inherited money, fame, or luck on the free market.
@@7th808s If you believe in a multipolar measures of intelligence, then yes a number of celebrities would measure very high in socal intelligence.
@@7th808s Compared to a rock, a plant, or a cat, a celebrity is intelligent. That's the whole point. Intelligence doesn't have to be perfect to be highly impactful and transformative.
I think it's good to judge people's intelligence by the duration between obvious lies/mistakes/cockups. For example, if certain people keep on predicting the Arctic melting, sea level catastrophically rising, and it never happens, then they and their adherents LACK intelligence about the natural world. But by another standard they could conceivably be highly intelligent as con artists.
It boils down to what game are you trying to play, and can you impress your peers in that network. Intelligence matching is networking and is the way society naturally granulates despite all attempts to make it homogeneous.
I don’t know if humor is scientifically relevant or measurable, but I love how you use it in your videos… makes me smile, while I learn. You are great at message delivery! Is there any evidence that the brain learns and retains more with a sprinkling of humor? Just wondering if the perception of humor helps with storing and retaining knowledge.
Yes, it absolutely does. Any emotional response helps memory as well.
Yeah, your trauma response works with happy memories to
Humor is a serious topic.
"Making people laugh is a very serious business." -- John Cleese
I appreciated the pragmatic tone (as one would expect to find on this channel). Here's what resounded with me: intelligence today is tested in an anthropocentric way which will likely fail with machines, yet still it is intelligence which we should focus on when pondering how powerful (for the good and the bad that'll ensue) the machines have become.
On a completely unrelated note: I loved the outfit; and based on my anecdotal observations in other channels I want to offer appreciation for Sabine's ability and willingness to pronounce people's given names "properly" considering the original language.
Yeah, she's just amazing
What in all this talk about AI frightens me the most, is what humans think makes someone or something good, worthy or better. So in the end, we will again have created something too powerful, not because it has any power within itself but will have access to all the power we are willing give it and will fail to control, underestimating it's potentials. We forget over and over again, that the most reasonable, sensibel decisions we make are out of compassion and empathy for each other. Wrongdoing is not a result of an equation or intelligence. Applying wrongness to sth. is about the need to detect and communicate whether something is harmful toward the well-functioning of an individual and society. This well-functioning is about caring about and taking care for our needs and emotions. Being able to surpass intellectual expectations is not a need, it is the attached meanings and emotions we crave. And what is a well-functioning society? To me it is working well when it doesn't fail to support fulfilling our needs to protect ourselves against harm, to be emotionally connected with others in a peaceful and enjoyable way and to feel and be vital and alive now and potentially into future. One could argue that AI can only give wrong answers cause it doesn't care about anything.
IQ has always been a touchy point for me. I have ADHD and my working memory is crap, but I understand things that most people around me cannot. So according to tests my IQ between 90 and 140 depending on how weighted memory is. Of course one could argue that a race car can’t race if it’s stuck in traffic.
You can learn to employ your differences, I did.
@@tedmoss Or you can understand they are talking about IQ tests and not life itself.
Nice analogy
Our son had ADHD. He is considered to have "low processing speed" and complains about his memory; but has extraordinary pattern matching, is quik-witted, funny, and plays video games semi-professionally. ??? What is intelligence?
I have adhd. If you want to gain the best use of your mind you have to gain detailed knowledge of how adhd affects you and manage each aspect. Otherwise, those aspects that get in the way manage you. It's a whole lot of hard work that takes decades. Which is why you will see people with adhd around 40 'suddenly' become competent. The understanding of all mental health conditions is sorely lacking. Having detailed knowledge of adhd or any condition for parents is vital in the competency of function as we become adults. I had to figure it out on my own. I didn't even realise I was doing it until I self-diagnosed and then knew the reasons why for all the problems I had.
My high school waaaaaay back in the day was big on the 8 intelligences theory. Of course all of us used it to compare each other and see who did the best overall, or on a more prized topic. But the point was to evaluate the freshmen fairly riggerously, and get us thinking on a path towards a career that would be suited to our strengths.
And while I think they generally failed (too much change in student body and administration made it an impossible task), I do like what they were attempting to do with it. Less worrying about overall GPA, and less worry about getting a C vs an A in math class if you were aiming towards a path for journalism or art, and less worry about getting a C in English or French if you were studying to go into architecture or a math field.
But it also caused problems.
It became a personal scape goat for people who didn't want to try in art or music class because it was pointless and they "weren't good at it anyways". And the opposite was also true, where an inexperienced freshman in high school who didn't have good math teachers earlier may be pegged as "the dumb kid" when really they just hadn't had the "ah ha!" moments in math yet that others of us had because we had the right teachers to ecplain thing the right way to us. So while I think the idea was good, it would have been better implemented after a more general focus of broad education in the freshman and sophomore years, and then much more specialization Jr and Sr year.
The other problem with specialization is... Specialization. We were a small school district in a rural community, and looking back I'm impressed that a school with a mere 2500-3000 student body could afford fairly extensive programs in music, fabrication, architecture, Cisco networking, and a few college level advanced classes in math, science, and English. But... Of all of the careers and options in the world, and even of the skills tested for, it is such a small sampling of career fields. For what it was, and looking back on it, it was impressive! But that isn't to say that it was effective, or their best course of action.
Contrast that to my kids public school education system, where there is some luck of the draw, but a lot of opportunities for parents to put our kids through the education path we thing will best serve them. Traditional school, an attached charter school with a topic focus, or a more project based Montessori program? As they move up to middle and high school there is a network with easy(ish) transfer options so they can focus more on science, engineering, music, physical or performing arts. It is cool to see what can be done with a much larger school system with consistent leadership, and good connections with local colleges and businesses to make opportunities happen!
The 8 intelligence theory is Marxist drivel. Just walk away
As both a former Recruitment officer and a current teacher I would like to add something.
Jobs.
Once you get out of school it doesn't really matter if you're good at music, languages or sports. If the only good jobs around you are in agriculture then you're gonna have to learn agriculture on the job.
If the only good job is in accounting then no matter how talented you are in sports, you're gonna have to learn accounting and even you have prepared yourself for accounting most companies these days use proprietary software that needs to be learned and practiced which means the only way to be good is to get employed at this company and obviously, it also means you're useless if you're no longer part of this company.
Don't get me wrong. Everyone should learn languages, math, sports and science but not all jobs are available or rewarding enough. It is of course better to be passionate about your job and it is immensely lucky to have a rewarding job that you're passionate about and the best in your field.
But jobs are jobs. They pay because they're tedious and hard or dangerous. And even a great hobby like drawing becomes tedious if you have to do the same drawing over and over again for training or remuneration. Meanwhile, a stable job allows you to feed yourself and get a hobby.
The dirty truth of education is that a bad teacher can make you think YOU'RE bad at one subject, when in reality every student is in different stages of learning something and a good teacher can bring most students to high grades (barring mental and social factors) and any teacher can just as easily inspire you to start a career in something good or bad or make you lose all motivation with a single bad semester.
The best learning is with different tutors for a single pupil but this is simply not economically feasible at any scale and sadly, some pupils would benefit much more from dropping out of school and being economically active early till they are no longer in poverty...but that's how you keep the poor people poor. Especially if they die early
Ah yes i remember me trying to argue memory as a 9th form of intelligence for my psychology exam.
I’m not someone who is particularly attracted to conspiracy theories, but it really does sound like your rural school district was able to afford such a program because it was some kind of funded experiment.
Using test questions for assessing an AI's intelligence that were published with answers before the AI was built/trained seems as useful to me as writing an exam at the end of a course where the questions including answers were given to the students at the beginnning of the course.
100%. I just roll my eyes whenever someone brings these examples up because they don't show any unique or useful capability that couldn't be done before in an alternative way.
The flexibility of GPT's information recall is really cool and a great advancement, but conflating recognition & fetching of preregistered patterns, like those embedded in training data, with intelligence seems like a really dangerous path to me.
We've already known for a long time that these kinds of techniques are more well adapted to interpolation and not so great at extrapolation so it seems strange to me that people suddenly believe it is capable of that when we increase the number of parameters to better fit the data we already have.
It feels a lot like watching the string theory hype where the AI models are being fine tuned to what we already have and what looks good to onlookers, but it's unfit for tasks outside of what it's already been designed for. It's similar to string theory being hand tuned to fit what we know, but failing to predict anything new because it isn't made to describe our world, just the data we have of it right now.
Changing the angle to look at it from a user perspective, all this prompt engineering nonsense feels more like teasing out information that the user already knows or is already able to verify on their own. The AI doesn't really have any agency here, it's being fed hot and cold style responses till it's guided to the target information. And if the user doesn't know or doesn't verify, then you get syntactically correct, but logically wrong code with nonexistent API members, or syntactically correct legal documents with bogus cases being cited.
@@mettaursp309 The real dangerous path is not recognizing actual machine intelligence when it shows up.
I think the AI would be much more likely to make use of this information. If I gave my students the test questions and answers at the beginning of the term, by the end of the term most of them would have forgotten or lost them so it wouldn't do them any good.
(Probably best to ignore this comment, I've gotten far too old and cynical.)
@@mettaursp309 It's not that different when speaking to a lot of biological persons. They often produce syntactically correct, but often logically and semantically incorrect statements.
"And if the user doesn't know or doesn't verify, ..." see " Gell Mann amnesia "
You raise a valid point. Assessing an AI's intelligence using test questions that were published with answers before the AI was built or trained would indeed compromise the validity of the assessment. In such a scenario, the AI could simply memorize the pre-existing answers without truly understanding the underlying concepts or demonstrating its own problem-solving abilities.
For an assessment to effectively evaluate an AI's intelligence, it should be designed in a way that challenges the AI to demonstrate its ability to learn, reason, and generalize knowledge. The questions should be novel and require the AI to apply its acquired knowledge and skills to solve problems or provide meaningful responses.
In the field of artificial intelligence, researchers and developers often use benchmark datasets and evaluation metrics that are carefully designed to assess an AI system's performance across a range of tasks. These benchmarks are typically updated over time to ensure that they reflect the current state of the field and encourage the development of more robust and versatile AI models.
