Yann LeCun was saying 3 years ago that because token generation is auto-regressive, its mathematically impossible to generate a long stream of coherent tokens, because errors amplify exponentially.
and then models learned that they can back track and error correct
I think it was largely the introduction of tool calling that allowed models to mitigate the issue of errors amplifying exponentially since it allows the model to understand if what it generated is correct or has issues that it needs to address. This addresses the potential lack of or low quality of world model by being able to reference the current state of the world.
I've definitely realized this phenomenon after a few occasions of erroneously trying to rely purely on instructions to get an LLM to do a thing or take on a role, especially without persistent cloud-based sessions that have internal checklists and other opaque guidance. They're essentially poor at self-managing, but can do really well when they are limited in scope/context and are worked into a sort of state machine that guarantees they perform certain tasks predictably. They won't always do those tasks the exact way you expect them to, but at least they actually do them, and because of that they are more likely to have the correct prior context to perform the next task better. Because they are so prone to selectively ignoring directions, that can quickly send them down an incorrect path that compounds on itself.
By self-modifying the software. Currently the model harnesses only allow the model to modify its own prompt (which could be considered a really weak kind of learning), but theoretically, a model could design and train its own replacement and run that, continuously improving itself. I’m not sure if LLMs will be able to do that but the static hardware has nothing to do with it (since the bits on the harddrive aren’t static).
this is profoundly false. AI not only can learn, it is built entirely from learning. The field is called machine learning after all.
Not only that... AI is NOT only learning during the training phase... LLMs learn in real time the minute you talk to it. It learns something and saves those learnings in a context window or somewhere else if you want it to exist beyond the context window.
All of the above runs on static hardware. Don't understand how someone can say a profoundly wrong statement and get voted up.
Correct me if I'm wrong, but if a profound insight is gathered in session 1 with user A and stored in context A1, this might be available to user A in session 2, if that still has access to context A1, but won't be available to user B in any of his/her sessions until that NN is retrained with input which includes at least some of the information from context A1.
Also with the stochastic parrot thing. If you say just the right thing to the right human and the right time, they'll very predictibly say their favorite movie/book quote or song lyric, like some sort of parrot.
Besides "smart", the headline also conflates AI with LLMs. The real, non-clickbait title is "Yann LeCun, founder of AMI Labs, is developing a new AI system"
It is just so bizarre compared to my everyday experience also.
I never ask Opus or Fable a question and think "what a stupid response".
Quite the opposite. It has actually raised the bar of what I consider an intelligent response to my inquiry. So much so that most responses from humans on most subjects to most forms of inquiry seem stupid and not really thought out.
I’m sure you’re very intelligent and capable so I suspect we work in different problem spaces if you have not seen this, but I definitely think the responses are at times very very junior and I find myself having to explain first principles. Fable less so, but Opus routinely will make very naive assumptions about retry logic, protocols it supposedly has training material on, and it will very often miss the forest for the trees.
This isn’t exactly saying how stupid anyone is but I’d definitely have been concerned about a human’s reasoning ability and understanding of logic if they’d given me similar answers.
Everyone nowadays seems to only think of AI as LLMs or maybe also stable diffusion. People want to ban games with AI in them, when by definition every NPC is following some kind of AI algorithm.
They literally interview another person in it and mention a lot of other labs doing this kind of research including Google. Yes, he's the main starting interview but this is not really clickbait or a marketing piece.
The article seems to define "smart" as being good at spatial awareness and navigating a body through 3D space and such. Thus, a mice is smarter than an LLM.
That's the first time in my life I hear this definition. Until now, the word "smart" has meant doing exactly the things LLMs do, and mice don't.
I guess it is a sign we are re-evaluating what makes humans special.
While we should be careful of a bias, it is also a good practice in the scientific method to review definitions that may have been not precise enough.
For example, initially, a "planet" was just a big body in space. Then when people started to see more and more nuances, the definition just refined, and some objects stopped being called "planet".
