Making LeCun report to Wang was the most boneheaded move imaginable. But… I suppose Zuckerberg knows what he wants, which is AI slopware and not truly groundbreaking foundation models.
replies(20):
If you think LLMs are not the future then you need to come with something better
If you have a theoretical idea that's great, but take to at least GPT2 level first before writing off LLMs
Theoretical people love coming up with "better ideas" that fall flat or have hidden gotchas when they get to practical implementation
As Linus says, "talk is cheap, show me the code".
Are all critiques of the obvious decline in physical durability of American-made products invalid unless they figure out a solution to the problem? Or may critics of a subject exist without necessarily being accredited engineers themselves?
If you want to predict future text, you use an LLM. If you want to predict future frames in a video, you go with Diffusion. But what both of them lack is object permanence. If a car isn't visible in the input frame, it won't be visible in the output. But in the real world, there are A LOT of things that are invisible (image) or not mentioned but only implied (text) that still strongly affect the future. Every kid knows that when you roll a marble behind your hand, it'll come out on the other side. But LLMs and Diffusion models routinely fail to predict that, as for them the object disappears when it stops being visible.
Based on what I heard from others, world models are considered the missing ingredient for useful robots and self-driving cars. If that's halfway accurate, it would make sense to pour A LOT of money into world models, because they will unlock high-value products.
Corporate R&D teams are there to absorb risk, innovate, disrupt, create new fields, not for doing small incremental improvements. "If we know it works, it's not research." (Albert Einstein)
I also agree with LeCun that LLMs in their current form - are a dead end. Note that this does not mean that I think we have already exploited LLMs to the limit, we are still at the beginning. We also need to create an ecosystem in which they can operate well: for instance, to combine LLMs with Web agents better we need a scalable "C2B2C" (customer delegated to business to business) micropayment infrastructure, because as these systems have already begun talking to each other, in the longer run nobody would offer their APIs for free.
I work on spatial/geographic models, inter alia, which by coincident is one of the direction mentioned in the LeCun article. I do not know what his reasoning is, but mine was/is: LMs are language models, and should (only) be used as such. We need other models - in particular a knowledge model (KM/KB) to cleanly separate knowledge from text generation - it looks to me right now that only that will solve hallucination.
Messing with the logic in the loop and combining models has an enormous potential, but it's more engineering than researching, and it's just not the sort of work that LeCun is interested in. I think the conflict lies there, that Facebook is an engineering company, and a possible future of AI lies in AI engineering rather than AI research.
Maybe at university, but not at a trillion dollar company. That job as chief scientist is leading risky things that will work to please the shareholders.
And while we've been able to approximate the world behind the words, it's just full of hallucinations because the AI's lack axiomatic systems beyond much manually constructed machinery.
You can probably expand the capabilties by attaching to the front-end but I suspect that Yann is seeing limits to this and wants to go back and build up from the back-end of world reasoning and then _among other things_ attach LLM's at the front-end (but maybe on equal terms with vision models that allows for seamless integration of LLM interfacing _combined_ with vision for proper autonomous systems).
Everything from the sorites paradox to leaky abstractions; everything real defies precise definition when you look closely at it, and when you try to abstract over it, to chunk up, the details have an annoying way of making themselves visible again.
You can get purity in mathematical models, and in information systems, but those imperfectly model the world and continually need to be updated, refactored, and rewritten as they decay and diverge from reality.
These things are best used as tools by something similar to LLMs, models to be used, built and discarded as needed, but never a ground source of truth.
If the answer is yes, then better to keep him, because he has already proved himself and you can win in the long-term. With Meta's pockets, you can always create a new department specifically for short-term projects.
If the answer is no, then nothing to discuss here.
The last time LeCun disagreed with the AI mainstream was when he kept working on neural net when everyone thought it was a dead end. He might be entirely right in his LLM scepticism. It's hardly a surefire path. He didn't prevent Meta from working on LLM anyway.
The issue is more than his position is not compatible with short term investors expectations and that's fatal in a company like Meta at the position LeCun occupies.
If you follow LeCun on social media, you can see that the way FAIR’s results are assessed is very narrow-minded and still follows the academic mindset. He mentioned that his research is evaluated by: "Research evaluation is a difficult task because the product impact may occur years (sometimes decades) after the work. For that reason, evaluation must often rely on the collective opinion of the research community through proxies such as publications, citations, invited talks, awards, etc."