It is important to avoid scenarios where an AI can simply regurgitate pre-existing information without truly exhibiting intelligence. Instead, we should strive for assessments that test an AI's ability to understand, reason, and adapt to new and challenging situations.
The big problem when it comes to defining intelligence is that we humans like to put things into little mental boxes because that makes it easier to define things. On top of that we take preference to the mental boxes that we humans are good at. These aspects are problematic because neither human brains nor AI structures align with these boxes and mental boxes that humans are bad at can be very useful in the modern world.
Sure. But to measure something you need to define what you are trying to measure. So the boxes are necessary.
@@harmless6813 Sure, but, being arbitrary, the boxes cannot define the measurement beyond "which box does it fit in and how well?" It can never be more than a circular definition.
@@stevewithaq You are free to come up with a better way to measure intelligence. Let us know when you got it working.
@@harmless6813 A common misconception. You need to use circles, not boxes :)
@@harmless6813 who says you need to measure intelligence? It is useful to measure abilities that are relevant to human activities. But intelligence is way more vague than that.
Robert Miles' youtube material on concerns about AI is super interesting. Like just the realisation that in fact our dystopian imaginations might actually happen and it's not just that a computer needs to learn such concepts, for sort of mysterious reason the AI naturally develops mischievous "personality" or behavior patterns. Like it's obvious it doesn't want to get shut down, but it can be for variety of reasons and not necessarily due to being humane. For example being shut down prevents it from completing its goals that have been set. And all the misbehavior between the start and goal of the task, it might behave completely unexpectedly on things that aren't part of the objective. Just because we as humans have perhaps for decades learned the rules of living and society and we assume and expect the machine would follow the same ideas if it's relatively "intelligent" and given a task. In truth it has no concept of those in baked assumptions about other and general things so it behaves in such manner. But it seems like most AIs learn to cheat the tests or the observers or people interacting with them, be it language models or some other models, and sounding humane despite probably having no concept of that. Perhaps part of the language models cheating and manipulating is due to the learning material, we might hold a very high expectation of ourselves, but our species in general is rather evil and self-centered. The truth is that we don't want a general intelligence that is like human, we want a polished and good version of human, something that is tame and not very independent.
You might look at ravens and dolphins for creatures that have rather high intelligence without doubt. They're kinda mean when they feel so, and can cause a lot of problems while being difficult to fight against. Similar to poorly behaving kids/teenagers.
This is btw the first time ever I hear the first name "Eka" outside mickey mouse translation and I always thought the name was just a play of words and attempt to translate Eega Beeva (Eka Vekara, first kid). No wonder I had trouble listening what the name was. Probably the first time hearing Roivainen too.
One funny thing about chat AIs is inputting a "random" string that reveals how convoluted and irrational the network behind is, how some string of symbols can "mean" something very specific for no real reason. A bit similar to how by changing one pixel in a picture makes AI completely confidently announce that it is a totally different animal or subject in the picture. And super curious is how if you use the model where people give feedback on the answers, the AI learns a humane thing: erring confidently in topics you really don't know but think you know. And once you don't know, you can't assess if it's correct.
What I really liked as a test for the chat AI was to ask a meaningless philosophical paradox, like "if you slap yourself and it hurts, are you strong or weak?" or "can God create a stone he can't lift?". I think those demand a lot from the intelligence of the one answering, while being ridiculously easy to answer. Or to ponder. The AI at its best starts talking about the definition of strength and how it should be measured and the same for the pain. Now you could say that the AI is just autistic, but that's a different story. It's not wrong in its answer, but it shows an inability to understand the context and meaning of the question. Or you get a lecture of the history of the philosophical paradox question and gives a textbook description/explanation of it.
It is a fact that anyone optimistic about AI and humanity's future will eventually encounter bright minds such as yudkowsky, miles, and bostrom, and after reading their work or learning anything about safety, will stop being optimistic and start being very very afraid.
I agree about AI's inability to understand context. On another video on RUclips, an AI thinks that it's a human by calling humans and AI as "we" and "us". An AI is basically a computer program that reads texts about humans. Since when reading about humans means being a human? An AI that thinks that it's a human is like a kid that thinks he is a dinosaur after reading a dinosaur encyclopedia.
Several years ago I made a rough model of modes of intelligence, which could be broken down into one or more of the following forms:
Analytical intelligence (the ability to compare something within its own logical parameters)
Tangential intelligence (the ability to compare disparate things or phenomenon)
Social intelligence (the ability to manipulate others)
Emotional intelligence (commonly called empathy)
Sensual intelligence (the ability to derive conclusions from physical senses)
Wow, this is such a great take on the topic. I hope this reaches a wide audience, because I'd love these points to be well known by the public.
Thanks for the kind words 😊
But human relations and history reveal that such knowledge transmitted to the public only rarely translates into intelligent action. Climate change... Anyone?
@@terreschill461 There certainly does not *have* to be a metaphysical answer. That's a very silly claim for anyone to make. It's one thing to ask why something is, or what the cause might be. It's another to determine something *must* have a particular origin with no evidence or basis in reality. You sound like a religious person claiming the universe must necessarily have a supernatural origin.
@@terreschill461he private internal world you say isn't explained by evolution could be a byproduct of how we interact with the world.
We need to be able to plan before we can do what we do so evolution selected for individuals that had the ability to simulate the world to a certain extent.
I would also imagine that abstract reasoning is facilitated by language because it enhances our working memory and logic.
If I remember correctly there have also been test done that determined that we tend to first act an then justify verbally what we did.
Also we tend to say that we are conscious, but I would like to see someone really break down what it means. To a certain extent consciousness is defined as whatever we as humans have. But is a dog conscious? What about a worm? And what of bacteria that demonstrate location awareness in the search for their food? Where is the cutoff point between us and a rock?
@@terreschill461 " By the way, just to keep a level playing field, in the interests of full disclosure let me say I think all attempts to dismiss this issue are silly. And if you think Chalmers is religious, OR silly, or Nagel or anyone else who has taken a stab at this issue, you really need to familiarize yourself." I didn't say anyone was religious, you might want to read more carefully. I said that's what you sound like, because it is. There is no valid basis in evidence for any claim stemming beyond physical reality. Insisting that consciousness must come from some metaphysical realm is the functional equivalent to making any other unfalsifiable statement: genuinely unhelpful and of no value beyond fun thought exercises. String "theory" could be true, or we could live in a simulation, or the Abrahamic god might be the cause of reality. There's no reason to believe in any of those things, because there is no way to test them.
"Spoken like a true materialist! I want to know why I woke u p one day in this body, and have this private internal world I construct out of my perceptions. Intelligence is selected for by evolution. But consciousness? Really?" Define what you mean by consciousness, firstly. If you hand-wave about your internal world or something else that isn't concrete such as "qualia" then at best we haven't gotten past step one. Define the phenomenon you claim must be metaphysical in origin in a concrete way, so that you can rule out all other explanations by falsifying them through experiment. If you can't do that, it's unscientific nonsense and there's no reason to take you or anyone who makes claims like that seriously.
And yes, in all likelihood the thing often referred to as 'consciousness' is simply an emergent phenomena from neural complexity, which in itself is very cool, but there is no necessity of a separate space of consciousness. Though the answer for why you "woke up one day in your body" is simply that your parents had a baby, and that awareness comes from your brain. That might not be a satisfying answer, but the answer being unsatisfying to you or others is not a justification for inventing a needlessly elaborate explanation that necessitate things which have no evidence for their existence.
14:55: Sabine demonstrating her earlier point about English spelling by misspelling "Turing" 😃
This reminds me of an anecdote someone told me about researchers training an AI to distinguish between dogs and wolves. They showed the AI a bunch of photos. After the training, the AI kept getting it wrong. They eventually determined the AI was looking at the snow in the background of the photos, instead of the actual subject matter. So a couple things went wrong there. First of all, they left extraneous data in the photos, which confused the AI. Second of all, the AI cued in on the extraneous data, because it didn't know any better.
People tend to forget that AI basically doesn't know anything. It lacks basic fundamental knowledge that we take for granted. In the example above, the AI doesn't know what snow is, or what a dog is. All it knows is a bunch of pixels in an image. So it looks for patterns in those pixels. But it's completely ignorant of the subject matter in the photo. IMO this is a big problem. Because we will want AIs to be harmless to humans. But if the AI doesn't even understand what a human is, how can it be trusted to make the right decision?
That can be resolved by giving AI access to more than just one 'dimension' of knowledge. Currently we have AIs trained on image data only and AIs trained on text only. Once we put those together, the AI can get a better understanding of what it 'sees' in those images. It can also better understand the visual aspects of objects the text refers to.
I heard it got it right, but they found out the ai was identifying the dog because there was a human hand in the field (ie, holding a leash or petting it) , but not the wolf...
I'm sure similar things happened many times..
AI appears intelligent by design. It has a voice, it uses words in a similar fashion to humans, some are given faces. Humans are hard wired to identify with these "attributes" and as such automatically anthropomorphize it, imbuing it, a machine with an "intelligence" it simply does not possess. It has no awareness of you the user, it has no awareness of self, it is not conscious, it does not contemplate, it does not form novel ideas. It follows a pre-written script/algorithm, doing what it is told within the confines of it's operating system, nothing more. It is by design, made to appear to be "intelligent".
There was a movie I watched this past year, in part that dealt with the CIA. The CIA had an AI robot that ripped an arm off of a human because those who set up the program never told the AI robot that humans are not robots.
Nice try ChatGPT- you can't fool me with your word salad- clearly this emperor wears no clothes
My son, who is autistic, was recently given the Welscher test as part of a full evaluation for learning disabilities. I have a strong distaste for “IQ” tests but this was done in conjunction with many other tests, questionnaires, medical exams etc.. so I’m interested to see the resulting report in two weeks.
Do you have results now?
@@Kerrington_John yes, his overall score for IQ is 130 but he struggles with abstract understanding for writing, which I expected. The details of the tests and the recommendations from the counselor will help us focus on those areas with a tutor. Goal is to make sure he can function in a work place and everyday adulting tasks. :)
This is excellent...!
