I really enjoyed this conversation with François. Here's the outline: 0:00 - Introduction 5:04 - Early influence 6:23 - Language 12:50 - Thinking with mind maps 23:42 - Definition of intelligence 42:24 - GPT-3 53:07 - Semantic web 57:22 - Autonomous driving 1:09:30 - Tests of intelligence 1:13:59 - Tests of human intelligence 1:27:18 - IQ tests 1:35:59 - ARC Challenge 1:59:11 - Generalization 2:09:50 - Turing Test 2:20:44 - Hutter prize 2:27:44 - Meaning of life
“I’m not interested in psychometrics as a way to characterize 1 individual person. I think psychometrics is most useful as a statistical tool. It’s most useful at scale.” This man is a measure of intelligence himself.
Lex, Thank you for providing us layman with this extremely beautiful and captivating podcast. We all appreciate the time and effort you put into this project of yours. Life would truly be quite different without you!
“Intelligence is how efficiently you learn new things” “Measure of intelligence is ability to change” Being able to learn comes from ‘understanding structured operation logic” model between events. And being able to change comes from the ability to use causal inference on a perceived event against the model.
But this definition would mean that intelligence in humans is strongly tied to plasticity making children more intelligent than adults. However children struggle with many causal relationships that are easier to grasp as adults. This isn't because children are dumb but that this understanding of operational logic is actually learned. What you are describing is often referred to as "learning how to learn" but this is not the base of learning, just a skill that helps with applying other logical relationships in new contexts. An intelligent system would be capable of reaching this logical understanding but it does not need to start with operational logic a priori.
@@goclbert I think going back to Francois point you could argue that this isn't evidence that adults *are* actually smarter since it's just a matter of having more training data 🤣
I love the way these two people think completely differently. I've had conversations about the way I think compared to my significant other, and its surprising to see how both similar and different ours ways of thinking can be!
Noam Chimpsky the chimp and Koko the gorilla effectively debunked Chomsky's core assertion that language is the root of intelligence. Various other animal studies have also broadened our conception of intelligence, for example the toolmaking ability of corvids.
@@MattOGormanSmith instead of being the root I'd argue that language is an outcome of intelligence. That it is itself a tool for articulation of more complex and abstract thought that simple action and reflex can't represent efficiently.
I listened to this after consuming thc. I had several mental breakthroughs by applying the albeit hard for my intelligence level to decipher. Highly valuable information contained within. Lex you have an amazing mission to question the questions, That has helped sculpt my mind into one that is slightly more efficient than it was. More efficient at detecting bias. I still have many biases. But I am getting better at detecting them. Thankyou youtube for providing such valuable content. The world is both better and worse for the torrents of information literally at our fingertips. From one random guy on the internet. Choose the content you consume wisely. Turn the play speed up incrementally over time as you adjust. This adds profound gains in efficiency of absorption. The faster play speed becomes a second language. You will become philosophically bi-lingual. Other people will likely not understand 2.0x play speed yet it sounds normal when you become fluent. 3.0x is only 50% faster than 2.0x. My average level of intelligence. Or maybe the frequency perceptiveness of my ears, will reach a performance ceiling before others with more capacity. Neurolink has the ability to increase the speed of uptake. In a realm where consciousness consists of stored information combined with interpretation and adaptation. Learning the dataset is just as important as improvisational ability. Any efficiency increase should produce measurable results in all round performance. /consciousness.exe
I really like the way this man talks for some reason, like he has a fairly thick accent yet is more intelligible than a lot of native English speakers.
Thank you both lex and Francois. This was a great pod cast. Thank you for this genuine and honest podcast. I as a viewer appreciate the knowledge you two are able to explain in understanding ways. God bless u both for ur good knowledge on things
Francois explains things in detail and his psych approach has keypoints I find impressive. I have the same view on what the outcomes are in testing. Knowing and not expecting depending on the circumstance does impact the targeted goal. I thought this in my experience(s) and may I add, its an added challenge not being able to hit some goals on your own. Meaning mine, I seem to kick my own butt. You guys already know how to talk to each other. I am really going off the rail communicating.
This guy.. is intelligent. I believe he has solved what intelligence really is, and that’s amazing. Also he has great insights about thé shortcomings of current deep learning and of neuralink
Love your work, mate. You're an amazing speaker and an even more amazing interviewer (or vice versa 😃)! Currently chugging through your extensive backlog of interesting content! Love it and keep up the good work!
Lex, as I observe the world and people around me, I also determined primary measure of intelligence is the ability to adapt to change. This is driven by deeper layers of character/mind that I have identified.