I would not be surprised if there is a bias that pushes some people to redefine "intelligence" away from machine, but I would not be surprised if there is a bias that pushes some people to ignore newly discovered nuance and put into the same "intelligence" bag things that are in fact very different. I personally can see how LLM are not really "intelligent", and I don't think it is a good idea to say: well, yesterday we said the minimum criteria is X, now that we noticed that X can be reached without really doing the real thing, let's just ignore that and pretend it is the same thing.
(: the biggest clue for me is to use an early model, and see that it sometimes looks very intelligent, and then sometimes you can see that it gets it wrong in a way that shows that it never "understood" it at all. Newer models are better, but because it is an iteration on the same bases, the increase of performances cannot really due to replacing the things that "looked smart by aren't" by "real smart", but more replacing the things that "don't look smart" by "look smart by aren't")
Yeah I think if we are looking at it through that lens, the problem is in the term _intelligence_ in itself. Psychology and biology could not even pinpoint what exactly makes for _intelligence_. There isn't really a precise definition yet so it's just natural that definitions tend to shift.
I don't think we even need to go into tech and AI for an example. The intelligence or lack thereof of pets surprise us. Sometimes a cat is surprisingly smart when it is able to open a door to get food it wasn't supposed to. But then same cat gets bamboozled by walls and simple optical illusions. We generally expect that if something/a human is smart enough to do the former, then it shouldn't be dumb enough to fall for the latter.
Coming back to AI, this dissonance is how AI-generated images are detected for example. If a human can render something so well, you wouldn't expect them to make small but nonetheless elementary line art mistakes.
For me, "smart" means doing things less based on instinct. Things humans can do but mice cannot, things mathematicians can do but normal people cannot, etc.
Considering the unit distance conjecture was disproved by OAI's model last month, I think maybe LLMs should count as "smart".
Is a brain also inference? I know that an LLM works very different from the brain, but I wonder what makes a brain more capable of thinking. Is it the long term context? Is it a different type of neuron activation?
Left behind how? It's been transformers since 2016 and not much actual progress in basic architectures has happened 10 years later. I'm honestly struggling to see how you can be left behind in this field.
Obviously, transformers architecture is just one of the ingredients. Otherwise we wouldn't be seeing competing labs in this race. I also read all his interviews as a marketing material.
We're past the point where there's a feasible argument that there is an AI winter coming.
The models work remarkably well for several classes of problem that seemed impossible a few years ago. They're not going away. There will still absolutely be a lot of ups and down and crazy stuff that happens in AI, but it won't be that AI almost completely stops being developed/funded for a decade or more. The biggest risk, I think, is regulatory capture; it's what Anthropic and OpenAI seem to be aiming for with their scaremongering about how capable and dangerous their models are. That'll put a damper on the industry for everyone except the companies that bribe the right people.
Not likely. Take with whatever grain of salt you'd like, but that was largely a property of development being academicized and subject to things like grant cycles, research topic fashionability trends, and institutional structure. It would be wrong to assume it's some baked in thing that's guaranteed to happen independent of how development looks.
But _AI today_ is heavily subsidized by investor capital in the same way investors subsidized social/mobile/big data/VR/blockchain in the past. It's not unlikely "AI" would get a soft taboo in the same way as if you just presented a mobile-first, big-data driven, VR social media app today.
Which, judging by the terrible PR optics AI has nowadays, could unfortunately seep into academia too. Fund grants wouldn't want their names associated with anything with "AI" in its name even if it's a return to expert systems.
You're mixing different things. Mobile first is integrated into new services to the point that they either are mobile first, or they have a design system which includes mobile as a surface. VR has a wide user base (MQ2 sold as well as the original Xbox) and is involved in both manufacturing design and simulation, and is hardly an academic taboo, even if the "main" topic of discussion is elsewhere right now. Blockchains are being integrated into financial infrastructure even as some people make snarky commentary about it. Sometimes optics is just an optical illusion.
Fair enough. Mobile and social became ubiquitous and are now table stakes. But my problem with VR and blockchain---even allowing for the fact/assumption that they are still relevant---is that they never lived up to their hype. They never became ubiquitous as mobile and social. They don't inspire investor confidence like they did in the past, if at all. AI, if it survives the public and regulatory backlash, could be headed to the same understudy role.