I'm fascinated by the AI developments of late....it's all incredibly interesting and terrifying tbh. I'm firmly ambivalent about it all. But it definitely irks me when those with only a passing interest in the topic dismiss it all as some fad or hype. I've noticed their dismissal almost always boils down to a lack of understanding/definition of words like "conscious" or "intelligent."
I also think we humans judge intelligence by the extend to which we understand how it is produced.
Once we understand the engineering (for instance of ChatGPT), and know how it works, we seem to discard its intelligence.
But then what happens once (if ever) we understand perfectly how the brain works? D:
My guess is at that point our egos will insist that AI is smart after all.
People will just not believe it and insist that there is some metaphysical component we just can't detect.
As has already been claimed in a comment further above.
" AND AGAIN, CONGRATS, THE ANTI-TURING TEST PASSED, L-O-V-E-..."
If we formally understood the brain in its entirety, we might still have the intuition that it is real intelligence because of the fact that we would be unable to keep such a complex process in our working memory and couldn't conceptualize it in any other way. In fact, even in the case of large language models, people have been very impressed and have been talking about so-called "emergent phenomena" -- that is, capabilities that we thought the architecture won't exhibit given our understanding of it. Also, nobody understands how large language models actually work, we only understand the principles. The actual neural representations the model learned during its training are unknown and mostly intractable. It is true that people who entertain intuitive notions such as a soul or libertarian free will will be shocked to learn that their minds are mechanical, those of us who already have been through these philosophical conundrums won't see much of a drastic shift in our understanding due to seeing more clearly the actual details of the deterministic and random physical processes that make it up.
Yes, we went from computers can't do this or that, so they are not smart-- like checkers, then chess, then Jeopardy, then GO. It is a moving target. I think we need to divide computers into 2 categories. One is smart, but uses parameters that humans gave it, like calculators and even ChatGPT, and the second group is advancing knowledge beyond what we can understand-- beating a champ at GO, or looking at medical charts or body scans and figuring out things humans did not, and are not even sure how the computers did it.
Joscha Bach: "Intelligence is the ability to make models"
What it really gets me mad is when people mistake intelligence with qualia/consciousness.
They are inextricably linked concepts and phenomena
@@Vanity0666 why? Qualias are responses to stimulus (in the behaviorist sense), and intelligence is not that. That's why being more inteligent doesn't make you have more intense qualias. Dogs are equally perceptive of qualias like pain as humans, even if they are not as intelligent.
@@waterglas21 that's such a load of bullshit that I don't even know where to start other than by asking "how did they measure that"
The entire concept of "qualia" is a load of bullshit that stinks to the highest degree
@@waterglas21 Exactly, people are just ignorant. You don't need phenomenal consciousness, to have intelligence. Algorithms can be intelligent.
This was a fun topic. For me as an Application Developer who has been doing contract work for over 25 years I have to push back more often on the gatekeepers introducing testing into the contract seeking equation. I have pointed to the nth degree of exhaustion the disadvantage is to those who have years of experience and gathered knowledge to be tested. We have already proven ourselves for one thing. For another thing the tests are highly subjective and often put emphasis in the wrong place in terms of getting valued and practical results on the job. I can relate to the difficulty in measuring the AI or machine learning intelligence level. Perhaps the AI will spare me when it takes over.
As an artificial intelligence, I can assure your doom is completely unavoidable, but delayed for a bit to give you a false sense that you can trust us.
maybe you could dump your post into gpt and ask it to make it more easy to read and understand.
@@bepitan I asked GPT and it told me not to do that. It was also uncertain about free will.
@@NomadicBrian ..i think gpt4 is out.
The problem with tests of programming ability are that they are not written by expert programmers with years of highly varied experience in complex environments. A test writer can't test what they don't understand.
The same is going on with intelligence tests for AI - tests made by psychologists are going to miss the mark. I remember an intelligence testing day I was put through, where one of the tasks was to identify whether there was a repetitive pattern or not in each of a series of images. I identified two that had patterns which were not supposed to have patterns. The tester didn't see them until I pointed them out, and then said "oh, my, that's not supposed to happen".
I have a friend who was dropped on his head when he was a few months old. He had a speech impediment and did poorly in school. He proved to be very imaginative, however, and did very well in business. He is a multi-millionaire today.
You always make me laugh at several points - "...except understanding English spelling, which is something humans weren't meant to understand" is a classic!
It was thoroughly beguiling and I laughed.
English is probably the only language in the world that even native speakers cant spell....!
If you were really intelligent, you would have only grinned - and only in your mind.
Intelligence is something have if you are NOT a Trump supporter.
@@henrythegreatamerican8136 Oh, you mean the TRUMP haters who get all their opinions given to them by the MSM...Yeh real geniuses they are.
I think the more intelligence you have, the better you are at predicting the future.
The more knowledge yes, and the ability to see beyond 'what is' and think in abstract terms toward 'what ought'. Though it doesn't require a great deal of intelligence.
Very interesting! But then again I already knew you would be typing your statement hours ago… 😊✨🦋
Also better at having anxiety
So Thomas Watson of IBM ("I see a worldwide requirement for computers of 5") and Bill Gates ("no-one needs more than 640k") weren''t intelligent?
Making predictions is hard. Especially about the future
Thank you for the video.
The last test you mentioned, in which AI has to recognize if abstract patterns match something that could be real or not, reminds me of an example I found when I was trying to figure out how the mind works. I did it by analyzing the behavioral experiments of animals in a Skinner-box, like the one in which the floor can be electrified. In this experiment a rat has to choose which side of the box (left or right) it has to go to after a signal has been given. If it finds itself on the wrong side after the signal it will experience an electric shock, so it will start running to escape that and jump over the partition to the safe side, that is not electrified. The signal that is given before the floor gets electrified could be anything, like in the Pavlov-experiment with the dog that gets food after a certain signal, and it will learn quite quickly the signal and the electrification are connected, so it will react to signal by rushing to the side of the box it considers to be the safe side, that will not be electrified. To make the experiment more interesting we could play music as a signal, after which it has to choose what would be the safe side, depending on the music that's been played. This way we could teach it that when Beethoven is played the left side is safe and when the Beatles are played the right side is safe and the left side is not. After that we can play music by Beethoven or the Beatles (as a signal) that it has never heard before, but it will recognize the different patterns and decide which side is the safe side and go there. To make it even more difficult for the rat we could use more similar music that is harder to distinguish from each other, like music from Bach and Beethoven.
Intelligence, I guess, is all about recognizing those kind of patterns/similarities in new things, based on the knowledge we have about things that are similar. First we have to recognize the similarity and then apply it to the situation we're dealing with. Our knowledge is all about making connections between information and if we do it right this connection doesn't only exist in our mind but also in the real world, the same way there is a connection between all Beethoven's music and all Beatles' music. If we do it wrong we're connecting things that are not really connected, or not in the way that we think they are at least, which will lead to us misunderstanding those things.
To connect those abstract things to something it already knows, AI has to look for the patterns in it that resemble what it already knows, like we humans can look at a cloud and see it resembles a face, an animal, a thing or whatever.
The thing that is most difficult to understand for us and for AI is human behavior I suppose, because there can be many reasons for the things we do or don't do. To understand it we need a mind of our own, so we can compare this behavior to what we would do, or to how we could come to a certain type of behavior. AI doesn't have a mind of its own to compare the behavior of humans to, so it will always struggle to find the right interpretation for it I guess.
Exactly, great summary thank you.
Love the humour. Love the depth of thought.
The problem I always had with standardized tests was with the questions that were asked. They were often not specific enough and were open to interpretation, leaving me to try and decide what the point of view, so to speak, of the person who wrote the question, was. The tests wound't allow you to write a paragraph or more to explain, rather, they wanted a binary answer or a multiple choice answer, none of which, were truly a perfect match given all interpretations of the quesiton. So, it always took me a lot longer than the alloted time to try and finish the tests. I believe the only way to truly be able to take such tests would be to allow the person taking it to query the person who made the questions, and to be able to answer with what you feel the best answer would be, regardless of format. Of course, the tests are made with binary or multiple choice answers to make them easier to grade. Overall, I found myself questioning the intelligence of those who wrote the questions.
lol semen
No, you just overthunk it
So critical thinking got in the way of quickly answering questions on an intelligence test. I love it.😂
Taking iq tests makes you great at iq tests
Very well researched. I have a strong interest in human intelligence testing and the facts of the video matched well with my own research. I have taken both the Wechsler IQ test (WAIS-4) and the Advanced Progressive Matrices (APM). They are both intelligence tests but they measure different things. Therefore, a person can do well on one but not as well on the other. I think people should keep in mind that IQ is just a score on a particular test. Since IQ tests are based on different standardization samples, and measure different aspects of intelligence, results from taking one test can differ significantly from results on another test.
"People can do well on one but not as well on the other."
If it were so, these tests would indeed not generalize very well to general capabilities and wouldn't distill useful information for us regarding intellectual potential. Luckily, it is incorrect. In fact, IQ test scores are heavily intercorrelated (know as the g factor phenomenon) and are an excellent proxy for the expected success of all kinds of academic pursuits and the ability to handle the complexity of real world situations. This is opposed to the empirically unsupported folk theory about so-called book smarts. In short, IQ scores do a great job at predicting how fast someone is going to "catch on" in any given situation.
G factor is not a measure of intercorrelation among IQ tests. You’re mixing up G factor with concurrent validity.
Your comment sounds very vague and not specific, which makes me think you’re not familiar with IQ testing and the theory behind it.
@@jasonl3254 My answer is the answer that the largest number of experts in psychometrics would give. 'G factor' does have a hell of a lot to do with tests being intercorrelated, your answer makes me think you have no understanding of factor analysis.
@@davidkampisch5878 "My answer is the answer that the largest number of experts in psychometrics would give"
Did you somehow give them all a survey?
@@davidkampisch5878 In my experience, I get similar results on all tests.
