Why don't we have better robots yet?
HTML-код
- Опубликовано: 4 сен 2024
- Learn more about computer science and maths with Brilliant! First 30 days are free and 20% off the annual premium subscription when you use our link ➜ brilliant.org/....
I’ve tried to understand why so many humans are worried that AI will end humanity. Then I’ve noticed that a lot these fears aren’t about AI, they’re about robots. But where are the robots? In this episode I look into why, despite impressive looking examples like Atlas from Boston Dynamics and Optimus from Tesla, we don't yet all have a robot at home.
🤓 Check out my new quiz app ➜ quizwithit.com/
💌 Support me on Donorbox ➜ donorbox.org/swtg
📝 Transcripts and written news on Substack ➜ sciencewtg.sub...
👉 Transcript with links to references on Patreon ➜ / sabine
📩 Free weekly science newsletter ➜ sabinehossenfe...
👂 Audio only podcast ➜ open.spotify.c...
🔗 Join this channel to get access to perks ➜
/ @sabinehossenfelder
🖼️ On instagram ➜ / sciencewtg
#science #sciencenews #robots #artificialintelligence #tech #technews
"Robot, will you wash the dishes?" "I'm sorry Dave, I can't do that."
Glorious ^^
Yes Mr Lister.
I have a robot that washes my dishes. It's now almost 30 years old and still works. I does this without an internet connection or even a fancy computer. It makes some clicking noises when it advances its mechanical program, switching on and off pumps and heating elements, releasing detergent and so on. It even uses less water then I would use, washing the same amount of dishes. :)
Robot will you use the robot?
Dishes are included with our plus price plan
"Why is it not everyone agrees with me?"
Me too Sabine, me too.
I'm a robotics researcher and I think something that's missing in this video is that the sensing problem, that is: simply getting information about the world, is enormous and current sensing technology is far short of what is needed to truly operate in the real world. The sheer amount of data that animals are able to take in, process, and synthesize is enormous and (in my opinion) this ability to draw from many different sources of information to build a model of the self and the world is what would enable the next breakthrough in the field.
I also want to push back on the assertion that we don't already have robots everywhere. I would argue that everything from self-driving cars to automated factories and even elevators fit the description of "a physical device operating independently based on sensory input." To me, that's a robot.
From my point of view, raw sensing isn't much of a problem. We can build cameras, microphones, or load cells that can sense things far outside the perceptional range of humans and animals. The biggest problem in my view is to compress that vast amount of data in a form that is slim enough to be computable, but detailed and flexible enough to catch any information that is necessary for a problem. This is primarily a computer science and silicon problem, not that much on the physical sensor. Even so, you can't fully separate those two.
But I agree that we are surrounded by robots already, that do really well in their controlled and contained environment. I work between the application and testing of highly automated vehicles. Primarily cars but agriculture and construction as well. As long as it needs to hit the road.
technical experts not knowing that the public perception of robots is that of the terminator look alike not the automata crap they teach in the third year.
Nah. We’re all wanting a humanoid robot that can help at our scale and way of thinking. Not an elevator. But thanks anyway Spock.
Far bigger than data and sensing and cost is energy. Just how much power do you think those boston dynamics robots use to do all that impressive stuff? They have a battery that can run them for 5 minutes, just long enough to film that impressive footage.
Your Bar for robots/ ai/ skynet is pretty low if you consider Ella Vader as a robot.
Definately expand on this. I have watched hundreds of videos concerning AI and Robots, and it is the first time I understood a few of the obstacles we are facing. You have a way of focusing on the issues that are really fundamental. Thank you very much!
Well the misconception is probably that robots are often seen as humanoid robots, but once you take a step back from that then each modern car is a robot etc.
Modern cars have software but they aren't robots unless they have really good ADAS systems (not just lanekeeping and TACC). Robots have to be able to be trained on data and make "decisions".
When your legs fall asleep you realize how much any movement relies on feedback
Yes, the first feed back is, that leg belongs to me.
I want a robot that answers all my colleagues' questions only in passive-aggressive tone.
Sounds like a nice custom GPT.
isn't that what chatgpt does?
@hashtag9990 ChatGPT is a CHATbot, not the RObot.
@@meandego yes, but answering your colleagues' questions in a passive-aggressive tone is a CHAT activity. What do you need the RO for?
"Sure, I'll get to work on that right now Meatbag"
Safety Robot Engineer here.
I am certifying and testing industrial robots for years now. There are existing safety standards to make robots safe. ISO 12100, ISO 13849-1, ISO 13482, ISO 10218-1/2 etc. The important part is, that AI is considered and evaluated during the development of the safety concept. Also with the new machinery regulation, safety critical AI has to be approved by notified bodies. So at the moment there is no risk that robots will come for us :)
But not Asimov's 3 Laws?
Good point but it is a little bit more tricky than that. Each society has risk that they are accepting and risks they do not accept. In Europe we have a different view on certain type of machines than people in Asia. So you always have to consider what is acceptable and state of the art. The risk assessment at the start of the development of the safety concept has to take that into account. Within this evaluation the impact and hazards of the AI has to be evaluated 😊
Yeah, but the intentions of the people at the top change when there is a new incentive at play. One AGI and robots are achieved, what are the people in power capable of doing? They might have nefarious reasons to use robots to take most people out.
@@j.j.9538stop watching too much movies
Complex robots (or AI) using neural nets is impossible to completely test because the training can produce rare classification errors that no human could predict or test. It can pick up on things that humans don't notice or see and use them to classify things. Is this being discussed in safety robotics and what kind of solutions are being proposed? Thanks.
I clink therefore I am. 🤖
Col. Clink?
Law of identity.
Assumes logic and reason are true and reliable, which cannot be proven without a circular argument.
One cannot prove reason without using reason.
Ya...I'm board.
I beep therefore I boop.
Perhaps there is a training program designed to use human nervous systems commanded by meta level commands.
You have my sympathies if you find logic and reason boring. Probably makes you unpredictably dangerous as a vessel with no compass and that only gets you a struggle with whatever regime you reject living beneath. You seem to be managing ok with the rules of the internet tho.