I'm using "AI" broadly here even if the current investor darling is just LLMs because, well, the term AI has been front and center of all promotions and investors and the general consumer public isn't really a discerning bunch. So I stand by my prediction that a "soft taboo" is likely where investors and consumers shy away from anything even remotely AI. The consumer backlash has arguably already started.
The vast consumer adoption and ongoing involvement seems to point the other way, though. I think a lot of the appearance of backlash is on (specifically anglophone, mostly) social media, which is going through a somewhat reactionary phase regardless.
I'd be interested in the "retention rate" for these two products. I wouldn't be surprised if the average original Xbox was used 2 orders of magnitude more than the average Meta Quest, which is collecting dust on some shelf.
I'd wager the typical MQ2 owner is someone with 20 hours of Beat Saber on it and 5000 hours total on Steam or PS.
Yann LeCun was saying 3 years ago that because token generation is auto-regressive, its mathematically impossible to generate a long stream of coherent tokens, because errors amplify exponentially.
and then models learned that they can back track and error correct
so much for "mathematically impossible..."
Stop attacking Yann. I would say like 90% of the HN crowd was parroting Yann too.
I think it was largely the introduction of tool calling that allowed models to mitigate the issue of errors amplifying exponentially since it allows the model to understand if what it generated is correct or has issues that it needs to address. This addresses the potential lack of or low quality of world model by being able to reference the current state of the world.
I've definitely realized this phenomenon after a few occasions of erroneously trying to rely purely on instructions to get an LLM to do a thing or take on a role, especially without persistent cloud-based sessions that have internal checklists and other opaque guidance. They're essentially poor at self-managing, but can do really well when they are limited in scope/context and are worked into a sort of state machine that guarantees they perform certain tasks predictably. They won't always do those tasks the exact way you expect them to, but at least they actually do them, and because of that they are more likely to have the correct prior context to perform the next task better. Because they are so prone to selectively ignoring directions, that can quickly send them down an incorrect path that compounds on itself.
> and then models learned that they can back track and error correct
You mean "Human developers learned ...", yes? Or was there really an all AI-driven, self-improving aspect to this?
Well, LLM networks don't have a 'back track and error correct' component in the design, AFAIK.
You insinuate here AI "learned".
I doubt that this was AI self-improvement.
Was there a particular change to the network or the way that it was trained that introduced the 'backtrack and error correct' mechanism?
do you have a problem with this field of research being called "machine learning"?
Also, almost any argument against LLM intelligence also applies to humans.
I very commonly see someone make some small mistake and end up going in the wrong direction, “accumulating stupid” as they go, sometimes for years.
Humans can learn.
AI can not.
For those disagreeing: please explain how a static hardware can learn.
By self-modifying the software. Currently the model harnesses only allow the model to modify its own prompt (which could be considered a really weak kind of learning), but theoretically, a model could design and train its own replacement and run that, continuously improving itself. I’m not sure if LLMs will be able to do that but the static hardware has nothing to do with it (since the bits on the harddrive aren’t static).
idk, how does voice recognition learn my voice? How can I install programs when the hardware is static?
this is profoundly false. AI not only can learn, it is built entirely from learning. The field is called machine learning after all.
Not only that... AI is NOT only learning during the training phase... LLMs learn in real time the minute you talk to it. It learns something and saves those learnings in a context window or somewhere else if you want it to exist beyond the context window.
All of the above runs on static hardware. Don't understand how someone can say a profoundly wrong statement and get voted up.
Correct me if I'm wrong, but if a profound insight is gathered in session 1 with user A and stored in context A1, this might be available to user A in session 2, if that still has access to context A1, but won't be available to user B in any of his/her sessions until that NN is retrained with input which includes at least some of the information from context A1.
Also with the stochastic parrot thing. If you say just the right thing to the right human and the right time, they'll very predictibly say their favorite movie/book quote or song lyric, like some sort of parrot.