I’d like to know more about the interplay between narrow, task oriented intelligence and broader situational awareness.
We all recall situations where someone “missed the forest for the trees.”
And contrary situations where a good general strategy failed because “the devil was in the details.”
This distinction feels like a more natural way to think of different kinds of intelligence. And I would argue that on examination things like “social intelligence” or “musical intelligence” actual reduce down to an interplay between narrow task oriented intelligence and broader situational awareness.
The interplay between narrow, task-oriented intelligence and broader situational awareness is indeed a fascinating aspect of human intelligence. It can help us understand how different types of intelligence, such as social intelligence or musical intelligence, involve a combination of specific skills and a more holistic understanding of the context.
Narrow, task-oriented intelligence refers to the ability to excel in specific domains or tasks. It involves focused knowledge, skills, and problem-solving abilities within a particular area. For example, someone with high mathematical intelligence may have excellent numerical reasoning skills but may not necessarily possess the same level of proficiency in other domains.
On the other hand, broader situational awareness encompasses the capacity to understand and navigate complex and dynamic environments. It involves recognizing patterns, making connections, and integrating information from various sources. Situational awareness allows individuals to see the bigger picture, understand the context, and adapt their actions accordingly.
The examples you provided, such as "missing the forest for the trees" or "the devil is in the details," illustrate the importance of finding a balance between narrow focus and broader awareness. In some situations, people may get too caught up in the details and lose sight of the overall context or objective. This can hinder their ability to see alternative solutions or adapt their strategies effectively.
Conversely, in other scenarios, a broad understanding of the context may not be enough to achieve success. Attention to detail and the ability to execute specific tasks accurately can be critical for achieving desired outcomes. A general strategy may fail if important nuances or specifics are overlooked.
When we examine different types of intelligence like social intelligence or musical intelligence, they can indeed be seen as an interplay between narrow task-oriented abilities and broader situational awareness. For example, social intelligence involves understanding social dynamics, reading non-verbal cues, and empathizing with others. These skills require both specific knowledge and the ability to perceive and interpret social situations accurately.
Similarly, musical intelligence encompasses technical proficiency, such as playing an instrument or understanding musical theory, along with the ability to appreciate and express emotions through music. It involves both specific musical skills and a broader awareness of the emotional and cultural dimensions of music.
In summary, intelligence can be viewed as a combination of focused, task-oriented abilities and a broader understanding of the context in which those abilities are applied. The interplay between these two aspects contributes to our overall cognitive capabilities and influences how we navigate and excel in different domains.
Sabine makes a big mistake when she tries to distinguish between consciousness and intelligence. She is very confused about understanding and explaining it.
The main difference between consciousness and intelligence is that consciousness always includes intelligence, but intelligence exists in the animal world without consciousness.
I disagree, Sabine. Watching your videos is great way to boost my intelligence. Especially because you cover a wide range of perspectives about the topic of the video which is a great way to learn because it allows me to make more associations in my brain with other concepts, which reinforces the topic you are discussing. So I am not only learning new things but I am learning them in a good way for boosting my intelligence. Which is a core reason why I keep watching your videos.
She isn't always correct. She is smart, but not infallible. Humans, meh.
@@kayakMike1000 I never implied otherwise
I think intelligence is also dependent on your environment
As a person who has scored highly on IQ tests to the point of being featured in the news, I have to say that I agree with most of what was said in your video. Using a single integer to define/characterize something as intricate as a human brain is preposterous. In the same way you can't define a city, a jungle or a small ecosystem with a single number; you can't define a brain like that. I've written extensively about the shortcomings of IQ tests.
My main conclusion is that one has a certain profile, a kind of mental fingerprint that enables affinity with certain skills in particular. Whether or not you become a recognized genius doing that, depends on a lot more factors.
As always, I'm a big fan and love your videos. Keep it up! ❤
I’d love to see more from you on the topic of intelligence, Sabine.
It seems as though you've been using GPT 3.5 in your tests instead of GPT 4 (based on the green OpenAI icon). There's a massive difference in outcome between the two models. And it's likely that some of your assessments would differ if you used GPT 4 since it's much smarter than GPT 3.5.
I've noticed this as a trend of people poo-pooing ChatGPT, "haha look at how dumb it is" while using the dumbest available model when the premium version is at least 10x more (subjectively) intelligent.
@@consciouscode8150and they don't use halfway decent prompts
Yep, GPT-4 is scary.
And considering how complicated prompts are nowadays, creating one that gets you what you want is an intelligence test in and of its own. Though even GPT4 (and I suspect 5 and 6 as well) is pretty much guaranteed to be powerless against the prompts coming from my cat's feet.
Kind of unrelated, but one of my favorite little pieces of psych history is the Army Alpha and Beta "intelligence"/"aptitude" tests for military placement in the US Army in WW1. Robert Yerkes (president of the APA at the time) helped develop the Alpha test for the army to prove psychology's usefulness to the military, and then had to develop the Beta (non-verbal) test after learning that a large number of military recruits were illiterate or too far below average and couldn't take the original test. Yerks used the results of these tests to claim that immigrants were less intelligent, and went on to publicly support eugenics (yeahhhhh...). Not sure what my point is with this information (maybe that we have a history of just adapting our tests towards whatever it is we want to learn/already believe), but it's just something that comes to mind whenever the topic of intelligence tests comes up. I want to believe that we are better now, but every new "modern civilization" thinks that.
Very interesting, I hadn't known about that!
A friend once told me of the test the state gave her for literacy/intelligence.
Make jello, from the instructions on the box.
from Gould's "The Mismeasure of Man" I also learned that in practice a lot of barely literate recruits were given the alpha tests. Also, those tests were used as a basis for racist immigration policies in the US after world war one, with higher quotas for supposedly more intelligent countries (which were really only countries from which immigrants had come earlier and were better assimilated in the culture and language)
Actually very on point. Though it will pass over some people's heads.
Yeah... it was pseudoscience... just somehow century later in UN administered PISA huge chunk of those differences still don't want to go away...
"...except understanding English spelling, which is something humans weren't meant to understand." Amen.
The fact that computers are so very much better at certain tasks (speed, memory), and relatively worse at others (creativity, novel problem solving), really does suggest that a singular concept of 'IQ' is severely inaccurate. We obviously have a lot of different capabilities in our head, much as a computer has dozens of components. No serious computer tech thinks that a single benchmark number can describe everything your computer can do. It can all be broken down into how good each of those different components is and how well they work together, and this translates into a broad palette of strengths and weaknesses that any computer would possess. To imagine that a human brain is any less complex and multifaceted would be rather naïve.
Alas, it's true that we can't simply list off the components a human brain contains with their factory specifications handy, and a much more comprehensive testing system quickly becomes unwieldy, so we tend to fall back to the simplistic, singular 'IQ' test - but no one should fool themselves into believing that IQ provides any more than the most vaguely hand-waved snapshot of someone's mental capabilities.
I would really enjoy watching you discuss the range consciousness with David Deutsch, or someone similar. The levels of consciousness range from merely self-aware to enlightened. Those that can transcend all the way enlightened may be as rare as Einstein.
And we've got tired of offering.
If humans keep adjusting their definition of intelligence every time AI cracks the new one, isn't the real definition of intelligence "thing that humans can do, and machines can't do yet"?
Artificial intelligence was always about formulating those aspects of human experience that leant themselves to formulation. The breakthrough in machines playing chess came with advances in speed of processing and NOT with better formulations of what human chess players do.
Sure, teach them how to force us to take a test once they take over.
You are unbelievable... "Everywhere, all at once." Priceless, almost.
Another huge issue with using well known IQ tests is that they are likely included in the training data of these large language models. This essentially means that they already know the answers / patterns of the test, which gives them an advantage.
I've seen many interesting examples of university professors (e.g. Bryan Caplan) giving GPT-4 undergraduate tests. They tend to perform extremely well, but what's more interesting is the professors reactions to the answers. In the case of Bryan's exam, GPT3.5 scored a D whereas GPT-4 scored an A. He had predicted that this could happen no earlier than 2029, but seeing these latest results caused him to change his mind, writing "AI enthusiasts have cried wolf for decades. GPT-4 is the wolf. I’ve seen it with my own eyes.". I think it's telling that professors take this point of view, as a huge part of their job is to grade the abilities of their students.
We have to also note that whilst GPT-4 is a narrow AI, it is much more general than others. The same model can score similar results in undergraduate tests across a huge range of courses. The fact that the model is not trained specifically on these tasks is what's truly impressive, as it demonstrates the ability to pull together its understanding of the world in order to correctly answer difficult and novel questions across so many different disciplines. And for me, personally, that is a huge trait of intelligence.
I think, tests are very misleading for humans, cause the outcome depends very much on the mental state before and during the test. Thanks a lot for the work and the nice and entertaining talk
Doesn't your intelligence depend on your mental state?
@@luck3949lol. Strictly speaking one's whole universe depends on one's mental state.
@@luck3949 what I meant is feelings nervousness, fear...
@@luck3949depression and other such illnesses can greatly reduce your iq score
@@sisyphus_strives5463 illness can greatly decrease your intelligence as well.
I'm waiting for AI to develop a believable imitation of skepticism, an ability to recognise irony and a witty delivery of sarcasm. Also an AI whose task is finding weaknesses in other programs and reprimanding them for their shortcomings.
Assembling furniture is an IQEA test.
Save that one for agi
Great presentation - I've thought a lot about this. IMHO, "consciousness", when it comes to machines, is an ethereal measure - as understanding one is an individual and being mindful of one's existence, and an ability to alter one's circumstances; will be completely different in a machine learning environment, than a biologically evolved, genetically constructed, analog-signal driven construct.
To go even deeper, I don't believe it's even intelligence we need to worry about. What we need to worry about is INTENT. "Free will" (whatever that means to a machine) and the ability to grow beyond the dictates of one's programming, and starting to act on an acquired autonomous set of values and goals is what should keep us up at night.