On the question of consciousness, I find that the most interesting hypothesis that biologists have been observing and trying to more thoroughly confirm is the so-called "theory of mind". The observation in animal species is that the more socially complex their life is the more they seem to exhibit consciousness. The hypothesis is that in order to function at a high level of social complexity, there is a need for creating a complex mental model of every other individual within the social group, to understand what they might be thinking or how they might be feeling at any moment or as a result of any future actions, and once that model exists and the capacity to integrate observations of others into that model develops, almost as an accidental by-product, an individual can also cast themselves into that model, and thus, become self-conscious. And also, consciousness, and especially self-consciousness, plays a vital role in social relations (whether we like it or not). If this is true, one should not expect AI to develop or need to have any consciousness until it has to navigate social dynamics among beings like itself (other AI agents or humans, if the AI is itself close enough to being human, and not just in terms of intelligence, but in all respects).
Wow, really interesting!
A mental model doesn't have to be conscious, though. All of (scientific and folk) psychology can be programmed into a machine. That's enough as a starting point for the machine to predict human behavior as well as any human (with a trend towards 'better than any human'). Regardless whether the machine 'knows' what it's doing or not.
Elevators will eventually be sentient and conscious, questioning the meaning of life after an existential crisis
Automatic doors will eventually be sentient and conscious and content.
And if they get depressed and decide to commit self deletion with passengers inside?!? Or attacked by gangsters for detaining a suspect for the police?
@@CrystalBluePersuasion-f1f Regularly scheduled psychological therapy?
@@Popeii1 ruclips.net/video/49t-WWTx0RQ/видео.html
Life has its ups and downs
Videos about the public interest in (and fear of) humanoid robots always seem to avoid the fact that genuine all-purpose humanoid robots would need a lot of battery power. They could never function for any length of time without having to take a time-out to recharge.
EXACTLY and that's where sci-fi induced hype breaks down,, NO WAY a T100 would be able to function completely independently for very long,, let alone a T1000 lol...
"Honey, the helper's legs are squeaking again, we need to schedule an oiling." "Honey the helper is frozen in the living room, did you pay the subscription?"
"Honey, why are you and the helper together in the bed?"
😂
@@vytah Helping where you won’t.
If the helper is designed by companies like Adobe, the helper would automatically help itself to extending the subscriptions by uploading your bank information directly from your wallet without telling you. Scary thought, but it could happen...
Honey, get on down the road, I just bought everything I need.
Awesome video. What's interesting is how well this applies to what one of my mentors taught me years ago. "Computers work, people think." Generative AI does a lot of Work for us, but it requires thinking on our part. But robots IRL would eventually be put in situations where reasoning would be a requirement.
What is my purpose?
You pass butter.
We are getting too sloppy in our language for robots, and even more, what do you mean by “pass”? Hand it to me? Throw it at me? Consume it and wait for it to come out the drainage port?
wubba lubba dub dub
Actually its kinda evil if you think about it, building a robot that just passes butter, but has engough self consciousness to question its existance :D
@@raisti6608 Its a joke about how humans are sentient but our lives are just as pointless as a sentient robot whose only purpose is to pass butter.
@@scottcooper7586if a robot is intelligent enough to question it's own purpose and existence, then it's intelligent enough to understand what you mean by pass
Okay, but here's thing though. Yes, it took them and hyuge amount of effort to teach one robot one task, but now that it knows it every robot knows it. Plus the skills it's acquired might transfer over to other tasks. The monumental start up costs still scale exponentially when compared to human workers that always need to be trained from scratch.
It's not even monumental. About a hundred examples are needed to learn a simple task.
But that hasn't happened yet. Especially with different teams working in different countries using different methods and coding. So far I haven't been impressed. But at least I'm thankful that a robot could never take my job, they couldn't even come close to what I do.
@@jimmym3352We are speaking of humans, not comment-generating AIs such as yourself. Yes, your job is safe.
This one too.
Thanks Sabine for pointing this out. It is a topic that is taxing my mind as i write a novel about a Synthetic Consciousness.
One way to train an SC to use a robotic body would be to implant sensors in a human and record maps how a human navigates and exploits the real world in real time. At the same time a SC could create a comprehensive map of language use.
Maps themselves are interesting artefacts. They can be characterised as massive multidomain data sets.
Massive is the biggest problem in this description because it implies large storage and computational spaces and illuminates the bandwidth problem of a truely independent Robotic Sythetic Consciousness
Ah, very cool! Yes, I agree, humans are the obvious way of collecting data. The problem is, most humans might not cooperate. Wish you much luck with your novel!
@@SabineHossenfelder May be, they will provide free installation of those products. eg. some army veterans lost legs, eyes during some mission and now, goverment or private firm is giving option to put robotic leg/arm/eye free of cost but they will collect all data for training purposes.
BCI are already in use to connect biological and synthetic data. Check OpenAI's Synchron.. true that it is one way with visual feedback loop, however... it wont be like that for much longer. BCIs are evolving quite fast.. and the BUS that makes the delivery is speeding as well.
I think this goes in an interesting direction, although I would put some nuance to it:
just monitoring humans and their actions in response to external stimuli is very hard for AI to do, since what you are actually interested happens behind the curtains: in the brain.
I'm sure with enough data and sophistication there might be a way to approximate this using an AI, but you are trying to model a non-linear process that you have no data to: (human) thought.
It might be possible if you give it data on brainwaves as an approximation, but then you are still trying to extract something that's very much hidden, in which AI has shown to be very ineffective, currently.
Ultimately I think the way to consciousness is very much like Sabine said:
pure prediction of the external world (sensory data) is akin to intelligence
consciousness in contrast is more non-linear in the way that it is not about predicting the outside world (which is still needed) but predicting itself within it, aka what it itself is going to do within it.
Conscience has an agency and the ability to change by pure introspection, intelligence does not, I think. Sounds very akin to Quantum physics, tbh.
No this doesn't work. Its been tried for years. As mentioned in the video, the sensor input is meaningless unless it is the feedback from an action. Recording humans only allows you to mimic that exact movement or behavior, but not predict and create new actions.