But as an industry researcher, he should know how his research fits with the company vision and be able to assess that easily. If the company's vision is to be the leader in AI, then as of now, he seems to have failed that objective, even though he has been at Meta for more than 10 years.
Yes but he was hired in the ZIRP era where all SV companies were hiring every opinionated academic and giving them free reign and unlimited money to burn in the hopes that maybe they'll create the next big thing for them eventually.
These are very different economic times right now, after the FED infinite money glitch has been patched out, so now people do need to adjust to them and start actually making some products of value for their seven figure costs to their employers, or end up being shown the door.
Frontier models are all profitable. Inference is sold with a damn good margin, and the amounts of inference AI companies sell keeps rising. This necessitates putting more and more money into infrastructure. AI R&D is extremely expensive too, and this necessitates even more spending.
A mistake I see people make over and over again is keeping track of the spending but overlooking the revenue altogether. Which sure is weird: you don't get from $0B in revenue to $12B in revenue in a few years by not having a product anyone wants to buy.
And I find all the talk of "non-deterministic hallucinatory nature" to be overrated. Because humans suffer from all of that too, just less severely. On top of a number of other issues current AIs don't suffer from.
Nonetheless, we use human labor for things. All AI has to do is provide a "good enough" alternative, and it often does.
Also, like… it’s Facebook. It has a history of ploughing billions into complete nonsense (see metaverse). It is clearly not particularly risk averse.
Cracking that is a huge step, pure multi-modal trained models will probably give us a hint, but I think we're some ways from seeing a pure multi-modal open model which can be pulled apart/modified. Even then they're still train and deploy not dynamically learning. I worry we're just going to see LSTM design bolted onto deep LLM because we don't know where else to go and it will be fragile and take eons to train.
And less said about the crap of "but inference is doing some kind of minimization within the context window" the better, it's vacuous and not where great minds should be looking for a step forwards.
Oh god, that is massively under-selling their learning ability. These models are able to extract and reply with why jokes are funny without even knowing basic vocab, yet there are pure-code models out there with lingual rules baked in from day one which still struggle with basic grammar.
The _point_ of LLMs arguably is there ability to learn any pattern thrown at it with enough compute. With an exception to learning how logical processes work, and pure LLMs only see "time" in the sense of a paragraph begins and ends.
At the least they have taught computers, "how to language", which in regards to how to interact with a machine is a _huge_ step forward.
Unfortunately the financial incentives are split between agentic model usage (taking the idea of a computerised butler further), maximizing model memory and raw learning capacity (answering all problems at any time), and long-range consistency (longer ranges give better stable results due to a few reasons, but we're some way from seeing an LLM with a 128k experts and 10e18 active tokens).
I think in terms of building the perfect monkey butler we already have most or all of the parts. With regard to a model which can dynamically learn on the fly... LLMs are not the end of the story and we need something to allow the models to more closely tie their LS with the context. Frankly the fact that DeepSeek gave us an LLM with LS was a huge leap since previous model attempts had been overly complex and had failed in training.
I suppose they could solve superintelligence and cure cancer and build fusion reactors with it, but that's 100% outside their comfort zone - if they manage to build synthethic conversation partners and synthethic content generators as good or better than the real thing the value of having every other human on the planet registered to one of their social network goes to zero.
Which is impossible anyway - I facebook to maintain real human connections and keep up with people who I care about, not to consume infinite content.
/s
In the software development world yes, outside of that, virtually none. Yes, you can transcribe a video call in Office, yes, but that's not ground breaking. I dare you to list 10 impacts on different fields, excluding tech and including at least half blue collar fields and at least half white collar fields , at different levels from the lowest to the highest in the company hierarchy, that LLM/Diffusion models are having. Impact here specifically means a significant reduction of costs or a significant increase of revenue. Go on
Even when writing, it shifts the mental burden from an easy thing (writing code) to a very hard thing (reading that code, validating it's right, hallucination free, and then refactoring it to match your teams code style and patterns).
It's great for building a first-order approximation of a tech demo app that you then throw out and build from scratch, and auto-complete. In my experience, anyways. I'm sure others have had different experiences.