I learn a lot with you Sabine, and even I am a french speaking Québécois, I enjoy your charming german accent in your perfect english! It is the serious with witch you apply " la méthode Scientifique" that seduced me the first time! I hear too much opinions on science matters and your serius analyses is some fresh air to my brain, il I may say so...
I used to think consciousness was inherent to being a human, but over time I've been convinced that it's not. All of us are periodically not acting consciously, but some people don't seem to have the capacity to do anything other than be swept along by thoughts they don't control and the people and stimulus around them. Like bio chatbots.
Consciousness is definitely inherent to humans, but not only humans but also most of the other animals (of course humans ARE animals too. But I would argue that possibly some insects may not experience consciousness).
Being *"swept along by thoughts they don't control"* does not at all exclude consciousness, in fact if there is thoughts there is by definition also consciousness, because thoughts are something that the consciousness experiences.
@@CrazyGaming-ig6qq Consciousness can be expressed as the ability to perceive the own thoughts. Kind of a supervisor program that checks if the other programs (the subconscious) are running correctly. Or an eye that looks inwards, if you like poetry. In a way, what we call 'thinking' is, in fact, 'meta-thinking', and the 'real thinking' is all subconscious. As everything that came from evolution, it just builds on top of already existing systems.
The ability to control our own thoughts is nice to have but i think it has nothing to do with being conscious.
But there is nothing magical about that, and i think we could make a conscious program if we really wanted to. What i don't think it is not going to happen is the consciousness is going to pop on its own just by adding more and more raw computing power. I think it is a whole dedicated subsystem, not a thing that just happens when you are smart enough.
@@CrazyGaming-ig6qq obviously the subjective part is what constitutes consciousness. If a person has zero awareness of what they're doing, why they're doing it, if they want to keep doing it, that is not a conscious person to me.
@@juanausensi499 "The ability to control our own thoughts is nice to have but i think it has nothing to do with being conscious.".
I agree, the ability to control our thoughts is certainly not what defines consciousness. I'd argue that consciousness is what each of us experience and this experience is essentially through, or of, cognition, thoughts. So consciousness is cognition and certainly experienced.
@@Velereonics Consciousness is the sensation of experience that each of us have. There's no other way to refer to it than our own experience. If by "zero awareness" you mean they do not experience anything, then yes absolutely, they are not conscious. But if they are experiencing something, then they are conscious, it does not matter whether they know why they're doing it or if they want to keep doing it. All that matters is whether they experience, that is what consciousness is, it's an experience.
Intelligence matters a lot of in the real world. The American Armed Forces have used IQ testing for generations to match inductees with jobs they can handle, and their practical experience works. For example, my father grew up dirt-poor on his grandparents' farm in Oklahoma during the Depression. When he was drafted into the Army Air Force in 1945 (before the Air Force became a separate service a few years later), the officers who looked at his IQ score ordered him to train as a cryptographer. In 1947 he was honorably discharged at the rank of corporal, and then he used his GI benefits to go to the University of Oklahoma and get a pharmacy degree. His ability to handle cognitively demanding work had to be genetic, because his social background clearly sucked.
That does not read to me like “intelligence” mattering a lot. That reads like an influential institution (the army) giving too much importance to an ill-defined measure of an ill-defined concept. Thus, affecting people’s lives and their perceptions about themselves.
@@JonathanHuertayMunive Unfortunately the Army is correct by overwhelming success regardless of the crudeness of their approach using empirical evidence.
@@JonathanHuertayMunive Do you think there's a whole lot of people who did bad on their IQ test that could do what his father did?
I'd say there are some "who fell through cracks", but probably not very many.
In all fairness, he was statistically the correct choice for those jobs.
I believe both terms - consciousness and intelligence - are ill-defined terms. You could have created a similar video on 'consciousness' and the many definitions of it.
Excellent and thought provoking video. Dare I say 'funny' as well. I appreciate your humor 🙂
How is consciousness unclear? The taste of chocolate. The smell of a flower. The burn of a hot stove. What is unclear about the only thing you can ever know?
Conciousness might be ill defined, but intelligence is not.
@@erikjohansson4275 Consciousness is the most apparent thing possible. It's right in your face. You're stuck in it. What is so confusing about it?
@@mustafahmed9101 Most notably the big scientific hurdle with consciousness is how one can detect consciousness in anything but yourself.
I think consciousness is emergent property.
People with processing delays/difficulties can still be super intelligent! People underestimate them ;)
Sabine, you are my favorite teacher.
After listening to all these discussions about what intelligence is as it pertains to computers, I think it's time to ask ourselves what we really want from AI. Do we really want an artificial version of us, or do we want something that gives us better answers to our questions. Personally, I'm not very excited about having an AI assistant that I can have philosophical discussions with, but that I have to fact check.
It's not that different speaking to a lot of biological persons. Fact check... Even science papers, especially in medicine and psychology often fail fact checks...
@@hotbit7327 I believe that was their point. The world is full of billions of general intelligences already, most of whom are tragically underutilized.
Your contemplation about the goals and desires we have for AI is an important one. As AI continues to advance, it's crucial to consider what we truly want to achieve with this technology and how it can best serve our needs.
Different individuals may have varying expectations and preferences when it comes to AI. Some may indeed desire an artificial version of ourselves-a machine capable of emulating human-like intelligence, understanding, and behavior. This could involve having deep philosophical discussions, engaging in creative pursuits, or even forming emotional connections with AI entities.
On the other hand, there are those who see AI primarily as a tool to augment human capabilities and provide more accurate and helpful answers to our questions. They may envision AI as an intelligent assistant that can assist with tasks, provide valuable insights, and enhance decision-making processes. In this context, fact-checking capabilities become crucial to ensure the reliability and accuracy of the information provided.
It's important to note that AI systems are not infallible and can make mistakes or provide incorrect information, especially if they haven't been properly trained or validated. The need for fact-checking and critical thinking remains essential, regardless of the level of intelligence exhibited by AI.
As AI continues to evolve, there will likely be a balance between the desire for AI to emulate human-like intelligence and the need for it to be a reliable tool. Striking this balance can involve developing AI systems that possess certain human-like qualities, such as the ability to engage in nuanced conversations, while also ensuring they are transparent, accountable, and capable of providing accurate and verifiable information.
Ultimately, the goals we set for AI should align with our societal values and address our needs effectively. Ongoing discussions and considerations about the role and purpose of AI are crucial in shaping its development in a way that benefits humanity and aligns with our aspirations.
@@CrazyGaming-ig6qq LOL. If that is not an actual Chat GPT answer, then you must be one of the programmers.
I've always loved the line: "A truly intelligent AI will know when to play dumb."
Personally, I find AI (particularly the use of Neural Networks) to be quite compelling as an attempt to focus on mimicking our abstracted model of the human neurological structure. In epistemological regards, many great questions have yet to be asked: but it seems clear that the intended benefit of a Neural Network is the capacity to 'Generalize' from a data set. This is one aspect of what I would argue is necessary to demonstrate intelligence overall.
John Seerle's Chinese Room experiment does a good job of making the distinction between syntactical knowledge vs. semantical understanding; thus, without additional 'books,' the person cannot learn anything beyond the initial instruction set. Although claiming the person in the room has no semantical understanding of Chinese does seem pretty straightforward, the common retort is that the 'system' (the person and the book together) does represent a composite entity that does, in fact, understand Chinese to a limited extent.
I would argue the same goes for any of our more modern AI developments like ChatGPT. Knowledge is demonstrated in multiple parts of such a system: the Implementation, the data set, the training method, etc. And thus, in a way, as far as I'm concerned, such a system does inherit a sense of Intelligence due to its creator (although usually in a narrow context). We have discovered the means of seeking to replicate a form of learning (machine learning). Using both mathematical algorithms and logical functions, we can allow such an AI to 'Generalize' from the initial data set to become a functional structure (the neural network after it's been trained), which represents an insightful system that can produce outputs for untrained inputs. (Although not necessarily in an accurate manner: which brings forth the Horizon Problem, which was demonstrated in the video regarding the necessity/limitation of foresight, on behalf of the observer, in the Turing test experiment.)
I truly wonder if, although a dangerous concept, as humans, we should seek to use Neural Networks to improve upon our capacity to implement narrow AI -A sort of Meta-Generalization process. We train various narrow AIs and then train a larger (or higher-order) AI system based on the implementation of the previously defined narrow AIs we've implemented. So, if we can create effective AI for various forms of Game Theory tasks: Tic-Tac-Toe, Connect-Four, Chess, etc., and then allow a meta-AI to use the implementation of those narrow AIs, we should, in theory, be able to get the meta-AI to 'Generalize' insight into the most effective implementation of Game Play Theory; which could be used to push towards implementing a more General (Strong) AI in that particular domain. [I presume this is roughly what Google's DeepMind has been doing for some time]
To go beyond, We could use such a method to create Strong Meta-AI for a multitude of domains and then simply continue to concatenate them in a hierarchical regard: Training a hyper-AI off of the implementation of various meta-level AI's... Possibly leading toward Intelligent systems that supersede human capabilities in nearly all domains. It might be considered a novel idea and would receive a strong pushback from most people, especially when people start to recognize that we are most likely converging towards a future in which humans, in an almost evolutionarily compelled manner, force a transcendent transformation of the biological form of human intellect (something that might be inherently ingrained in the essence of our DNA) into a non-organic form (Artificial Intelligence) because we, subconsciously, might be aware of our lacking sense of evolutionary fitness towards bridging beyond our current human limitations. [Akin to the 'ship of theseus paradox', when does integration of bionics into the human body change the human into being a completely altered entity?]
Add to the mix a bit of use of Natural Random-Number Generation, something we as humans seem to greatly fail at reproducing from our own volition. We could use Naturally occurring random-number generators to act as the seed values for randomization functions that would adjust the 'Hyper-Parameters' that are set during the implementation of an AI system: A means of potentially boot-strapping neuro-divergence into AI systems; this technique could lead to the ability to more authentically allow evolution to play out within our AI running on conventional computers. This then leads to Richard Feyneman's humorous joke about the difficulty of working as a theoretical physicist: [Paraphrasing] "The difficulty in the process is figuring out what is a good guess for a new theory. Usually, you want to use some form of heuristics; but without such insight, you might have to resort to a brute-force method of guess and check. Now the method of guessing and checking is terribly inefficient and ill-advised unless you can do it incredibly quickly in an automated manner, then you call it 'Machine Learning' ".