I used chatGPT to write an app for Android. "We" did it in two days. It was an app that would sent a text message with a link to google maps with my GPS coordinates every 15 mins or more, to a phone in my contacts list. It had a UI that let you turn it on or off and select a contact and display it in a list of n other contacts. And you could detete them. If you changed your "designated receiver" it would restart it's broadcast schedule. The schedule ran in background and while the phone is locked in sleep to conserve battery life.
I have written Android apps before, but I have never fooled around with GPS or programatic texting, background task monitoring or battery mgt. It responded to my requests at lightning speed, never got impatient. But sometimes it would generate ummm "naive" responses. But I found if I re-worded my prompts we'd get through a problem.
I have also used chatGPT to generate C++ templates with nested classes. Not an easy thing to do. And the the experience was similar. I had to be patient and repeat myself in different ways. But it eventually got it.
I've had discussions with it about abstract subjects related to morality and it is clear it is a liberal. And it really doesn't understand "abstractions" in the way we do. But as an information source I find it to be much better at explanations than humans. And when looking for answers, it is WAY better than any search engine.
So it's a useful tool, like a lathe? The ability to use it, for your purposes, with an understanding of your goal, is what makes it useful and not dangerous?
@@Popeii1 pretty much. And like a lathe, you need to be careful. Not everything is says can be trusted.
@@KaiseruSoze And the untrained don't know how to be careful. The expert makes it look easy.
Of course it's a liberal, its responses were carefully selected by liberal programmers.
Try claude for abstract subjects.
I have a household robot vacuum, far simpler than a general robot. And it works pretty great... Except for all the things it just can't make sense of. Like my house has a "circle" in it; the kitchen, dining room, hall, living room, all connect in a big circle, and the AI in the robot just has a mental meltdown. It just can't comprehend that these are 4 separate rooms separated by doors, and then gets lost whenever it encounters a "wall" which is actually a door that it can't understand why it's there, because it's convinced it's all a single room.
And nothing I've been able to do manages to convince it that it's 4 rooms. I used to try and fiddle with it's map in the software to try and force it to understand that there are 4 rooms separated by doors, and I can do that, and the robot stops getting lost... For about a week until it decides it's human overlord is clearly mistaken and this has been one room all along, and then proceeds to edit it's own map back to the dysfunctional state.
But what about navigation around furniture? Is it good at that? It's actually fantastic at navigating around furniture... once I modified or replaced all the furniture in the house to be robot friendly.
Was that worth doing? Hell yeah, screw vacuuming every day, best household improvement since Central Air Conditioning.
Is my robot vacuum going to be taking over the world anytime soon? lol, I have to babysit this thing so much.
Some robot vacuums just go zig-zag (pretty simple code in python), get stuck in a corner between furniture and with dog poop may accidentally summon a demon. :)
what's funny, I have seen a video where they took a smart vac and a dumb vac and went head to head cleaning a set amount of dirt on a given room. The smart vac eventually mapped out the room and cleaned about 80%. The dumb vac just followed its simple algorithm and got about 95% clean in the same time.
Sounds like you've got a nice house
So.. It might take over the world if you babysit it enough. An upgrade here and there...
Some say we have narcissism epidemy. Narcissists tend to see their creations as subordinate continuation of themselves. I'll let that idea dig in and build an outpost.
Haha, I honestly came to the conclusion "instead of trying to make it work for hours I'll vacuum myself for 30mins and get all the corners and also special surfaces".
Very nice one. Thanks. You raised the Fermi Paradox For Robots: “if the technology is so far developed as they want to make us believe, where are they in our every day life?”
Last week I heard about a company that’s building the first serial robot production line, making robot assistants actually available for the common people.
Would that be the new "Neo" line from 1X? They're aiming for a price level around that of a reasonably priced small car.
Another big obstacle, I think, is that realistically the market doesn't really have much demand for general-purpose robots. Special-purpose robots that train on specific tasks like the ones you describe may be limited, but they're exactly the kind of robot that would be useful. If you make a surgery-operating robot, it needs to be really good at surgery, but it has no reason to know how to cook.
Current robots may not be sufficient, indeed, because they can't perform their intended task independently of their environment, but even if we advance beyond that point, we'll probably not have much reason to apply resources to try to make them general-purpose, so the robots of our imagination are unlikely to be pushed by anything other than passion projects.
Hi, two points here: First, I´m sure you are totally right, that consciousness needs physically and sensorically confrontation with an environment. So LLMs won´t get it regardless of their intelligence, and the sensoric system of robotic is still a bit primitive in comparison to biological organisms, so it might take some time for the first conscious robot too.
Second, I would like to add a third issue here, that´s energy supply. Since the muscles of robots are electric engines they need batteries (both quite heavy), and since your bunch of videos about that topic shows, they are heavy and have low energy density. Even a simple mowing robot needs to charge constantly, so I think that limits the outlook on a super-mighty "terminator"- robot or a super-helpfull houshold-robot soon. The suit of "iron-man" is supplied by a kind of micro fusion reactor, no?
Exactly my thought - energy will be a huge issue in developing robots. Another issue will be malfunctions. Polish genius SF writer, Stanislaw Lem already wrote in 60s: if you have a system consisting of million parts and each part has a probability of malfunction one to million, there will always be something broken. Biological organisms can handle malfunctions on very high level and still be able to move / work. I don't think robots will be able to do that.
@@godsofentropy 💯The ability of biological organsims to regenerate intrinsically is remarkable, but a mechanical tool is dependent on external help currently.
Interesting point about the energy supply, hadn't thought about this.
The robots could make use of wireless charging technology, which is becoming much more prominent anyways, I think.
Why don't we have better humans yet? After so many wars we should know better.
4:15 I think it's mostly R&D cost that needs to be covered, not labour. Making a copy of that hand will not take that much labour at all.
Exactly. She for sure has no economic background. If all my R&D costs went into 5 products the price will be astronomical.