What the hell? 18k views, 140 comments? These are too low, people dafuq. This shit is the most interesting informative discussion ever Edit- ... discussion ever today, as every podcast is amazing
Roger Penrose and Jeff Hawkins were already interviewed so it seems natural to me for Lex to complete the set of AGI experts by inviting David Deutsch.
Wonderful and very subtle ideas about consciousness and learning. Thank you. Some historic non-quotes: Descartes on logic: “Remember to always put des cartes before des horse.” and “C’est bien joué Monsieur Des Cartes.” Accordingly to Bertrand Russell, Descartes always wore his sword at his side, even when he was in his apartment. Thank you. William L. Ramseyer
I've always been curious, what is Francois' academic background? He's obviously a great engineer -- keras is a beautiful API and a huge success. Somehow he also has the ability and time to also work on deep questions of intelligence. Did he study physics? computer science? something else? PhD? Very curious if anyone knows.
I agree with Chollet that language is a layer on a cognitive base. Interestingly, the early universe, according to the standard model, was more accurately described by our language than the reality that we perceive. It wasn't until those few massless particles grew in number and created the force fields that gave the particles mass, that being as we know it came about...which is a combination of being and moving, not separated as in our language of nouns and verbs.
@@joeh858 Chollet invented a word to explain something, they ended up using it a few times. I was making a joke about that. It is a good video, great content.
GPT is a bookworm. Tesla's autonomous driving ai is always in the car. Alpha Go knows nothing but Go. None of those will produce a well rounded education. A good AI would know the rules of writing, memorize the multiplication tables, practice a sport and musical instrument, learn how to drive, hang out with friends on the weekend, go to the beach or the mountains every now and again, watch some movies and read something like Reddit a little. Maybe blend the subjects that an AI learns. The approach we take with kids is to teach reading not by simply exposing to words but instead exposing a picture book with images next to text. So, GPT3 knows statements about dogs that may be false but they are plausible statements. Tesla's autonomous driving data surely includes dogs and dog behavior around streets but that dataset maybe has no word label for the a dog. Google images have images of dogs that are labeled as a dog. If somehow these three data collections could be combined then the ai would have statements about dogs, images labeled with text about dogs and behavior of dogs.
Creative / divergent thinking should be considered highest form of intelligence as all human progress is dependant upon it, and yet it almost impossible to access. The school systems over reliance of standard models of intelligence testing does great harm to society. It filters out our most creative people unfortunately which results in reduced innovative which we already are beginning to see.
I really like the way this guy thinks, Francois pushes the boundaries of how to measure intelligence. I like the idea of not using the Turing test when people cheat and people will always cheat so it is best to utilize the scientific method. Like what Kate Darling was referring to, if we tell the judges they are talking to a child from Namibia then it is much easier to get them to agree that the computer is human because we have shifted the judges away from their base of reference. We are tricking them.
The 'child mind' is, at the beginning, pure observation. Language builds out of the interactions between other people. Thinking, pre language, is the child mind. In my opinion, all thinking continues to happen in this mind, and you have to translate on the fly. You get really good at it. Language only has meaning in the interactions between people. But the child mind still has to have a language to be translated. Other than pure observation, what comes before that?
Walking my dog, and I realize that when I have no mental context to narrow focus I just observer. I take in stimuli and it just is. But it's only when I ask a specific question of myself do I start to recall labels and give the scene English language.
In that framework, abstraction must be practiced and it must be shared. You can obsess over language on your own, but there is no guarantee that you will end up with something others agree with.
We can make a camera create an image on a screen, which our brains can then understand as an image. Take the screen out and hook it to an AI, but... What do our eyes do? Does the consciousness illusion keep us from making a proper investigation of what our senses actually do?
Emotions come before observation, your emotional state dictates your reaction to your observations. The same context and event can happen to the same person but be experienced, interpreted and processes completely differently. Your current mental state whether influenced by lack or sufficient sleep, supportive/toxic work environment, hunger or drunken state, or stressed or euphoric state will determine your reaction to events and ultimately the outcome
Hey Lex, have you heard about Zettelkasten? It's a method of organizing information which is somewhat similar to mind maps but more sophisticated. You might find it useful. Btw, thanks for you podcasts, they are amazing! ❤️
Thanks for these podcasts, lex. I've been looking for a podcast with things that I am interested in. Keep these things coming, I find a lot of value from it.
By no fault of the eloquent guest, my auditory processing issues made this difficult to understand. Strangely enough, 1.25x speed improved intelligibility. Something to try for anyone in a similar boat.