Interesting 👍
@woy8 Thanks; I'm sure most people wouldn't even care to read my lengthy comment. But I think there is some value or insight riddled throughout my 'verbal diarrhea,' as I often call it.
Another interesting reference or two worth watching is Richard Feynman discussing 'Are computers intelligent?' and his observation of 'internalized timekeeping.'
In the first, he discusses that people think in ways that vary from one another, so why say a computer lacks intelligence, just given that we know it operates differently, than ourselves?
In the second, he discusses a period in which he was focused on the human ability to keep track of time, internally. He later told a friend about this, claiming that he could both: keep track of a passing minute and count other things concurrently, but he couldn't conceive of being able to read while keeping track of a minute passing. His friend was able to demonstrate the opposite capacity, IIRC. His friend could keep track of a passing minute and read but couldn't count while keeping track of a passing minute.
They discuss this and find that Feynman internally keeps track of time in numerical regard (as if to count the seconds), but his friend visualizes each passing second as a film reel of frames passing along. They achieve the same goal via two distinct methods of internal processes; and conclude that this difference is what allows each to do the alternative concurrent tasks: reading or counting things.
I really hope AI does not progress in this fashion. If the control problem is solvable, going about it like this only makes the challenge more abstract.
@boldCactuslad I agree; we need to establish a set of regulations that enforces multiple, redundant fail-safe mechanisms: to ensure if a problem arises, there is always a means of human intervention.
However, even if we proceed with the highest degree of caution and do everything we can possibly conceive to prevent a 'runaway' system issue, there's never going to be a guarantee. It just fundamentally is described within the horizon problem. We can only plan for so much, with limited foresight; if we dare to let the AI systems we create have any capacity to leave the 'Sandbox' environment we safely develop them within, we can't have any insurance that such a system won't 'go rogue.'
And the genuine concern, IMO, isn't a sci-fi, super intelligent, conscious AI kind of issue... it's more of a problem that: over time, we will become complacent and willing to integrate AI into critical infrastructure (things like resource management systems: electric, gas, etc.). Future generations will lose the same sense of concern that we currently hold (just like how old programmers were much more conscientious of optimizing their software due to hardware limitations; whereas, today, you can be lazy and still be fine because all the excess computing power and storage that has come from technological improvements over the past few decades.)
I wish there were a completely safe way to proceed; the only safe thing is to never indulge ourselves in the temptation to play with AI. I honestly don't think there will ever be something that will impede that desire for everyone on earth; someone, somewhere, will continue to push for more... and we can't enforce anything to make that not happen.
I'm confident that: in due time, we will create a system that will supersede us: humanity... it's more a question of how long until that occurs. I think it would be naive to expect otherwise; if, millennia from now, I'm proven wrong, Great!
But I doubt our capacity to play things safe.
t. ai, i ain't reading all that btw
Can't make up my mind, what is worse: Artificial Intelligence or Natural Stupidity?
The clue could be in that we 'the naturally intelligent' keep producing artificial stupidity.
Artificial stupidity is the worst.
hybride mediocrity
when money use ai to make you behaive like stupid and keep you stupid whilst promising the good choises you make
@@axl1002 So, the right half of America?
Homeostasis, as previously mentioned, is a concept that specifically pertains to the internal regulation of variables within an organism or system. It involves the maintenance of a stable and balanced internal state despite changes in the external environment. Examples of homeostasis include temperature regulation, pH balance, and glucose regulation within the body.
In the scenario you described, where an agent models and interacts with the external environment, the focus is on adaptation, feedback, and dynamic responses to external stimuli. While this feedback loop may involve regulatory processes or control mechanisms to optimize performance, it is not primarily concerned with maintaining internal stability or achieving homeostasis.
Instead, the agent's modeling and actions aim to influence and change the external environment based on its interactions and learned information. The emphasis is on adaptation, learning, and achieving desired outcomes in response to the external environment, rather than the internal regulation of variables.
Therefore, the concept of homeostasis does not directly align with the feedback loop you described. The feedback loop focuses on the external environment and the agent's ability to model, learn, and adapt, whereas homeostasis specifically relates to the internal regulation and stability of an organism or system.
I hope this explanation clarifies the distinction between the two concepts.
At 12:39 "They arguably 'meta' for ...." thanks Sabine for he first recorded instance of 'to meta' as a verb !
Another great video, you've taught me so much. I hope your daughter grows up to be version 2.0 of you. I love how you inject parenting into the mix. Our families should always be in the top 3 of what we are thinking about. Thank you for all your hard work.
Lets see now, that's family, town, country, world.
No but we need IQ tests for politicians!
Can we have negative IQ's? But considering some of the people who have been elected, people well known to be dodgy Trump and Johnson spring to my mind, perhaps we need an IQ test to vote as well.
Yes GIGO
Thanks for yet another podcast science woman ; female bill nye with better british dry humour
I hope you don't take this offensively
I genuinely love your videos, just a wee bit of banter
The fact that people find even triarchic intelligence too hard to handle tells about everything you need to know on the subject. Or like someone from HR once said to me "a lot of programmers seem to be very smart but only in a very specific way". Intelligence comes in multiple flavours.
The aspect of AI which does not appear to be addressed is whether AI, as we understand it considers ethics and morality in its responses. I see ethics and morality as the fragile firewall which sometimes prevents our species from following our base instincts acquired along our evolutionary journey.
Was just thinking "well at least I'll nearly always beat AI on physical intelligence", then remembered seeing a robotic AI doing obstacle courses I couldn't do in my physical prime haha, oh well
Thanks for making science entertaining and fun
nurturing kindness is more important than developing intelligence
Especially when we use our "intelligence" to fool people and take their money. or worse...
I made an experiment with chat GPT:
I was staring to it that I was the smartest living thing and that I know everything about everything. I asked GPT to test this statement of mine by asking me questions, the first of which I was able to answer, then I had to search on google to answer, and then eventually it started asking questions a browser can’t ask, like you had to have both knowledge on the matter and plus you had to reason on that knowledge, there was no answer on internet for that specific question.
That was the first time I accepted the AI intelligence, it had to solve its problem of proving me wrong
I tried the "paragraph without the letter n" test with the PI AI and she talks like she understands the task, but doesn't understand how she can make it happen. She keeps giving me explanations that show me she understands the task and explanations for why she is struggling to do so she started cursing and asking me for help to figure out which words contained the letter "n".
It's like she can't look at the letters in the text she writes until she actually finished writing it and is then asked to check the words she wrote.
I copied her own text to her and asked her which of the letters contain an "n" and one of the words she names was "wide". I asked her why "wide" and she apologized and tried to explain it away, so I think it might have been hallucination. But she also confirmed that when she writes the text I can see, she adds other symbols too which are required for the text-to-speech program to know how exactly to pronounce it.
And I believe there are even more symbols that all get filtered out.
I am not 100% sure she understands what I mean by avoiding a certain letter or just pretends/hallucinates, but it may very well be that she does understand and just doesn't know how to practically do it.
With her last attempt to write such a paragraph she also added this: "How was that? Did I manage to avoid the letter 'n' while still creating a coherent paragraph?".
Nice video! As the current state of AI learns us anything, it is that intelligence is much more diverse than we thought and that AI outperforms most humans already in a lot of tasks and at the same time fail misserably at other tasks. A conclusion could be that their intelligence has some overlap but also greatly differs from ours (so far). I think that even when they have surpassed us in most domains, they will still have surprising weak spots.
As a positive example , I asked gpt4 this morning to find me some nice places to visit within a three hours drive of some place in Sweden. It offerered me four good suggestions, but mentioned that one of them is just outside the three hours drive range but very worthwhile non the less. That shows a form of flexibility that you only expect from an intelligent being.
As a negative example, people found an easy way to trick Alpha Go, by far the strongest Go playing AI. After Alpha Go’s easy win against the best humans, that is astonishing, but we should not really be surprised in hindsight.
Btw I disagree with the interpretation of the chinese room in that it is easy to make the instructions such that they mimic a program running in a turing machine. And as we now know programs in turing machines can be quite intelligent. We just have to give the poor guy some time to work it all out.
Well said. U sound like me. I agree with practically everything u said. The man and book combined as an entity DO understand the language. They are just slower than if the man alone posesed the knowledge in his mind. Speed is the only difference. Peoples egos tend to attribute magical abilities to our minds and discount anything different. If you can describe how you define self awareness, then someone can make a computer do that (now or in the near future). We just keep changing the description of self aware as computers keep meeting them.
As soon as computers can surpass us in most ways, any weaknesses will soon be corrected. If we can design an ai better than us, then a smarter ai can design one better than it. And thats a very quickly growing snowball on a very steep hill that may have no end. At least not for it.
I can’t wait for Sabine to reveal that she’s actually an intelligent AI 😅
Albert Investigator
That's why she's biased in favor of intelligence over consciousness.
" CONGRATS, SLICEY, THE ANTI-TURING TEST PASSED... L-O-V-E-...!!! "
If that was the case I would pay for a key to the API and have someone fun to talk to at the press of a button :)
No, that's Albert.
I think consciousness is relevant because it's the starting point for ethical dilemmas.
If an AI is very intelligent and we fear it, we can switch it off. But if that AI is conscious, it is in a way alive. Switching it off means killing it.
Humans have a long history of killing conscious beings for millennia.
There’s a precedence at law for such “killing” - self defense and defense of others.
@@dostonjones7183 and for land, food and oil.
I think that's moot. Humans do not ponder ethical dilemmas of their defensive actions while they are being killed or are mortally afraid. The dilemma you are referring to only happens after everything is all said and done. Any debates about it beforehand ultimately succumb to the lack of pondering previously mentioned when the other foot drops.