[dead]
Besides "smart", the headline also conflates AI with LLMs. The real, non-clickbait title is "Yann LeCun, founder of AMI Labs, is developing a new AI system"
It is just so bizarre compared to my everyday experience also.
I never ask Opus or Fable a question and think "what a stupid response".
Quite the opposite. It has actually raised the bar of what I consider an intelligent response to my inquiry. So much so that most responses from humans on most subjects to most forms of inquiry seem stupid and not really thought out.
I’m sure you’re very intelligent and capable so I suspect we work in different problem spaces if you have not seen this, but I definitely think the responses are at times very very junior and I find myself having to explain first principles. Fable less so, but Opus routinely will make very naive assumptions about retry logic, protocols it supposedly has training material on, and it will very often miss the forest for the trees.
This isn’t exactly saying how stupid anyone is but I’d definitely have been concerned about a human’s reasoning ability and understanding of logic if they’d given me similar answers.
Everyone nowadays seems to only think of AI as LLMs or maybe also stable diffusion. People want to ban games with AI in them, when by definition every NPC is following some kind of AI algorithm.
They literally interview another person in it and mention a lot of other labs doing this kind of research including Google. Yes, he's the main starting interview but this is not really clickbait or a marketing piece.
The article seems to define "smart" as being good at spatial awareness and navigating a body through 3D space and such. Thus, a mice is smarter than an LLM.
That's the first time in my life I hear this definition. Until now, the word "smart" has meant doing exactly the things LLMs do, and mice don't.
I guess it is a sign we are re-evaluating what makes humans special.
> I guess it is a sign we are re-evaluating what makes humans special.
Always has been: https://en.wikipedia.org/wiki/AI_effect
Tangentially: https://en.wikipedia.org/wiki/Moravec%27s_paradox
While we should be careful of a bias, it is also a good practice in the scientific method to review definitions that may have been not precise enough.
For example, initially, a "planet" was just a big body in space. Then when people started to see more and more nuances, the definition just refined, and some objects stopped being called "planet".
I would not be surprised if there is a bias that pushes some people to redefine "intelligence" away from machine, but I would not be surprised if there is a bias that pushes some people to ignore newly discovered nuance and put into the same "intelligence" bag things that are in fact very different. I personally can see how LLM are not really "intelligent", and I don't think it is a good idea to say: well, yesterday we said the minimum criteria is X, now that we noticed that X can be reached without really doing the real thing, let's just ignore that and pretend it is the same thing.
(: the biggest clue for me is to use an early model, and see that it sometimes looks very intelligent, and then sometimes you can see that it gets it wrong in a way that shows that it never "understood" it at all. Newer models are better, but because it is an iteration on the same bases, the increase of performances cannot really due to replacing the things that "looked smart by aren't" by "real smart", but more replacing the things that "don't look smart" by "look smart by aren't")
Yeah I think if we are looking at it through that lens, the problem is in the term _intelligence_ in itself. Psychology and biology could not even pinpoint what exactly makes for _intelligence_. There isn't really a precise definition yet so it's just natural that definitions tend to shift.
I don't think we even need to go into tech and AI for an example. The intelligence or lack thereof of pets surprise us. Sometimes a cat is surprisingly smart when it is able to open a door to get food it wasn't supposed to. But then same cat gets bamboozled by walls and simple optical illusions. We generally expect that if something/a human is smart enough to do the former, then it shouldn't be dumb enough to fall for the latter.
Coming back to AI, this dissonance is how AI-generated images are detected for example. If a human can render something so well, you wouldn't expect them to make small but nonetheless elementary line art mistakes.
It depends on how you define "smart".
For me, "smart" means doing things less based on instinct. Things humans can do but mice cannot, things mathematicians can do but normal people cannot, etc.
Considering the unit distance conjecture was disproved by OAI's model last month, I think maybe LLMs should count as "smart".
It's inference. It's really good at generating stuff when the example base is extensive. Like for non-esoteric coding.