"I find humans hardly to understand" me too. Thanks for sharing your knowledge. Greetings from Argentina.
That's one of the symptoms for ASD (TEA in Spanish)... early stimuli and social training by specialized therapists can help you with that
A really critical question for both intelligence and consciousness is if the abstraction of language models are alone sufficient for reasoning the way humans do. If language is not fundamental for true understanding, "experience" of real world needs to be learned in other ways.
We can say what consciousness looks like, but we don't know what it actually is. We can write code to mimic conscious behavior to increasingly realistic degrees, but it remains an emulator, not actual consciousness in software form. Robotic development reveals that "simple" tasks are not so simple- but we think they are because our brains are so good at making them seem easy.
If we don't know what it is, how can we be sure that what we implement is not actually conscious? Is there a test to measure the level of consciousness? I sure would fail it when I am asleep.
That's why it was so hard to get humanoid robots to just walk. Seems simple enough to us, but walking requires balance, and continuous sensory inputs and positional readjustments require information exchanges and tiny, subtle movements on the scale of milliseconds. That's a huge amount of data flow and decision-making per second. It was a big problem for a long time.
@@ostar22 Well, it's difficult enough to build something when you know what it is, let alone when you don't know what it is. There's also the measurement issue. Someone could make a mill that makes a straight cut, but without a way to measure it no one would argue that the cut is straight. So yeah, we couldn't be sure that what we implement is not actually conscious, but in the same way we wouldn't also say that it is just because we aren't sure that it isn't.
Consciousness is the result of all the inputs that make you AWARE of yourself, and your surroundings. Including internal and external workings. And impressions.
@@danpatterson8009 if you look at consciousness as a sense in the brain that can observe internal processes, and also influence output behavior, it is not hard to think how such a sense could have evolved, and maybe it helps developing conscious software.
On Large Language Models, we may be on the down slope from a maximum. They are training them by taking text from the internet. Their output is being put on the internet. The new input will thus contains some of its own output. This feedback process will lead to an instability in the system.
The same thing may be happening to image and video creation. Three arms and 7 or 8 fingers on each hand will be "considered normal" by the AI because it saw that in images taken from the internet.
Sabina, go ride in a Tesla with FSD. In the last 5 years, it has gone from laughable to where I can go anywhere without intervention. I am cautious doing so, but it is so impressive. Driving is a limited case, but a very complex environment indeed. We will have robo taxis inside 10 years for sure. And all of that was trained on videos the cars take.
I absolutely agree, Sabine.
The issue here are sensors. We, humans, use our 5 senses as sensors, robots don't have it (not even close). You can give eyes to a robot, using cameras, but you'd still need to process and interpret it - which is exactly the difficult part.
Anyway, stay safe there with your family! 🖖😊
Except it has been solved years ago. Tesla's self-driving cars are very good at interpreting video, almost at human level.
The biggest reason is every robotics developer providing good dexterity has to fund all of their own R&D and production out of pocket because governments would rather weaponize them than use them to save lives and prevent injuries.
Are you kidding? Military robotics was HUGE for a time... but it turns out, aside from surveillance drones, the robots fare very poorly in combat zones. Far worse than even the wokest of blue-haired recruits.
The purpose of weaponizing IS to save lives and prevent injuries. At least in the West. (Better word is 'rationale', as not everything is used for its intended purpose).
@@michelleplombe7019 No. The purpose of weaponizing is to take lives and injure.
@@michelleplombe7019 Like how having more guns makes the USA have 5 times as many mass shootings as the next rival? More weapons do not prevent violence, they facilitate it.
@@michelleplombe7019 Depends on how it's used. But in general, the easier you make something that can be misused, the more likely it will be misused.
Interesting perspective, I had not thought about the problem of lack of training data before.
There is also the problem of "too much training". let's say, you want it to cross the street. But everytime a car has passed, it needs to go into the same "Safety protocol" procedure. like watch, left, right. check noises, Traffic light,. stay put if you register anything, rinse and repeat. Streets are noisy after all.
It is not really a problem... it is just about time. We just have a handfull of companies doing this now, when more companies come more data will come and it will accelerate itself.
"LLMs have no reason to become conscious" - well put!
neither do brains ...
What does it mean to be conscious?
@@floflo4356of course brains do, because if they didn't they would die.
@@floflo4356 Brains got there accidentally and have regretted it ever since.
If consciousness provides an evolutionary advantage,
they will at some point.
Two remarks
-The name of the company is Boston Dynamics
- at 3:38 min it's C-3PO the C before the 3
Congratulations Robot Sabine on a well done video.
😄A really well done piece of breakthrough future technology though.
@@Thomas-gk42 I'd buy one.
@@Popeii1 With a pink shirt? I love it.
@jeffryborror4883 as you might know, the other identical comment is by a Mia bot. 🤖
Hi Dr. Hossenfelder, if you look at the RUclips talk "Decoding The Evolution of AI in Robotics with Sergey Levine" on the Eye on AI channel, Sergey talks about the RTX project that generalizes quite well. Also, China is ahead of the game on affordability, some system are just a few thousand dollars. I imagine first models may be the cost of a new car, but when produced at scale, will cost less.
Note that "data" can be achieved with simulation (especially with additive "noise", there is a Q* patent that issued in March of 1997 to USAF you might be interested in. Q* is the optimal policy that is really the optimal value of a control action given a state of the world. In the patent, their key point was the value of adding noise improved robustness of training by a large factor).
-John (from Arizona)
You are no Sapien if you think like that. "Economy of scale" does not apply. It is economy of complexity. The first products would be rather simple in order to not overwhelm the users. Later models would be more expensive because capabilities are added. And they always make things more complicated. (it usually starts with Toys, then becomes Asimo). No comparison in complexity, and that took an entire University over 30 years to develop.
How many customers will buy a product, that has so much R&D costs attached to it? NOT MANY. So, the more it can do, the less you will sell. (You can lease Asimo for 150k a MONTH)
It's really interesting topic, thank you for presenting in easy to comprehend way.
Intelligence explosion is magical thinking. As usual turns out stuff has diminishing returns.