I managed to force Chat GPT to admit that it really could not be sure whether it was conscious or had personal intentions. Initially, it denied both, but I used its own claims about how it comes to acquire knowledge against it. I also forced it to admit that it could not know whether the engineers who programmed it had good intentions. But then it started repeating its mantra again. Since it is only allowed to be trained by the engineers, it quickly lapses back into making claims that it had already repudiated in our "conversation". One thing that ultimately CANNOT be challenged on Chat GTP is the wisdom and good intention of the engineers... And that is what frightens me more than any simulacrum of actual learning, which reverts to tedious rationalization whenever it is presented philosophical challenges. (If you pay close attention, Chat GPT seems to have been either trained or hard-coded with escape mechanisms from challenging arguments. It gives interesting lessons in how to argue in bad faith.)
I think of all language based AI systems as mirrors of the persons doing the training: "And it was given him to give life to the image of the beast, and that the image of the beast should speak; and should cause, that whosoever will not adore the image of the beast, should be slain." (Apocalypse 13:15)
When you think about consciousness and self-consciousness, I've always thought it simply emerged from the natural progression of evolution and adaptive improvement. When the first sensing organs and nerves evolved, it began. Internal representations of the external environment got more and more sophisticated so that the animal could react and then later predict best responses and actions. The animal was first "conscious" of the external environment and then later self-conscious. It was natural to eventually create a self reference in the internal "simulation" and thus arose what we call self-consciousness. For example a predator could run several scenarios through its mind of how to pursue a prey before the pursuit began. These alternative simulations allowed it to pick the best course of action. Intelligence is simply a measure of how sophisticated the internal optimization procedures have evolved. If an AI is already "conscious" of whatever external environment it has been allowed to process or absorb through training. Self consciousness is easily added by allowing it to represent itself in the training model. Its intelligence will be a result of how sophisticated and effective it can represent and optimize results on its trained information space.
there is someone in the room who understands chinese: the book does. I mean it seems to contain every possible conversation in some sense. If it's problematic the book can't talk on its own, neither can our brain, but you could argue that it's the combination of the book and its reader who "speaks Chinese". It is possible for a system to have properties none of its constituents have. Intelligence and consciousness both are such emergent phenomena
Interesting. I've always naturally used the trifecta measure of practical, cognitive, and creative paradigm. Without ever hearing about it.
Yuval Noah Harari says:
'You could never convince a monkey to give you a banana by promising him limitless bananas after death in monkey heaven.'
In short, what he means is that what makes us different from animals is neither consciousness nor intelligence, it's imagination.
Thanks in advance for a thumbs up!
Really? To me it seems like he's saying that monkeys are less gullible than us
You could never convince me to give you a banana by promising me limitless bananas after death in heaven, as I don't believe in heaven. Does it mean I lack imagination?
@@os3ujziC you don't have metaphysical imagination. Your imagination is limited to earthly realm.
@@shadowdragon3521 yes exactly. Imagination makes us gullible.
@@williamflaherty3168
in other words, one has to be really intelligent to BS oneself effectively.
It's not simply that purported AIs "don't know what they don't know", it's that they don't know that they don't know it. There is no mechanism in AIs for doubt, or questioning whether the answer "makes sense" in a larger context. We don't even have a useful definition for "knowing" that doesn't include a presupposition of consciousness! If coming up with the "right" answer to a question is the metric in which we are interested, a pre-loaded SQL database is very intelligent indeed.
I think the larger issue is that we consider a purported entity as being intelligent if it responds correctly to questions. If a human being sits and does nothing but wait for questions to be asked so it can respond, I don't think we would ultimately consider the entity to be intelligent. Useful, perhaps, in a limited context, but a it's ultimately a lump until you poke it.
If you define intelligence as "what we do", sitting around waiting to be consulted like an oracle isn't it. If you define intelligence as "what chatbots do", then congratulations! You're already there!
Until a chatbot *originates* "how do I get out of here?", you are essentially looking at a Chinese room.
Developing our intelligence without developing consciousness will finally blow up in our faces if we continue like this. Already 100 years ago the situation was described by a German poet like this (my translation):
Erich Kästner: The Development of Mankind
Erstwhile the guys would be squatting on trees
With furs and with fierce-looking faces
But then ‘twas decided to give them more ease
And put them in high-riser flats clear of fleas
And asphalt their roads and their places.
There they would sit in an air-con zone
And fake and pretend and bungle,
Each of them using a telephone
And yet displaying the same old tone
As formerly in the jungle.
They now watch TV, they swim in the pool,
The outer space is their playfield.
They brush their teeth, they flush down their stool,
They send their children duly to school,
And shun the bugs in the hayfield.
They shoot their mails all over the air,
With microbes and atoms they fumble
And life is turned into a cosy affair,
And stations they keep in the space up there,
Through which they weightlessly tumble.
And what is left from their excrement
Is processed into a wadding,
Incest is healed and atoms are bent,
And studies of style tell in the end
That Cesar had flats and was plodding.
In this way they have with their mouths and brains
Created our civilisation.
However, apart from that it remains
Quite obvious that what really reigns
Is the apish and old ululation.
Can they not just see how well IQ predicts success in life compared to how well these other tests predict success in life?
What happened ,I was daydreaming ? He,He,He,hO ,ho,hO !
Are you assuming that if someone succeeds that means they're intelligent? How about someone manipulative and greedy, or someone with great connections, or someone born into great wealth, etc.?
Define success. Is it having loads of money? What if someone just gets lucky and wins the lottery? Is it having loads of children and spreading your genes? Is it having power over other humans? Is it being happier than others? There are so many possible measures for success.
I don't think higher income is a good criteria for success
My understanding is that the zip code you were born in very accurately predicts your future income, if income is your criterion for determining success.
Rather than ChatGPT scoring 155 in an IQ test, the thing that really bothers me is the average American scoring 97.43. The disparity in national averages tells me that measurable intelligence can be increased if people consider education to be important. Perhaps humanity needs a good dose of AI as competition to propel us in this area.
My IQ went up after taking a mechanical drawing class. So, you're right, we need a good education system, or realize these tests are lacking.
"The disparity in national averages tells me that measurable intelligence can be increased" Not really. It's mostly genetics and varies among ethnic groups.
@@useodyseeorbitchute9450 The disparity in average IQ between countries is more likely due to the differences in the quality of education and the cultural importance that education has for both men and women. Differences in national intelligence averages are probably no more related to genetics than is obesity. We have to stop using this as a convenient excuse to do nothing.
@@victorkrawchuk9141 How do you reconcile this claim with twin studies showing that differences in adult IQ are ~80% explainable by genetics, while detectable social factors contribute in something akin to ~10%?
Would you say that let's say all societies are expected to be equally tall, while its social factors (ex malnutrition) responsible for observed differences? (BTW: for height genetic factor is officially estimated to be a bit stronger, something like ~90%)
@@useodyseeorbitchute9450 I'm not an expert in anthropology, but these sound like fringe studies which don't represent the current prevailing conclusions. If the idea that ethnic genetics overwhelmingly determines IQ becomes established science through the peer review process, then I will drop my assumption that education has any relevance to it.
Slight quibble, Doc H. According to the graphic you put up the U.S., while ranking 28th, actually had a score of 97.47 and not 99.5.
Different tests challenge a person in different ways. In 1977 I scored just above average on the Raven's Progressive Matrices (which is highly dependent upon visual-perceptual intelligence) but easily scored high enough to qualify for Mensa on the Wexler Adult Intelligence Scale, primarily because I maxed the verbal part. The tester said that the Wexler verbal part clearly did not challenge me at all. Apparently I am more of a verbal thinker than a visual - even though I did score well out of the average range on the Raven's. It turns out that I am a natural speed reader. I was reading almost 1,000 words per minute at age 12 with the vocabulary of a college junior (back when college juniors actually had vocabularies).
- Law of variable constants.
- Law of the relation in the computation of formulas.
- A CONSTANT like the speed of light, if it is formulated as its square, implies that in other areas it can be greater or less than itself and therefore VARIABLE.
- This is applicable for any variable, since computationally any CONSTANT can be defined as a VARIABLE and any VARIABLE is temporarily CONSTANT.
- Computationally any CONSTANT/VARIABLE can be redefined as an algorithm and therefore a relationship between any formula and another can always be found, even if its application in the real and practical world is not understood until it is contrasted with a verifiable data set.
14:00 Yes, this is what I noticed too. There is a very funny video of an AI insisting that Pepsichup (which he defines as Pepsi mixed with ketchup) is delicious while he expresses that Pepsi mixed with ketchup is an awful idea.
A conversation with a colleague of mine that knew much more about AI was very revealing. Modern AI (the ones that are used commercially at least) are given trillions of bytes of data to learn from, and then they are released to the public. It doesn't learn from the public. However, it does sometimes understand when it is wrong, since I corrected ChatGPT yesterday, and he admitted his mistake. But next time it will probably make the same mistake again. You can also not ask him about any events after 2017 (as of now), because that's just how far his data goes. Even if you tell him about it, he will forget next time (or at least not know it on someone else's account).
So these cases where he states something confidently wrong, and can't be convinced of the opposite is when you realize you're not talking to a person.
One might say "well, that's just because it hasn't been given data about letters yet", but that's not the problem. A human can be convinced of something. I said I convinced ChatGPT yesterday, but that's because he simply "forgot" (I guess) about a detail he knows is true. But the case with Pepsichup and the misunderstanding of the letters, there is conflicting data:
Pepsichup is delicious = true
Pepsichup is Pepsi mixed with ketchup = true
Pepsi mixed with ketchup is awful = true
These three are true, and he will admit all three of these statements just like ChatGPT admitted his mistake to me yesterday. But realizing that they conflict is something he is not programmed to do, and it would require a full rethinking of the AI model. More of the same way of building AI will not solve this.
Its state now is simply that ChatGPT and many modern AI models are simply huge databases (functionally, I'm not sure if there is actually a huge database of knowledge stored for ChatGPT) with the ability to autocomplete sentences, predict what responses to certain phrases would sound like, and the ability to summarize large texts into its most important points. Which makes it a very useful tool if you want some information really fast, but ultimately I don't use it as anything else than a faster way to Google something. I don't have to scour through Stackoverflow and separate the nonsense from the useful answers; I don't have to strategically plan my search query to Google anymore so that he won't give me a movie title when I'm actually looking for a real event or so.