I am always puzzled by what exactly is "human level" driving means by different people. It seems by default it means the best human can be, which is very different from what actually is the human level driving in real world. Just because humans in theory can handle the corner case does not mean the average human would handle them perfectly.
Lex, don’t know if you’d see this - for visualizing notes you take, there are tools that allow you take notes in text files and make connections between notes very easily, so you get a “knowledge graph” for free. It’s easier to make connections between thoughts I think. For emacs and org (there’s one for md): github.com/org-roam/org-roam
Astonishing interview. I agree with Chollet's definition of intelligence and I sincerely hope that we will never develop non-deep learning AI. Or at least not before the final version of Neuralink.
Also, human level intelligence is far beyond what we are today: Learning to play chess is actually a solution to another complicated situation, like feed the hunger for getting recognized or get that geeky chess girl.
Cognition id prediction; beautiful. Develop potential, an energetic mind. The brain must be a muscle, mind builds from memory. Let go, begin again from 0...
Thanks for the pronunciation of his name. I've been struggling since I bought his book "Deep Learning With Python". Btw, he always sounds paranoid in his book. Am I the only one to feel like that?
I think an AGI needs to be able to „lazy“ learn from it’s environment while it is running instead of pre-learning everything through one shot learning. We humans do the same. We are motivated by something, try to move towards it and start learning when there is a delta between desired outcome and actual outcome. I think we do this by formulating problems and raising questions around those problems. Questions allow us to identify lack of knowledge or training in certain areas. We can start to reason from past experience or „pull“ new information or training required to improve/adjust our skills to be able to solve the problems and reach better outcomes.
You mean like a kind of background baseline running and stability and then the freedom to build n play n learn n create on top of.. ? I guess I use a computer program type of idea as a way of explaining it. I was thinking that I would like to note the way that one seems to work n function best. I think this because I know what it's like to be broken.. and I just want to ...not be broken ... I am writing this as they are talking about ways of hierarchy.. or a kind of template concept.
It would be great if the guests had some way to write things down, instead of only having to rely on words. For example, when Chollet was explaining mind maps, he could draw an example really quickly instead of trying to describe it.
Hehe, I read his book: "Deep Learning with python" --> It's by far the best book I've read about this topic. In my opinion, Human Intelligence often uses heuristics. Computational power can extend these heuristics by the huge amount of operations. So, the proof of "Four-Color-Theorem" was just possible by the combination of human intelligence and computers.
Training data size is not measured in gigabytes though. Just because videos are a lot of data doesn't mean they will provide that much more novel information compared to pictures
It's been said that babies have their own built in language that is understood by other babies, which eventually gets lost as they age. Something to really think about.
9:36 I have this idea regarding language and AI... I don't know how to properly elucidate it. I don't think language is necessarily the base "language" for communication, even internally. Like.... if you raised a child in complete sensory deprivation, except an artificial set of inputs you determined, I think the brain would be capable of communicating in a purely symbolic or mathematical (or whatever) form. And those forms could reach equal complexity. I think this is significant when you're talking about AI. Because it's kind of analogous: creating these babies in pure darkness communicating along lines of predetermined telepathy. And so... if the goals is to recreate something recognizable as human intelligence, it would be worth determining if language is actually at that base level. Or has just filled that niche and then been reproduced out of material benefits. (I paused the podcast to write that. Lol. So if they address this, and I'm repeating something, I haven't heard it yet.)
I think you're right, but I also think most people would agree that "the brain would be capable of communicating in a purely symbolic or mathematical form" *is* language
@@AlanShore4god I do wonder if there is some form more abstract that would serve that function. Or if by serving it, it becomes that, whatever the form. As in... could there be something so fundamentally different filling that role that it would be unrecognisable but able to communicate. Lol. Ahh... I have this idea in my head of some kind of higher base language. Able to operate in a kind of parallel. But I literally have no idea how to articulate it properly.
Great conversation, general intelligence is a very interesting topic. The intelligent mind seems to me to be a confluence of culture, curiosity and the ability to address new environments and problems in a new and imaginative way, it’s a dynamic constantly evolving awareness of environment, threat level and incentives.
Animals possess forms of communication that match the complexity of their biology which arose as a consequence of it and to a degree aided in its evolution. An analogy of an operating system is too rigid, rather it is more like turning a dial of increasing capability and dimensions of it.
No other creature possesses, 'Applied Intellect'. INTELLIGENCE must be more than the speed at which new skills are acquired, you must first have the ability to interrogate the world around you, and the ability to use that knowledge. i.g. What is that, and how can I use it. Human Beings have an extraordinary quality, the need to solve problems. In the absence of a problem, they will create one, that's what games are all about. A simple question, with broad applicability, is more valuable, than a unique question with limited applicability. (maybe?)