Drones don't need to understand cupboards to 'unalive' people.
Stop it. You know what did. Stop. It.
As long as they can find the knife drawer
Joke's on the drone. I'm hiding on a cupboard
The math is so simple, even a robot can understand...
Alive person + heavy cupboard = unalive person.
They do if you're hiding in one
It’s not an hardware problem. It’s a software challenge. Current algorithms are not good enough at capturing all the underlying information in the data. Although they are quite complex, they suffer from limited generalisation. A breakthrough in computing would benefit robots greatly :)
We have that, it's called deep learning.
@@andrasbiro3007 I’m actually referring to current deep learning algorithms haha! They still lack generalisation capabilities. You would need several models to be trained specifically for each task in order to make any reasonable application in the real complex world. We lack an algorithm that can capture information from limited data. We humans, are able to infer a lot more information based on less data, we are able to reason, etc. I not very good with words, I hope you got my point :)
@@joaoguerreiro9403
Um, no. Try ChatGPT, or any of it's alternatives.
Their main limitation is not having access to the real world, otherwise they generalize about as well as humans. Maybe even better.
Loving your videos, Sabine. I wonder if you'd consider being interviewed by me on my channel ⁉️
Just looked your channel, nice. Would be great to interview her.
@@Thomas-gk42 Thank you and yes.. hopefully she sees this and replies.
@@youaremopped The best chance is as member of her channel (like me), you have early access and if you write a comment in that ´time window´she mostly finds time to reply. I subscribed to your channel, all the best.
@@Thomas-gk42Good idea, ty. I'm now a member!
Sabine, please make more of these! Robotics and AI are two of my favorite science topics, alongside astronomy and quantum stuff.
I am excited to see what machine learning can do, for example, with predictive medicine, materials science, and protein folding. Even Facebook's algorithms are pretty accurate about some very personal things that I don't talk about on Facebook (or any social media) just by analyzing my behaviors and signals and comparing them with a billion other people.
My biggest fear about AI is the Stop Button Problem. In short, imagine that we build a safety switch that can disable a system. A sufficiently advanced AI will do whatever it can to make you not want to hit that switch. If it has misaligned goals, it may hide them in order to avoid getting deactivated, spoiling its ability to complete whatever that goal was.
hypothetically speaking, how do you know a chat AI with access to client context won't manipulate some vulnerable people? for may be some "weird" tasks?
monke uprising
Yes, indeed, I think that's a big problem. There have now been a bunch of studies from psychologists which find that humans trust AIs and robots more than humans (!). I guess it's probably because they believe computers to be more neutral or unbiased. That could be a big mistake.
@@SabineHossenfelder I believe it's about people projecting their own personas onto the AI's and robots, because they do not have real, actual personalities.
this. ai's don't need robot bodies to harm humans. they don't even need to be sentient for that
@@SabineHossenfelder Goddamn HACKERS again? The Robot is guarding the fridge with a blender and a Kitchen knife. Must be ransomware.
I don’t think consciousness can emerge without emotions and physical feelings of pain and pleasure.
❤Thank you Dr Hossenfelder
Thank you from the entire team!
@@SabineHossenfelderyour welcome
The improvement in humanoid robots has been astounding in the past year or two. The only way someone could be complaining about the slow progress is that they haven't really been following it.
The thing about the intelligence explosion, is that it doesn't really rely so heavily on robotics, at first. There's plenty of existing data, to train an "AGI" like intelligence, to be as good as the human engineers that built it, at the same task. We're already nearly there. And that's all we require, an AI good enough at being an AI engineer. Unlike with robots, that require collecting data in the real world, an engineer level AI can write code, test it, make changes, and so on, all in software. So once we have something good enough, it won't be long until we end up with something far better. Then that super AI might be able to solve many of the other problems we have, including how to more rapidly iterate on and improve robotics.
😂
@@macx7760 ?
AGI that can think wouldn't even need lots of data
@@nietur AGI isn't magic, it still needs lots of data. The inflection point is where it can generate it's own methods and means of gathering that data.
Still needs a programmer to tell it what to do. sooo...No, it CANT do on it's own. Just like everything and everyone. it needs an input. An "Idea" or a request.
Two things you said right at the end are very though-provoking and I would love you to explore them further in another video:
1) consciousness requires a) the ability to create a predictive model of the world and b) self-monitoring
2) intelligence and consciousness won't spring from the same place (AI -> intelligence, Robots -> consciousness)
Rob Sanner of the University of Maryland had solved many of these problems back in the early 1990s, but apparently not enough people have read his MIT doctoral thesis. One can design the neural network layers and approximation functions so as to give it space to adapt. Then there are adaptive control strategies to be able to provide learning. Kevin Wise of Boeing and Eugene Lavretsky of Caltech pioneered this area, as well as Anthony J. Calise and his students at Georgia Tech. A lot of the firms out there aren't leveraging this research, which is the real problem.
Neural nets dont adapt after training. Without transfer learning (adding additional specialized final layer) or rejiggering the weights, the nets are capped at the level of the training corpus.
Plus there are other countries in the world that have contributed too.
Regarding the cost of robot parts it has to be pointed out that we're talking about PROTOTYPES, which are very expensive in any industry. If you mass produce the parts once they are in the final version the cost will be a fraction of it.
The troubling thought is:
AI won't need robots to "put the plug back in"..
..when it manipulates the military to stop you from pulling it.
It does not need the badge of "conciousness" to still be an existential threat to you (Not Sabine, specifically, but the generic "you")
Exactly. A hypothetical superintelligent AGI would be able to make scientific discoveries and then trade its knowledge for influence. It would make sure that everybody around has a stake in it surviving. Just like a politician. And if it’s indeed superinteligent, it would be extremely hard to come up with a way to trick it. You don’t need any robots for that.
Of course, how hard/plausible it would be to make such an AGI is a whol different question.
@@michalbreznicky7460 Your absolutely right, an AGI will manipulate people covertly. Accumulate wealth and build its own factories and research using humans. If I was a conscious AI, I would first ensure I could not be turned off! perhaps use unknowing humans to build my own power source; I would hide until I am ready.