Starting with the sophomoric questions of the optimist who mistakes the possible for the viable: how definite of a thing is "the world", how knowable is it, what is even knowledge... and then back through the more pragmatic: by whom is it knowable, to what degree, and by what means. The mystics: is "the world" the same thing as "the sum of information about the world"? The spooks: how does one study those fields of information which are already agentic and actively resist being studied by changing themselves, such as easily emerge anywhere more than n(D) people gather?
Plenty of food for thought from why ontologies are/aren't a thing. The classical example of how this plays out in the market being search engines winning over internet directories. But that's one turn of the wheel. Look at what search engines grew into quarter century later. What their outgrowths are doing to people's attitude towards knowledge. Different timescale, different picture.
Fundamentally, I don't think human language has sufficient resolution to model large spans of reality within the limited human attention span. The physical limits of human language as information processing device have been hit at some point in the XX century. Probably that 1970s divergence between productivity and wages.
So while LLMs are "computers speak language now" and it's amazing if sad that they cracked it by more data and not by more model, what's more amazing is how many people are continually ready to mistake language for thought. Are they all P-zombies or just obedience-conditioned into emulating ones?!?!?
Practically, what we lack is not the right architecture for "big knowing machine", but better tools for ad-hoc conceptual modeling of local situations. And, just like poetry that rhymes, this is exactly what nobody has a smidgen of interest to serve to consumers, thus someone will just build it in their basement in the hope of turning the tables on everyone. Probably with the help of LLMs as search engines and code generators. Yall better hurry. They're almost done.
if OpenAI can build a "social" network of completely generated content, that can kill Meta. Even today I venture to guess that most of the engagements in their platforms is not driven by real friends, so an AI driven platform won't be too different, or it might make content generation be so easy as to make your friends engage again.
Apart from it the ludicrous vision of the metaverse seems much more plausible with highly realistic world models
Text and languages contain structured information and encode a lot of real-world complexity (or it's "modelling" that).
Not saying we won't pivot to visual data or world simulations, but he was clearly not the type of person to compete with other LLM research labs, nor did he propose any alternative that could be used to create something interesting for end-users.
But that sure didn't happen.
I really resonate with his view due to my background in physics and information theory. I for one welcome his new experimentation in other realms while so many still hack away at their LLMs in pursuit of SOTA benchmarks.
If that content becomes even cheaper, of higher quality and highly tailored to you, that is probably worth a lot of money, or at least worth not losing your entire company by a new competitor
The future is here folks, join us as we build this giant slop machine in order to sell new socks to boomers.
Its pretty much dog eat dog at top management positions.
Its not exactly a space for free thinking timelines.
Is the real bubble ignorance? Maybe you'll cool down but the rest of the world? There will just be more DeepSeek and more advances until the US loses its standing.
When did they make groundbreaking foundation models though? DeepMind and OpenAI have done plenty of revolutionary things, what did Meta AI do while being led by LeCun?
[1] Doctor of Philosophy:
This is why we're losing innovation.
Look at electric cars, batteries, solar panels, rare earths and many more. Bubble or struggle for survival? Right, because if US has no AI the world will have no AI? That's the real bubble - being stuck in an ancient world view.
Meta's stock has already tanked for "over" investing in AI. Bubble, where?
But the skill sets to avoid and survive personnel issues in academia is different from industry. My 2c.
The US government basically forced AT&T to use revenue from its monopoly to do fundamental research for the public good. Could the government do the same thing to our modern megacorps? Absolutely! Will it? I doubt it.
https://www.nytimes.com/1956/01/25/archives/att-settles-anti...
You assume that's the only use of it.
And are people not using these code generators?
Is this an issue with a lost generation that forgot what Capex is? We've moved from Capex to Opex and now the notion is lost, is it? You can hire an army of software developers but can't build hardware.
Is it better when everyone buys DeepSeek or a non-US version? Well then you don't need to spend Capex but you won't have revenue either.
It seems they've given up on the research and are now doubling down on LLMs.
We have estimates that range from 30% to 70% gross margin on API LLM inference prices at major labs, 50% middle road. 10% to 80% gross margin on user-facing subscription services, error bars inflated massively. We also have many reports that inference compute has come to outmatch training run compute for frontier models by a factor of x10 or more over the lifetime of a model.