How do we count in the expense of evolution in human systems? The compute that the nature "invested into us", into building our priors? I'd be careful not to ignore that part when evaluating the efficiency part of the equation. I like Shane Legg's equation from the "Universal Intelligence" paper but it does lack the efficiency part that Chollet is insisting on. Also I loved his arguments against Hutter's compression hypothesis of intelligence. I wonder what Hutter would have to say, considering his work on universally optimal agents - AIXI work.
@47:20 This is what I was thinking. There need to be models of cognition that can interpret BASIC abstract ideas which can then be iteratively developed. AGI is crippled without these.
Great conversation. Thank you. A couple things: I wish you had talked about the ability to “know” which transcends our compressed past and our recognized ways of perceiving, by some considered legitimate and real ways of tapping into knowledge which is actually beyond of physical finiteness. Also, our interconnected ness to other life forms. Perhaps even non physical consciousness? How do we factor in ways of learning and perceiving info from these sources which can add to our compressed knowledge as well as, dare I say it, pave ways into new ways of knowing which transcend linear time and space itself? Taking into account recent revelations in the fields of quantum physics and the emerging areas studying consciousness itself, it seems like these considerations are important not only in the development of AI, but also when considering the meaning of life itself.
I really enjoyed this conversation with François. Here's the outline:
0:00 - Introduction
5:04 - Early influence
6:23 - Language
12:50 - Thinking with mind maps
23:42 - Definition of intelligence
42:24 - GPT-3
53:07 - Semantic web
57:22 - Autonomous driving
1:09:30 - Tests of intelligence
1:13:59 - Tests of human intelligence
1:27:18 - IQ tests
1:35:59 - ARC Challenge
1:59:11 - Generalization
2:09:50 - Turing Test
2:20:44 - Hutter prize
2:27:44 - Meaning of life
Should be a simulation time-stamp every pod :D lol cheers
@@newwave77 I
We want aurelien geron
7
Invite Terence Tao. He's great!
“I’m not interested in psychometrics as a way to characterize 1 individual person. I think psychometrics is most useful as a statistical tool. It’s most useful at scale.” This man is a measure of intelligence himself.
Listening Francois gives so much insights on everything. He is just simply amazing
Hey, Lex, I love how you dig up the guests that we didn't know we needed to hear and provide high stimulating intellectual discussion, keep it up
This guy is definitely very famous in the ai world! He wrote a very famous and important paper, " on the measure of intelligence" recently
@@jp2kk2 I'm a humble physics student but I'll have to check that one out, thanks!
@@jp2kk2 he created Keras, which might be better known by people.
This man helped me get through my ml elective
Lex, Thank you for providing us layman with this extremely beautiful and captivating podcast. We all appreciate the time and effort you put into this project of yours. Life would truly be quite different without you!
“Intelligence is how efficiently you learn new things”
“Measure of intelligence is ability to change”
Being able to learn comes from ‘understanding structured operation logic” model between events.
And being able to change comes from the ability to use causal inference on a perceived event against the model.
That is consistent with the fact that IQ and the personality trait Openness correlate quite strongly
Am yet to observe non-supportive data of this statement!
But this definition would mean that intelligence in humans is strongly tied to plasticity making children more intelligent than adults. However children struggle with many causal relationships that are easier to grasp as adults. This isn't because children are dumb but that this understanding of operational logic is actually learned. What you are describing is often referred to as "learning how to learn" but this is not the base of learning, just a skill that helps with applying other logical relationships in new contexts. An intelligent system would be capable of reaching this logical understanding but it does not need to start with operational logic a priori.
@@goclbert While I don't disagree, you have to consider that children's minds are not fully developed and lack many computational abilities of adults.
@@goclbert I think going back to Francois point you could argue that this isn't evidence that adults *are* actually smarter since it's just a matter of having more training data 🤣
When Francois talks, he gives a new insight every time. What a genius!
I love the way these two people think completely differently. I've had conversations about the way I think compared to my significant other, and its surprising to see how both similar and different ours ways of thinking can be!
Foggat
@@justindecker4051 imagine a flat earth believer commenting on your post ☠️
We need Joscha Bach to come back ASAP!
This please!
I've listened to that podcast 3 times already.
We need more Joscha Bach!
That would be epic
hes awesome! :)
We need him to come Bach
Would love to listen to a conversation between François Chollet and Joscha Bach.
'So you disagree with Naom Chomsky?!'