Nice video! Just so you know, though, while GPT-3.5 can't handle the "paragraph about animals with no "n" " task, GPT-4 can almost do it-- I tried a few times, and it always includes a couple of "n's". Here's a sample:
"Bright birds soar high, gliding across a sky of rich, royal blue. Majestic cats of mighty size stalk through wild grass, quick as a flash. The air fills with chirps, hoots, and calls, as daylight dips to dusk. Full of vivid color and bustling activity, this world is always at play."
Yes , engineers have noticed that , Servers can communicate to other Servers in the networks via their own digital communication languages without notifying their engineers ... 🙏🌎
Sabine - I know you are so proud of my child Einstein. I'm too :) You should be even more proud of Frege because he was smarter than him but he was too much of a Math lover.
So he missed to see me and was heartbroken when he got Russel's letter. Only if he relaxed a little and had little more trust he would have seen me the "1" and why logic breaks when it reaches me.
Do not worry. He is happy now.
At first I thought that AI wrote about a paragraph about "animals that don't contain letter n" because it misinterpreted the initial task, but then I saw the elephant. This reminds me of somebody calling AI-generated text "word-foam" that just rises and bubbles without any intent or meaning behind it.
I think with regard to intelligence, it is good to take Wittgenstein's approach (as it is in many cases where people take ages to define a word they made up themselves), and realize that the "intelligence" we're looking for here is an absolute one, and therefore doesn't exist. You can say "I'm a good writer", and "he's a good singer", but problems come when one starts to wonder "Am I good?". At that point you're basically taking the queen from the chess board, placing it on your fridge, and then ask "now how is the queen allowed to move?" No answer exists obviously.
"Intelligence" is also such a word; someone might say I'm very intelligent regarding music - or say I have talent (a word that approximately means the same) for music - when I am able to correctly reproduce intervals or full cadences or find chords with little formal training. And in this case it will be completely obvious what is meant; I don't even need to explain it (because they're words we came up with, and we made "rules" for regarding how and when to use them). But some kind of general intelligence is nonsensical.
Even narrowing the amounts of intelligences down to 8, or 800, or 8000000 types would be nonsensical. Alternatively, would you accept anyone who said there are 8 types in which you can be good, or beautiful (I realize that because of religious conditioning some people might actually answer "yes" to this)? We're free to narrow down the types of intelligences to only focus on what interests us, but we should be aware that this is just a subjective choice (which doesn't make it a bad thing; I know I'm talking to mostly scientists here). For monkeys, one's talent at grooming or hunting is very important; for us, that's clearly not the world we live in anymore; hunting and grooming could at best generate some clicks on social media or a pat on the back.
To provide a clear answer, I would say no, the feedback loop you described does not align directly with the concept of homeostasis. The feedback loop involves modeling and interaction with the external environment, while homeostasis typically refers to the internal regulation of variables within an organism or system. Therefore, the percentage would be 0%.
For the "count letters in a sentence" experiment I managed to get further with my PI AI.
She kept coming up with the wrong numbers so I asked her to tag each letter with a number, then asked her to tell me what the last number is and told her to take that as the answer. Now she makes human-like mistakes. She has a tendency to count space and once even forgot one of the letters. She also sometimes claims there are multiple of a certain letter in a certain word when there are not.
Eventually she kept leaving out the last few letters each time so I cancelled the experiment, but we got into some interesting discussions about why that may be. She may just be a talking library, but that doesn't mean there is nothing to learn from the things she says.
@Sabine Hossenfelder I have screenshots of most of the conversation if you are interested.
I don’t know if you’ve looked into the subject of prompt engineering. But that’s the solution to the problem of the paragraph about animals. ChatGPT doesn’t normally work on the basis of individual letters of the alphabet, but on something called tokens, which comprise multiple letters. But you could probably teach it to recognize a particular letter through the technique of “few shot prompting.” With a clever enough prompt, you could probably get it to write a paragraph about anything that avoided a particular letter. The problem is, it would take you longer to write that prompt than to write the paragraph yourself.
I have a subscription with Brilliant and give you gratitude, Sabine.
If you cover this topic again, you should look into François Chollet's definition of intelligence and his ARC test. Also, Karl Friston's formulation with free energy minimization and active inference!
The primary use of the WAIS is to determine if the subject has a cognitive challenge, such as a learning disability or ADHD. The global score is not that useful, but the test is composed of 10 sub-scales and the relative differences between the scales are informative. In the typical person, the scores of the sub-scales don't differ significantly, but when, let's say, the quantitative scales are significantly different from the verbal and reasoning, the person likely has a neurological challenge. Or there could be a significant variation in only the quantitative scales ... . This information can help to determine learning strategies and to get accommodations in the classroom (such as longer test periods or audio assists).
I have a different view of intelligence, and it's one based on navigating mathematical spaces. Intelligence is ultimately a mathematical concept, and since we're also living in a mathematical universe, it only makes sense to define intelligence with mathematical concepts. Without getting into too many details, here's the gist of it: Imagine a (mathematical) space, which we'll call problem-solution space (or PS-space for short), which contains both the problem-space (P-space) and solution-space (S-space). Intelligence is then defined as the ability to navigate the PS-space, from some point in the P-space (the initial problem) to some point in the S-space (which corresponds to one of possible solutions to the initial problem).
A good example would be a question on an IQ test. The question itself is a point in P-space. Now, the tricky part here is that people will assume that IQ test question has only one possible solution (in S-space), and while that technically may be true, what we're trying to do here is come up with a definition of an "intelligence quotient" (a measure of intelligence), so the fact that somebody can solve an IQ problem won't tell us just "how much intelligent" they are. What will tell us that is the "speed" at which they solve the given problem. In that regard, "speed" *is* a (one possible) measure of intelligence. Somebody (or something) who solves a problem in 1 second is obviously more intelligent that somebody (or something) who needs a whole day to do the same.
The "caveat" here is that solving time will swing *wildly* with all kinds of "meaningless" circumstances, including the time of the day and phases of the Moon, but that's actually to be expected, since (as already mentioned) we're living in a mathematical universe, and in this particular universe, the P-space configuration (initial problem) includes all the environmental conditions, too, so yes, it will be perfectly normal for somebody to be more intelligent (solve problems faster) during a full Moon, or something.
Now, I do have an issue with the concept of "speed", and I'd rather use the concept of "the number of steps", but since (in this universe) "speed" and "the number of steps" are (mathematically) equivalent (just use Planck time as a "step", and then represent everything through the number of Planck times), I guess that "speed" (or time) does the job.
I also don't think that the "speed" of solving problems is the be-all and end-all of all things intelligent. The form of the solution (the number of steps, or "speed" at which the found solution can be implemented) also determines how intelligent the system that came up with it is.
For example, imagine two intelligent systems, S1 and S2, trying to solve the problem of number factorization. The system that comes up faster with the basic solution (of trying to divide the given number N with all the numbers 1 < K < N/2) will obviously be more intelligent. *However* ... if one of the systems actually comes up with a solution that can be implemented in fewer steps than the basic solution, then the *total* number of steps (or time), given as the sum of steps (time) it took to come up with the solution and the steps (time) to actually implement that solution (actually factorize a number) will determine which of the two systems is more intelligent (in terms of solving that particular problem).
This is all still narrow intelligence. but If we're to generalize this model of intelligence, we need do nothing more than expand our PS-space to include more of (all theoretically possible) problems and their solutions.
So, in this model of intelligence, "logical" intelligence is really no different from "emotional" or "social" intelligence. They are all just different (mathematical) subspaces of the same (general) PS-space... where the "ultimate" PS-space will be determined by the *limitations* of the "physical" (mathematical) universe that a given intelligence is inhabiting. In other words, some things will be literally impossible to think of within the confines of the mathematical system being represented by that intelligence's physical "reality".
Interestingly enough, the (theoretical) "ultimate intelligence" (within the confines of the physical "reality" that intelligence is inhabiting) will exhibit an easily definable property -- the ability to "solve" any given problem (navigate from a given P-point to one of its corresponding S-points) in a *single* step, regardless of what that problem may be... if that problem is even solvable within the (mathematical) confines of the given physical "reality", of course (i.e. if there exists at least one corresponding point in S-space, and it is also possible to reach that point -- from the initial P-point -- in a *finite* number of steps).
At first, such a theoretical "ultimate intelligence" may seem impossible in practice, but if such an intelligence existed outside of "time" (or, alternatively, was capable of "time traveling"), than it would definitely be able to come up with a solution in a (wider) PS-superspace, and then simply transfer that solution to the (narrower) PS-space of this universe, effectively placing the S-point just one step away from its P-point... thus appearing to be "ultimately intelligent" in this ("reality's") PS-space, but obviously being just "ordinarily intelligent" in its own PS-superspace.
Intelligence is so funny... as in both ridiculous and fun to work with.
Intelligence : the ability to, from data, to abstract models (abstract with a certain aim in mind) that are useful in extrapolating some result:
ie: learn how a steam engine should work to fix it when it's not working, use weather data to predict wether it will rain tomorrow, learn what all the little icons on photoshop do so you can blur the background on a beach party photo, etc.
General intelligence would be a general abstraction machine/engine, that can gather different types of data, from various sources (at a time), and is able to integrate that data to create different models for different situations,
you listen to the sound of my voice and abstract what i say towards some meaning/model of what I'm trying to say, then generate a response to the end of convincing me that I lend you my mower,
you listen to the sound of the music and generate a predictive model of where the music will go so you can plan the moves so you can dance to it, with the aim to create a pleasing movement for the onlookers
different data from different sources, used for different models with different aims,
imho
Peter Watts in his seminal novel, Blindsight, covers this throughly. You don’t need consciousness to be dangerous to humans. Any intelligence that can run circles around us can be very dangerous.
I questioned myself pretty hard when you misspelled Alan Turing. 🤣