Is a brain also inference? I know that an LLM works very different from the brain, but I wonder what makes a brain more capable of thinking. Is it the long term context? Is it a different type of neuron activation?
i guess inference engineering, like dpsark or dflash specific speculative decoding technqiues
Ha before reading the article I thought "this must be an interview of Lecun". A bitter scientist that hates he was left behind the revolution.
Considering all of the great research that has come from his labs (eg. DINO, Segment Anything) I don’t think that’s fair (no pun intended).
In what way was he left behind? If he wanted to actually work on LLMs all the AI labs would fight to get him
Left behind how? It's been transformers since 2016 and not much actual progress in basic architectures has happened 10 years later. I'm honestly struggling to see how you can be left behind in this field.
Obviously, transformers architecture is just one of the ingredients. Otherwise we wouldn't be seeing competing labs in this race. I also read all his interviews as a marketing material.
and CPUs have the same basic architecture since 2000. no progress happened, right?
What's next is more AI spam-slop. I already noticed this on youtube. Tons of short videos are AI sloppified, making youtube worse in the process.
AI winter.
We're past the point where there's a feasible argument that there is an AI winter coming.
The models work remarkably well for several classes of problem that seemed impossible a few years ago. They're not going away. There will still absolutely be a lot of ups and down and crazy stuff that happens in AI, but it won't be that AI almost completely stops being developed/funded for a decade or more. The biggest risk, I think, is regulatory capture; it's what Anthropic and OpenAI seem to be aiming for with their scaremongering about how capable and dangerous their models are. That'll put a damper on the industry for everyone except the companies that bribe the right people.
Not likely. Take with whatever grain of salt you'd like, but that was largely a property of development being academicized and subject to things like grant cycles, research topic fashionability trends, and institutional structure. It would be wrong to assume it's some baked in thing that's guaranteed to happen independent of how development looks.
But _AI today_ is heavily subsidized by investor capital in the same way investors subsidized social/mobile/big data/VR/blockchain in the past. It's not unlikely "AI" would get a soft taboo in the same way as if you just presented a mobile-first, big-data driven, VR social media app today.
Which, judging by the terrible PR optics AI has nowadays, could unfortunately seep into academia too. Fund grants wouldn't want their names associated with anything with "AI" in its name even if it's a return to expert systems.
You're mixing different things. Mobile first is integrated into new services to the point that they either are mobile first, or they have a design system which includes mobile as a surface. VR has a wide user base (MQ2 sold as well as the original Xbox) and is involved in both manufacturing design and simulation, and is hardly an academic taboo, even if the "main" topic of discussion is elsewhere right now. Blockchains are being integrated into financial infrastructure even as some people make snarky commentary about it. Sometimes optics is just an optical illusion.
Fair enough. Mobile and social became ubiquitous and are now table stakes. But my problem with VR and blockchain---even allowing for the fact/assumption that they are still relevant---is that they never lived up to their hype. They never became ubiquitous as mobile and social. They don't inspire investor confidence like they did in the past, if at all. AI, if it survives the public and regulatory backlash, could be headed to the same understudy role.
I'm using "AI" broadly here even if the current investor darling is just LLMs because, well, the term AI has been front and center of all promotions and investors and the general consumer public isn't really a discerning bunch. So I stand by my prediction that a "soft taboo" is likely where investors and consumers shy away from anything even remotely AI. The consumer backlash has arguably already started.
The vast consumer adoption and ongoing involvement seems to point the other way, though. I think a lot of the appearance of backlash is on (specifically anglophone, mostly) social media, which is going through a somewhat reactionary phase regardless.
> MQ2 sold as well as the original Xbox
I'd be interested in the "retention rate" for these two products. I wouldn't be surprised if the average original Xbox was used 2 orders of magnitude more than the average Meta Quest, which is collecting dust on some shelf.
I'd wager the typical MQ2 owner is someone with 20 hours of Beat Saber on it and 5000 hours total on Steam or PS.
AI climate change.
Human winter.
Its definitely not as dumb as MAGA crowd
Was it worth it to log out and create a new account just to post this?