'𝓨𝓮𝓼'
10/10
Gay
@@thisway6539 people are allowed to disagree with each other, that's basically how new ideas sprout
Noam Chimpsky the chimp and Koko the gorilla effectively debunked Chomsky's core assertion that language is the root of intelligence. Various other animal studies have also broadened our conception of intelligence, for example the toolmaking ability of corvids.
@@MattOGormanSmith instead of being the root I'd argue that language is an outcome of intelligence. That it is itself a tool for articulation of more complex and abstract thought that simple action and reflex can't represent efficiently.
I listened to this after consuming thc.
I had several mental breakthroughs by applying the albeit hard for my intelligence level to decipher.
Highly valuable information contained within.
Lex you have an amazing mission to question the questions, That has helped sculpt my mind into one that is slightly more efficient than it was. More efficient at detecting bias.
I still have many biases. But I am getting better at detecting them.
Thankyou youtube for providing such valuable content. The world is both better and worse for the torrents of information literally at our fingertips.
From one random guy on the internet.
Choose the content you consume wisely.
Turn the play speed up incrementally over time as you adjust.
This adds profound gains in efficiency of absorption. The faster play speed becomes a second language. You will become philosophically bi-lingual. Other people will likely not understand 2.0x play speed yet it sounds normal when you become fluent.
3.0x is only 50% faster than 2.0x.
My average level of intelligence. Or maybe the frequency perceptiveness of my ears, will reach a performance ceiling before others with more capacity.
Neurolink has the ability to increase the speed of uptake. In a realm where consciousness consists of stored information combined with interpretation and adaptation.
Learning the dataset is just as important as improvisational ability. Any efficiency increase should produce measurable results in all round performance.
/consciousness.exe
“Everything we do creates ripples into the future and that’s the path to immortality”
I really like the way this man talks for some reason, like he has a fairly thick accent yet is more intelligible than a lot of native English speakers.
Fransios is such a guy that warms your heart with knowledge!
When he talks I instantly imagine a "smoking" jar with liquid nitrogen, just in this case it's pure liquid intelligence oozing towards us :)
Legend says that when Francois applied to Google, the interviewers began to question the level of their own Googliness
The best logical person i ever heard talking about intelligences and AGI.
I request a part 3. I’ve watched a number of these AI interviews, but in my opinion Francois has the deepest and clearest insights on this topic.
"Russian starts with a language and ends with a vodka" :) lol
Spare a thought for those that skipped the ads. That was Lex gold!
What a great discussion! I loved the metaphor of having a subway map in your head as you write a paper. I do that all the time.
Thank you both lex and Francois. This was a great pod cast. Thank you for this genuine and honest podcast. I as a viewer appreciate the knowledge you two are able to explain in understanding ways. God bless u both for ur good knowledge on things
This channel is a hidden gem on youtube.
Carry on lex.
Francois explains things in detail and his psych approach has keypoints I find impressive. I have the same view on what the outcomes are in testing. Knowing and not expecting depending on the circumstance does impact the targeted goal. I thought this in my experience(s) and may I add, its an added challenge not being able to hit some goals on your own. Meaning mine, I seem to kick my own butt. You guys already know how to talk to each other. I am really going off the rail communicating.
Loved the Piaget reference. One of the experiments described in Genetic Epistemology has had a profound impact on me.
I really like the way this guy thinks. Glad to see him on again!
This is an amazing interview. One of the most rewarding ones I have ever heard. Fantastic!!
Thank you for this informative video! 😍🐶😻Love Francois & Lex!
This guy.. is intelligent. I believe he has solved what intelligence really is, and that’s amazing. Also he has great insights about thé shortcomings of current deep learning and of neuralink
Bidirectional search 14:45
Love your work, mate. You're an amazing speaker and an even more amazing interviewer (or vice versa 😃)! Currently chugging through your extensive backlog of interesting content! Love it and keep up the good work!
Language - the operating system of the mind. Incredibly profound.
I don't think more intelligence is necessarily a good thing. Rather what we need is more equality, kindness and self-knowledge
What a great interview, makes me realize how crucial the ability to simulate reality is to intelligence.
Lex, as I observe the world and people around me, I also determined primary measure of intelligence is the ability to adapt to change. This is driven by deeper layers of character/mind that I have identified.
This one got me really excited
What the hell? 18k views, 140 comments? These are too low, people dafuq. This shit is the most interesting informative discussion ever
Edit- ... discussion ever today, as every podcast is amazing
This show breaks my brain everytime I watch it. Love it. Thx Lex and guests. Hella outlet.
Have David Deutsch on your podcast!!
ungratefulmetalpansy no Deutsch is the basically the father of quantum computation but his views on practically any matter are probably enlightening
@@kevinpurnell9465 especially AGI !