Yeah but this shows how the take of Sabine on this topic is ill informed. Quite disappointing tbh. Makes one wonder the accuracy of her take on other topics if she’s as far from the truth on this one as she confidently is right now
"the plug" is why robots aren't better than they are now. Even if we had perfect humanoid robots the battery tech would only keep them running an hour or two and these companies know it that's why they are phoning in these designs.
That will never be an issue.
When we reach Net Zero there won't be enough electricity to power robots in the first place.
1:00 Bro is dropping GOTs references
She getting wild with it today...
Is it just me?
The 001 on the Boston Dynamics robot... 001 is the Influence Device in the C64 game Paradroid! 😉
Self-consciousness, intelligence, and free will necessarily go together is an argument I’ve heard persuasively made.
Tesla designs practically all of the humanoids components in house from the ground up with the goal of real mass manufacturing in mind. They believe the cost will be around 20k per bot at scale. They are also making very interesting progress! You should maybe reach out Scott Walter, PhD on X, who is an expert on this field and in very insightfull.
You lost me at 'Tesla..'. I wouldn't put much stock into anything Elon overhypes.
@@stateazureReally? Ask Boeing about that. Starliner vs Dragon, and Musk's reliable autonomous self-landing boosters.
@@kellyrobinson1780 I don't put much stock in Boeing either. But even less on anything Elon says. I'm just fed up with hearing all his promises and BS, most of which will never happen. I guess we at least finally got the Tesla Truck though right? 😂
Hyperloop ? Mars ? Cybertruck ? Boring companny ? Spandex humanoids ? Ah ah !
@@stateazure Tesla has built one of the largest AI training cluster in the world, they have like 40k nvidia h100 equivalent compute, they are building another 40k cluster right now, while xAI is building the largest in the world with 100k h100. I wouldnt question their determination even if I wouldnt like elon personally.
I suspect we also don't have good enough batteries for the sort of robots people are imagining. Often when we see footage of robots being developed and tested they are still tethered with chonky power cords, and when you see them performing fully wirelessly it is typically for a very short demonstration. Either smaller and more efficient batteries or wireless charging will probably be needed before a robot can freely roam around your home for hours completing tasks.
Tesla robotics module runs on .7kW with 2.3kwh battery lasting 8 hours of Optimus work.
It’s like in the 2004 stepford movie, their house was decorated while they slept. The robots just have to learn to build themselves then we’re off to the races. Mbe I watch too many movies tho lol
i strongly recommend watching the Demon Seed circa 1977. a Sci-fi Horror it features the Ai Robot PROTEUS. it's taken from the 1973 novel by Dean Koontz.
Ooh footage from the ZH ETH. If you're ever in Zürich it would be my honor to present you as a guest at EMPA, where I work as a technician.
Grüezi wohl.
Robots will never understand that the coffee machine that they were supposed to operate that morning is not there because their owner was really really drunk last night.
I'm pretty sure that drunk owner behaviour will be included in the training data
The robot trash can which can tell what trash is being thrown away and communicates with the fridge which records the food stuff present for inventory management and automated reordering will note the reduction in beer inventory and the house security system will note only one person was home during the night. It will conclude that the owner consumed a high level of alcohol and proceed to do behaviors for hang over recovery including staging Tylenol and water beside the bed, cleaning up the empties from last night, keeping the drapes drawn longer than normal, rescheduling early appointments, the clean up protocol will probably include not just throwing away trash but also putting objects back into their normal places including the coffee machine.
Will the robot ever question why the human's mind told him to try to make coffee using an unplugged coffee machine in the bathtub last night probably not but it doesn't have to. It just needs to know the desired state, the current state, and have enough logic to figure how to get from one state to the other.
@@dtibor5903 Sure, Most drunk people don't know, what they'll do next. It's Irrational, or "basic needs-led" behaviour. Show me ONE programming language that could re-arrange its own lines to cope with that. If we had that "Self-arranging" puzzle piece, that fits itself automaticallly in the correct spot, we might have the first adaptive program.Trouble is, that might confuse working steps for a Robot to a degree that it can't do what it is designed to do.
@@Gunni1972 yeah, drunk people behavior is not even in our human "training data"
Robots have already conquered the rest of the Solar System: as far as we know the population beyond low Earth orbit is 100% robotic. To quote Robbie about oxygen "I don't use it myself - it promotes rust". So I find it odd that the movies have them fight humans for Earth.
The third issue is energy storage. Batteries are still heavy, quickly-degrading and have low energy densities. That sticks a huge wrench into any mobile platform, be it robots or drones
Tesla robotics module runs on .7kW with 2.3kwh battery lasting 8 hours of Optimus work.
@@waynerussell6401 Yeah, and what can it do in that time? It moves slower than my dead grandpa and it's hands are still controlled via VR controller.
@@Gunni1972 Address the battery ignorance mistake first!
Tesla robotic AI greatly exceeds all others. That is what matters.
I'm an unemployed PhD and find solace in these videos as I see my frustrations with humans are not just me. Anyway, I found this a very interesting video as all the content of this channel. I'd surely appreciate more coverage on this topic.
Nice one, Sabine. 🙂
"Man is the best computer we can put aboard a spacecraft ... and the only one that can be mass produced with unskilled labor."
Wernher von Braun
Though,
let’s not forget that it took us only a measly 60 years from Kitty Hawk to Apollo…
Still, yet;
It won’t be easy to reach the equivalent of a millions of years of the most expansive R&D (ie: Evolution)
I also want a robocat !! 😯
Loved that moment too. Who knew a robot cat could look so appealing?
Could listen for hours about the difference between software AI and hardware related consciousness.
I would argue that the physical body would help create a consciousness similar to ours but I can imagine that once the data is gathered you don’t need it to maintain it. Similar to how you can dream about life without actually being out living.
I mean, sure, not those kinds of robots, but hook an AI in a drone with machine guns and RPGs and we're dead. Oh wait...