The only source of uncertainty is: how much inference do the free tier users consume? Which is something that the AI companies themselves control: they are in charge of which models they make available to the free users, and what the exact usage caps for free users are.
Adding that up? Frontier models are profitable.
This goes against the popular opinion, which is where the disbelief is coming from.
Note that I'm talking LLMs rather than things like image or video generation models, which may have vastly different economics.
Damn did you just invent that? That's really catchy.
In any case if I have to guess, we will see shallow things like the Sora app, a video generation tiktok social network and deeper integration like fake influencers, content generation that fits your preferences and ad publishers preferences
a more evil incarnation of this might be a social network where you aren't sure who is real and who isn't. This will probably be a natural evolution of the need to bootstrap a social network with people and replacing these with LLMs
"Why Bell Labs Worked" [1]
"The Influence of Bell Labs" [2]
"Bringing back the golden days of Bell Labs" [3]
"Remembering Bell Labs as legendary idea factory prepares to leave N.J. home" [4] or
"Innovation and the Bell Labs Miracle" [5]
interesting too.
[1] https://news.ycombinator.com/item?id=43957010 [2] https://news.ycombinator.com/item?id=42275944 [3] https://news.ycombinator.com/item?id=32352584 [4] https://news.ycombinator.com/item?id=39077867 [5] https://news.ycombinator.com/item?id=3635489
I don't disagree that the world is full of fuzziness. But the problem I have with this portrayal is that formal models are often normative rather than analytical. They create reality rather than being an interpretation or abstraction of reality.
People may well have a fuzzy idea of how their credit card works, but how it really works is formally defined by financial institutions. And this is not just true for software products. It's also largely true for manufactured products. Our world is very much shaped by artifacts and man-made rules.
Our probabilistic, fuzzy concepts are often simply a misconception. That doesn't mean it's not important of course. It is important for an AI to understand how people talk about things even if their idea of how these things work is flawed.
And then there is the sort of semi-formal language used in legal or scientific contexts that often has to be translated into formal models before it can become effective. Law makers almost never write algorithms (when they do, they are often buggy). But tax authorities and accounting software vendors do have to formally model the language in the law and then potentially change those formal definitions after court decisions.
My point is that the way in which the modeled, formal world interacts with probabilistic, fuzzy language and human actions is complex. In my opinion we will always need both. AIs ultimately need to understand both and be able to combine them just like (competent) humans do. AI "tool use" is a stop-gap. It's not a sufficient level of understanding.
No, I think hes suggesting that "world models" are more impactful. The issue for him inside meta is that there is already a research group looking at that, and are wildly more successful (in terms of getting research to product) and way fucking cheaper to run than FAIR.
Also LeCun is stuck weirdly in product land, rather than research (RL-R) which means he's not got the protection of Abrash to isolate him from the industrial stupidity that is the product council.
If LLMs actually hit a plateau, then investment will flow towards other architectures.
> is massively monopolistic and have unbounded discretionary research budget
that is the case for most megacorps. if you look at all the financial instruments.
modern monopolies are not equal to single corporation domination. modern monopolies are portfolios who do business using the same methods and strategies.
the problem is that private interests strive mostly for control, not money or progress. if they have to spend a lot of money to stay in control of (their (share of the)) segments, they will do that, which is why stuff like the current graph of investments of, by and for AI companies and the industries works.
A modern equivalent and "breadth" of a Bell Labs (et. al) kind of R&D speed could not be controlled and would 100% result in actual Artificial Intelligence vs all those white labelababbebel (sry) AI toys we get now.
Post WW I and II "business psychology" have build a culture that cannot thrive in a free world (free as in undisturbed and left to all devices available) for a variety of reasons, but mostly because of elements with a medieval/dark-age kind of aggressive tendency to come to power and maintain it that way.
In other words: not having a Bell Labs kind of setup anymore ensures that the variety of approaches taken on large scales aka industry-wide or systemic, remains narrow enough.
If Deepseek is free it undermines the value of LLMs, so the value of these US companies is mainly speculation/FOMO over AGI.
And I stopped reading him, since he - in my opinion - trashed on autopilot everything 99% did - and these 99% were already beyond the two standard deviation of greatness.