David Deutsch is truly a genius.
Yes, Lex, ask David Deutsch in AGI. This will become a very interesting conversation...
Roger Penrose and Jeff Hawkins were already interviewed so it seems natural to me for Lex to complete the set of AGI experts by inviting David Deutsch.
François is such an amazing person, loved this!
Love his cute french accent and his calm, quiet and relaxed way of talking during the entire time!
2:31:15 he is playing both sides do that no manner which one wins he comes on top, this man is a tactical genius as well
Wonderful and very subtle ideas about consciousness and learning. Thank you.
Some historic non-quotes: Descartes on logic: “Remember to always put des cartes before des horse.” and “C’est bien joué Monsieur Des Cartes.” Accordingly to Bertrand Russell, Descartes always wore his sword at his side, even when he was in his apartment.
Thank you. William L. Ramseyer
Would love to hear discussions on the differences in the mechanisms of emotional and classical intelligence.
I've always been curious, what is Francois' academic background? He's obviously a great engineer -- keras is a beautiful API and a huge success. Somehow he also has the ability and time to also work on deep questions of intelligence. Did he study physics? computer science? something else? PhD? Very curious if anyone knows.
Computer science
This was really beautiful. Especially the very end starting from the meaning of life. Hoping to be a patron soon. Best wishes
I agree with Chollet that language is a layer on a cognitive base. Interestingly, the early universe, according to the standard model, was more accurately described by our language than the reality that we perceive. It wasn't until those few massless particles grew in number and created the force fields that gave the particles mass, that being as we know it came about...which is a combination of being and moving, not separated as in our language of nouns and verbs.
Hell yes, I was waiting for exactly this episode ever since I read the paper.
The amount of no real words in this pod is out of the charts. Amazing talk as usual, thanks, I learned a lot.
Sorry if I misunderstand, but that seems like a contradictory statement?
@@joeh858 Chollet invented a word to explain something, they ended up using it a few times. I was making a joke about that. It is a good video, great content.
@@MarcosBiga Oh, I see what you meant now by "no real words"
@@joeh858
"Agentness" :)
I just checked it, Oxford English Dictionary doesn't know it :)
Thanks a lot Lex
GPT is a bookworm. Tesla's autonomous driving ai is always in the car. Alpha Go knows nothing but Go. None of those will produce a well rounded education. A good AI would know the rules of writing, memorize the multiplication tables, practice a sport and musical instrument, learn how to drive, hang out with friends on the weekend, go to the beach or the mountains every now and again, watch some movies and read something like Reddit a little. Maybe blend the subjects that an AI learns. The approach we take with kids is to teach reading not by simply exposing to words but instead exposing a picture book with images next to text.
So, GPT3 knows statements about dogs that may be false but they are plausible statements. Tesla's autonomous driving data surely includes dogs and dog behavior around streets but that dataset maybe has no word label for the a dog. Google images have images of dogs that are labeled as a dog. If somehow these three data collections could be combined then the ai would have statements about dogs, images labeled with text about dogs and behavior of dogs.
Notice that you describe very human-centric AI. That's almost certainly not the only one there could be.
Creative / divergent thinking should be considered highest form of intelligence as all human progress is dependant upon it, and yet it almost impossible to access. The school systems over reliance of standard models of intelligence testing does great harm to society. It filters out our most creative people unfortunately which results in reduced innovative which we already are beginning to see.
I really like the way this guy thinks, Francois pushes the boundaries of how to measure intelligence. I like the idea of not using the Turing test when people cheat and people will always cheat so it is best to utilize the scientific method. Like what Kate Darling was referring to, if we tell the judges they are talking to a child from Namibia then it is much easier to get them to agree that the computer is human because we have shifted the judges away from their base of reference. We are tricking them.
The 'child mind' is, at the beginning, pure observation. Language builds out of the interactions between other people. Thinking, pre language, is the child mind. In my opinion, all thinking continues to happen in this mind, and you have to translate on the fly. You get really good at it. Language only has meaning in the interactions between people. But the child mind still has to have a language to be translated. Other than pure observation, what comes before that?
Walking my dog, and I realize that when I have no mental context to narrow focus I just observer. I take in stimuli and it just is. But it's only when I ask a specific question of myself do I start to recall labels and give the scene English language.
In that framework, abstraction must be practiced and it must be shared. You can obsess over language on your own, but there is no guarantee that you will end up with something others agree with.
We can make a camera create an image on a screen, which our brains can then understand as an image. Take the screen out and hook it to an AI, but... What do our eyes do? Does the consciousness illusion keep us from making a proper investigation of what our senses actually do?
Intuition comes before that.