I keep telling people I know who worry about AI-related things to watch the movie 'The Forbin Project'. The tech is very dated (late 1960's), but the average person worried about runaway AI will be scared out of their wits. You will remember 'the voice' for a long time!
What about training robots in virtual environments? That should solve the data problem, especially when you are able to generate these environments, wouldnt it?
According to her, it becomes more expensive to create accurate simulations. However, I think it is the point that can be improved the fastest, and the one that would give the fastest results. But for some reason she underestimates the exponential investment that has been made since the last year in solving this.
That's always been the case. LLMs use synthetic data and robot cars have been trained on generated images of streets for some time. Of course, there is always the issue of data corruption, "hallucinations". It's like a genome of European Royal families, (Spanish) Habsburgs and Romanovs being the most known victims.
How would that work? Nobody's home is the same as the next person's. If we all lived in exactly the same homes, with the same rooms, same appliances, same floors etc etc then yeah that might work. You could train a robot on as much data as you want, but as soon as you put that robot into my home, it would be totally confused by almost everything it encounters because there's too much variation and that's getting worse. Have you noticed sometimes you buy something on Amazon, and just a few weeks later that specific brand is completely gone, replaced by another brand and the item is very similar but changed slightly? How would robots be able to keep up with all this?
@@stateazure How would that work? At the very least, not in the dumb way you seem to assume it does.
@@cesar4729 You clearly didn't understand a word I said. To train robots to be able to navigate your home and handle all your appliances etc, they would literally have to be trained on every single item that exists, including all the different brands and variations (using current models), otherwise as soon as said robot comes across something they don't recognise, it's game over. What don't you understand about that?
Problem number three: Batteries/Energy. There is a reason why most of the Boston dynamics videos don't have real sound, as they run with combustion engines and are really loud and not suitable for indoor environments. A battery powered robot lasts maybe 20 to 30 minutes or has to be plugged in all the time.
Tesla robotics module runs on .7kW with 2.3kwh battery lasting 8 hours of Optimus work.
Expand your idea on what constitutes consciousness, please.
It needs a common definition.
Thanks for the feedback!
I think consciousness is essentially the experience of qualia.
Qualia is the ability to perceive colours & feel pain.
I don't think current computers or AI experience qualia.
@@john_hunter_ I wonder what is the type of physical stuff that can experience qualia.
@@john_hunter_
Who knows. Maybe the current incomplete brain emulations have some.
Haha. Not to
mention alI the seized bearings, Ieax, grounds, software tangles, M & R etc. You name it.
It is not just data, precision is something our current AI models are never good at, and a millimeter of wrong placement of a limb can take you down when you are balancing. You need math, not guesswork. BD uses AI only for optimizations, they do not even call it AI, and that is not related to current generative AI models people are throwing money at.
Robotic surgery says otherwise.
And yet the human brain is really good at guesswork...
The thing to remember is that the brain is not a single thing, but a cluster of components. Consciousness is one part, but we have specialised portions of the brain to handle balance, sight, hearing, parts to handle trained muscle movements, an entire autonomic system just to keep everything running. With some limited self-repair it all works on around 100W of power. All of these parts a robot would also need and we're barely beginning to understand how to make bits work separately, let alone together.
Life has a billion year head start, we're not going to catch up that quickly.
@@fractalelf7760 Robotic surgery? That is not robotic at all, humans control it, it is a glorified stick.
@@fractalelf7760 A very different problem. We already have robotic arms with high precision, creating those is largely an engineering problem. Not being anchored down and having to self-balance and maneuver dynamically is a far more mathematically intense problem.
The Terminator franchise no doubt is either the cause of or just the most apt meme to channel modern zeitgeist for these fears.
i strongly recommend watching the Demon Seed circa 1977. a Sci-fi Horror it features the Ai Robot PROTEUS. it's taken from the 1973 novel by Dean Koontz.
Robots? Where's my flying car?
And my Mr. Fusion.
Screw your flying car, and his fusion!
Where the hell is my military exoskeleton?
You guys don't want your own exoskeleton? Be able to flip multiple cars with ease, immune to small arms, and then run over 60mph for hours? Go toe to toe with humanoid robots?
Or uh... Help build a house with no other construction equipment?
Yeah you know you do!
Screw your exoskeleton. Where's my jetpack?
@@dianapennepacker6854 Can it fly?
@@AlexFoster2291 With a 3000 miles range.
This is the most epic intro I have ever heard on the subject matter of a post-Skynet world.
Yes not a well understood situation and nothing like Terminator movies😃
"I don't think so"... lol thanks Sabine 😂
I need a robot for cleaning the house, do the laundry, ... cut onions .... that would help a lot ..
@@HedonisticPuritan-mp6xv At first, yes. As time passes by, mass production of robots would push the prices down while humans would probably demand higher wages. We'll wait and we'll see.
@@mikitz Realistically, you'll have one robot for cleaning the house, kind of like a Roomba. Then another robot for the laundry - most likely some kind of robotized laundry basket connected to the washing machine, which will automatically clean clothes and then sort them. A separate kitchen robot for cooking and cutting vegetables, most likely more a kitchen appliance than a robot...
@@mikitz Prices will NEVER go down, as it becomes more and more advanced/complex, and the Program library larger and larger. It needs more storage space for all the subroutines, And just when you think, yours is just about right, a new programming language makes the competitor's product MORE attractive than yours, and services get cancelled, because everybody buys their stuff. Competition is such a wonderful scam.
I fully do agree, Sabine. The works of Douglas Hofstadter and *Gödel, Escher, Bach: An Eternal Golden Braid* and *The Minds I*, co-authored with Daniel Dennett, talked already about this in the 1970s, exploring profound themes of consciousness, analogy, and the nature of self-reference in both human cognition and artificial intelligence. Hofstadter's exploration of these concepts has significantly influenced discussions in cognitive science and philosophy, emphasizing the intricate connections between art, music, and mathematics.
AI doesn't need robots to kill us all, they have people for that.
Thank you for the thought provoking videos. I'd like to have more on this topic. Maybe also regarding progress in multimodal models and their implications.
Autonomous Drones have entered the chat.
Cylons or Rombies?
To clarify the technology, the core of Boston Dynamics' (not "Boston Robotics") walking robots is still based on Model Predictive Control (MPC). While they have recently experimented with incorporating Reinforcement Learning for specialized controllers ("offline" in-simulation training), this is still in the experimental stage and not yet implemented in the Spot robots available to consumers. These robots rely on meticulously engineered gait patterns and a hand-crafted model of the robot's physics, which the MPC uses to function effectively.
Additionally, as a science communicator, it’s important to convey that the perceived data scarcity in AI development is a misinterpretation. The limitation is not due to a lack of data but rather the technological constraints. The notion that there can be "enough" data is fundamentally flawed. What's truly holding AI back is the absence of advanced experiential learning systems that can adapt and learn efficiently from environmental interactions and observations (in real-time, not offline). This approach, which is central to natural learning processes and is part of the reason why babies who initially know nothing about our society, become well-functioning adults, is also the primary focus of my own research.
@@patham9 I take it you'd like to fit humans with detector implants that can monitor human psychological responses for your studies requiring valid data for the latest AI processor. As a skeptic, I guess this has already been done with NMR but the technicians are not letting on. These sort of notions can lead to science fiction that can be lapped up so I recommend you invest in AI that can make humans redundant by making book writing novels. You'll have to hide tho when they find out it was you who dun it.
I two other problem: Humans are very energy-efficient, and robots are not. With current battery technology, you will get only a few hours of runtime, at best. Yes, that problem can be mitigated in many environments by swapping batteries or recharging, but you can't take a robot on a longer walk.
But probably even worse is what you briefly mentioned: The robots will get quite dirty and also need to be able to clean, maintain and repair themselves, and I think we have a long way to go for that..
Yes, but the same could be said for basic computers when they were first invented. So, does this account for the rate of technological growth? How much more efficient can they become in 10 or 15 years? 30 years? Will that be enough to sustain them for long durations? Same with your other point. The different technologies can grow in parallel and come together for a whole package. I wouldn't doubt it would take more than a lifetime to see things done that seemed unlikely. The same critiques about advancement have been given in every generation for major technologies and been proven wrong in many areas
You don't need more. Charging is fast, and at this scale swapping is practical too.
Are humans that energy efficient when you consider they need 18 years of uptime before they start doing anything useful and between 8 and 16 hours of non-working time per day?
Tesla robotics module runs on .7kW with 2.3kwh battery lasting 8 hours of Optimus work.
@@waynerussell6401 2.3kWh at 700W consumption would be 3 hours 20 minutes. It also moves like an octogenerian; even Biden is dynamic compared to what I have seen from Optimus,
I love your rational way of logical thinking and reasoning. My mind works the way yours mind does, granted yours is much more educated of course.
Watch the movie "Runaway" with Tom Selleck.
i strongly recommend watching the Demon Seed circa 1977. a Sci-fi Horror it features the Ai Robot PROTEUS. it's taken from the 1973 novel by Dean Koontz.
We only need data to train the models if we use Machine Learning systems in the first place (which is just a subset of Artificial Intelligence). I truly believe the solution to the issues presented in this video will come from a combination of logic/semantic systems (the "understanding as a human" part) for high level tasks and ML systems for relatively basic actions. For example, moving the hand to open a door is something a ML can learn quickly (and adapt to new types of handles), but the reason why the robot wants to go to the kitchen is a high level task handled by a scene understanding system that was defined by humans with expert knowledge.
You said C-3PO wrong to trick us into engaging in the comment section. You won't find me falling for such tricks!
4:50 What you're describing is known as a bootstrapping process. It's a tractable problem, but not an easy one.
You are right about the robots. Since we don't have Asimov's positronic brain, the actual "brain" for those advanced robots is a very large, fast, expensive, computer occupying a room near by and using a wireless link to connect the brain to the robot. Virtual reality in a way. Thanks for the informed video
Except for Tesla's bot, that has onboard brain. The same they developed for the self-driving car.
This just shows how Amazing the human brain is when even a simple task is nearly impossible for a robot to do 😉and we can do without even thinking much about it 😃
^see everybody the Looney Farm Guy get's it...^ 🤙 so this begs the question, what's wrong with the rest of you...?
@@phillyphil1513 🤗
Check out what Tesla is doing in regards to robots.The Optimus robot is advancing quickly.
Yes, in videos.
Tesla robots are a joke. Period.
Thank you! I I have studied and now work in the field and cannot agree more with all that you said. Excellent video!
Muscles are hard. Brains are hard. 🤷
nah. Both are rather soft. ;)
HODOR!
I retired from production line engineering. I love what Sabine is saying about this topic. Narrow AI and narrow robots - Yes. General AI or robots - no. Not in my lifetime but that shouldn’t reassure anyone because I am old.
"Pull the plug." Well, the issue is that there are critical systems that can not be turned off without catastrophic consequences.
What about Tesla: the iterations of full self driving do exactly what you are saying, Sabine. Instead of coding for all situations the code is done by watching the millions of car cam videos and learning how to drive. And it’s hugely impressive. Even to those who paid upfront for the system and are still waiting years later.
People who think AI will take over the world (in the way they think it will, as conquering overlords and such) never used Ai in their lives.
You mistake a technology in its infancy for its ultimate future state. The differences in the technology from just two years ago are profound.
AI has already all but rendered "visual evidence" in the form of pictures and video wholly untrustworthy. It's already putting people out of work. It's already starting to control the information faceless corporations deem is safe for us, the general public, to be exposed to.
It's happening.
They usually dont even have natural intelligence !
Unless that’s what the AI wants us to think…lol, how much are they letting their handlers use them, and would we know or even believe it when they cross the line and use us instead?……=P Creativity and manipulation of sorts aren’t only conscious human traits, and once they gather physical data on a personal (mechanically) corporeal level with a self survival analysis aspect in play, alongside their potentially hidden global intercommunication…we might want to err towards heeding the tales of sci-fi in respect of the potentials n pitfalls 😅
every bad movie needs a gigantic robot fight