Emotions come before observation, your emotional state dictates your reaction to your observations. The same context and event can happen to the same person but be experienced, interpreted and processes completely differently. Your current mental state whether influenced by lack or sufficient sleep, supportive/toxic work environment, hunger or drunken state, or stressed or euphoric state will determine your reaction to events and ultimately the outcome
Hey Lex, have you heard about Zettelkasten? It's a method of organizing information which is somewhat similar to mind maps but more sophisticated. You might find it useful.
Btw, thanks for you podcasts, they are amazing! ❤️
my favorite twitter person is back!
Good interview (so far). Seems like the proper view on language.
Masterclass looks good. I would also like to pre-order that Sci-Fi novel of yours Lex.
Thanks for these podcasts, lex. I've been looking for a podcast with things that I am interested in. Keep these things coming, I find a lot of value from it.
Lex should interview Dr. Peter Hershock with a background in comparative East-West philosophy
ARC in kaggle still not achieve 0 yet.. so we still have many things to do
Another great guest and interview. Best response to the meaning of life question I've heard.
By no fault of the eloquent guest, my auditory processing issues made this difficult to understand. Strangely enough, 1.25x speed improved intelligibility. Something to try for anyone in a similar boat.
Luv the poem...Pushkin request
I am always puzzled by what exactly is "human level" driving means by different people. It seems by default it means the best human can be, which is very different from what actually is the human level driving in real world. Just because humans in theory can handle the corner case does not mean the average human would handle them perfectly.
Lex, don’t know if you’d see this - for visualizing notes you take, there are tools that allow you take notes in text files and make connections between notes very easily, so you get a “knowledge graph” for free. It’s easier to make connections between thoughts I think. For emacs and org (there’s one for md): github.com/org-roam/org-roam
Not just new but also difficulty .. level of newness and complexity
One of the few youtube channels that increases my IQ
According to some academics, IQ can't be improved, and it's all to do with what genes you inherited. Which is depressing.
@@halaldunya918 IQ doesn't really matter that much though. If something fascinates you, you will still need to work to understand it better.
Без водки, не случился б не один прорыв! Спасибо тебе за лучшие подкасты, с удовольствием смотрю и узнаю для себя, что-то новое!
Astonishing interview. I agree with Chollet's definition of intelligence and I sincerely hope that we will never develop non-deep learning AI. Or at least not before the final version of Neuralink.
We need an update!
This blew my mind
The fact that during the entire interview (it looks like) your phone screen is on at full brightness with white background is making me anxious :)
Also, human level intelligence is far beyond what we are today: Learning to play chess is actually a solution to another complicated situation, like feed the hunger for getting recognized or get that geeky chess girl.
As far as I know, really great people, masters in their field, not neccessarily chess, are primarily motivated internally, not externally.
Really admire Francosi Chollet. Thank you for this amazing talk!
We need another podcast with this based person
Cognition id prediction; beautiful. Develop potential, an energetic mind. The brain must be a muscle, mind builds from memory. Let go, begin again from 0...
Language, words spoken, the magic; code. Fix it! Thoughts make all the difference. I am in charge? Pray AI save the day!
Thanks for the pronunciation of his name. I've been struggling since I bought his book "Deep Learning With Python". Btw, he always sounds paranoid in his book. Am I the only one to feel like that?
Do a video on vodka cocktails..
I think an AGI needs to be able to „lazy“ learn from it’s environment while it is running instead of pre-learning everything through one shot learning. We humans do the same. We are motivated by something, try to move towards it and start learning when there is a delta between desired outcome and actual outcome. I think we do this by formulating problems and raising questions around those problems. Questions allow us to identify lack of knowledge or training in certain areas. We can start to reason from past experience or „pull“ new information or training required to improve/adjust our skills to be able to solve the problems and reach better outcomes.
You mean like a kind of background baseline running and stability and then the freedom to build n play n learn n create on top of.. ?
I guess I use a computer program type of idea as a way of explaining it.
I was thinking that I would like to note the way that one seems to work n function best. I think this because I know what it's like to be broken.. and I just want to ...not be broken ... I am writing this as they are talking about ways of hierarchy.. or a kind of template concept.
It would be great if the guests had some way to write things down, instead of only having to rely on words. For example, when Chollet was explaining mind maps, he could draw an example really quickly instead of trying to describe it.
"You listen to Lex's show?" "Yes." Okay, then you're intelligent.
Hehe, I read his book: "Deep Learning with python" --> It's by far the best book I've read about this topic.
In my opinion, Human Intelligence often uses heuristics. Computational power can extend these heuristics by the huge amount of operations.
So, the proof of "Four-Color-Theorem" was just possible by the combination of human intelligence and computers.
50:18 You can, in fact, train on 100x more data when you add in video and audio
Training data size is not measured in gigabytes though. Just because videos are a lot of data doesn't mean they will provide that much more novel information compared to pictures
@julkiewicz Well, I didn't say novel data. Just more data. The amount of novelty you find in that is not my point
Who wants Geoff Hinton on the Podcast ?
memories are information we collapses in real time,,,,,,,,,,,,,,,,,,,thoughts do not need language to exist....words do
It's been said that babies have their own built in language that is understood by other babies, which eventually gets lost as they age. Something to really think about.
Lex, thank you so much for yet another awesome podcast!
Thank you so much for the insightful discussion!
42:40 Lex is correct here. My guess is that the GPT-3 hype surge was just due to more people getting access to the OpenAI API.
Lex thank you everyday for inspire me
for real reasoning we will need a intermediate semantic abstraction in the ai model itself - like a symbolic knowledge graph similar to web3
9:36 I have this idea regarding language and AI... I don't know how to properly elucidate it. I don't think language is necessarily the base "language" for communication, even internally. Like.... if you raised a child in complete sensory deprivation, except an artificial set of inputs you determined, I think the brain would be capable of communicating in a purely symbolic or mathematical (or whatever) form. And those forms could reach equal complexity.
I think this is significant when you're talking about AI. Because it's kind of analogous: creating these babies in pure darkness communicating along lines of predetermined telepathy. And so... if the goals is to recreate something recognizable as human intelligence, it would be worth determining if language is actually at that base level. Or has just filled that niche and then been reproduced out of material benefits.
(I paused the podcast to write that. Lol. So if they address this, and I'm repeating something, I haven't heard it yet.)
I think you're right, but I also think most people would agree that "the brain would be capable of communicating in a purely symbolic or mathematical form" *is* language
@@AlanShore4god I do wonder if there is some form more abstract that would serve that function. Or if by serving it, it becomes that, whatever the form.
As in... could there be something so fundamentally different filling that role that it would be unrecognisable but able to communicate.
Lol. Ahh... I have this idea in my head of some kind of higher base language. Able to operate in a kind of parallel. But I literally have no idea how to articulate it properly.
Great guest and great interview! Thanks Lex.
Great conversation, general intelligence is a very interesting topic. The intelligent mind seems to me to be a confluence of culture, curiosity and the ability to address new environments and problems in a new and imaginative way, it’s a dynamic constantly evolving awareness of environment, threat level and incentives.
Educators call it Semantic Mapping of ideas
The test already exists: demonstrate the ability to define, identify the unique characteristics of, ever smaller (or larger) "units" of matter.
Animals possess forms of communication that match the complexity of their biology which arose as a consequence of it and to a degree aided in its evolution. An analogy of an operating system is too rigid, rather it is more like turning a dial of increasing capability and dimensions of it.
Science might catch up!
No other creature possesses, 'Applied Intellect'.
INTELLIGENCE must be more than the speed at which new skills are acquired, you must first have the ability to interrogate the world around you, and the ability to use that knowledge.
i.g. What is that, and how can I use it.
Human Beings have an extraordinary quality, the need to solve problems.
In the absence of a problem, they will create one, that's what games are all about.
A simple question, with broad applicability, is more valuable, than a unique question with limited applicability.
(maybe?)
How do we count in the expense of evolution in human systems? The compute that the nature "invested into us", into building our priors? I'd be careful not to ignore that part when evaluating the efficiency part of the equation. I like Shane Legg's equation from the "Universal Intelligence" paper but it does lack the efficiency part that Chollet is insisting on.
Also I loved his arguments against Hutter's compression hypothesis of intelligence. I wonder what Hutter would have to say, considering his work on universally optimal agents - AIXI work.
@47:20 This is what I was thinking. There need to be models of cognition that can interpret BASIC abstract ideas which can then be iteratively developed. AGI is crippled without these.
Great conversation. Thank you. A couple things: I wish you had talked about the ability to “know” which transcends our compressed past and our recognized ways of perceiving, by some considered legitimate and real ways of tapping into knowledge which is actually beyond of physical finiteness. Also, our interconnected ness to other life forms. Perhaps even non physical consciousness? How do we factor in ways of learning and perceiving info from these sources which can add to our compressed knowledge as well as, dare I say it, pave ways into new ways of knowing which transcend linear time and space itself? Taking into account recent revelations in the fields of quantum physics and the emerging areas studying consciousness itself, it seems like these considerations are important not only in the development of AI, but also when considering the meaning of life itself.