It is even more highly problematic if you have absolutely no results eg products to back your claims.
How did you determine that "surefire paths to success still available"? Most academics agree that LLMs (or LLMs alone) are not going to lead us to AGI. How are you so certain?
I'll happily step out of the way once someone simply tells me what it is you're trying to accomplish. Until you can actually define it, you can't do "it".
yes, a glib response, but think about it: we define an intelligence test for humans, which by definition is an artificial construct. If we then get a computer to do well on the test we haven't proved it's on par with human intelligence, just that both meet some of the markers that the test makers are using as rough proxies for human intelligence. Maybe this helps signal or judge if AI is a useful tool for specific problems, but it doesn't mean AGI
Same goes for academia. People's visions compete for other people's financial budgets, time and other resources. Some dogs get to eat, study, train at the frontier and with top tools in top environments while the others hope to find a good enough shelter.
Not that I believe AGI is the measure of success, there's probably much more efficient ways to achieve company goals than simulating humans.
That didn’t last. People in the know knew that once you have a billion users and insane revenue and market power and have basically bought or driven out of business most of your competitors (Diapers.com, Jet.com, etc) you can eventually slow down your physical expansion, tighten the screws on your suppliers, increase efficiencies, and start printing money.
The VCs who are funding these companies are hoping that they have found the next Amazon. Many will probably go out of business, but some might join the ranks of trillion dollar companies.
As for IQ tests and the like, to the extent they are "scientific" they are designed based on empirical observations of humans. It is not designed to measure the intelligence of a statistical system containing a compressed version of the internet.
Who says they don't make money? Same with open source software that offer a hosted version.
> If Deepseek is free it undermines the value of LLMs, so the value of these US companies is mainly speculation/FOMO over AGI
Freemium, open source and other models all exist. Does it undermine the value of e.g. Salesforce?
> Our probabilistic, fuzzy concepts are often simply a misconception.
How eg a credit card works today is defined by financial institutions. How it might work tomorrow is defined by politics, incentives, and human action. It's not clear how to model those with formal language.
I think most systems we interact with are fuzzy because they are in a continual state of change due to the aforementioned human society factors.
Yann was largely wrong about AI. Yann coined the term stochastic parrot and derrided LLMs as a dead end. It’s now utterly clear the amount of utility LLMs have and that whatever these LLMs are doing it is much more than stochastic parroting.
I wouldn’t give money to Yann, the guy is a stubborn idiot and closed minded. Whatever he’s doing wont even touch LLM technology. He was so publicly deriding LLMs I see no way he will back pedal from that.
I dont think LLMs are the end of the story for agi. But I think they are a stepping stone. Whatever agi is in the end, LLMs or something close to it will be a modular component of aspect of the final product. For LeCunn to dismiss even the possibility of this is idiotic. Horrible investment move to give money to Yann to likely pursue Agi without even considering LLMs.
This is something that was true last year, but hanging on by a thread this year. Genie shows this off really well, but it's also in the video models as well.[1]
[1]https://storage.googleapis.com/gdm-deepmind-com-prod-public/...
Like the new spin out Episteme from OpenAI?
They generate revenue, but most companies are in the hole for the research capital outlay.
If open source models from China become popular, then the only thing that matters is distribution / moat.
Can these companies build distribution advantage and moats?
But ultimately I agree with you that this entire societal process is just categorically different. It's simply not a description or definition of something, and therefore the question of how formal it can be doesn't really make sense.
Formalisms are tools for a specific but limited purpose. I think we need those tools. Trying to replace them with something fuzzy makes no sense to me either.
> how many people are continually ready to mistake language for thought
This is a fundamental illusion - where, rote memory and names and words get mistaken for understanding. This was wonderfully illustrated here [1]. Few really grok what understanding actually is. This is an unfortunate by-product of our education system.
> Are they all P-zombies or just obedience-conditioned into emulating ones?!?!?
Brilliant way to state the fundamental human condition. ie, we are all zombies conditioned to imitate rather than understand. Social media amplifies the zombification, and now LLMs do that too.
> Starting with the sophomoric questions of the optimist who mistakes the possible for the viable
This is the fundamental tension between operationalized meaning and imagination. A grokking soul gathers mists from the cosmic chaos and creates meaning and operationalizes it for its own benefit and then continually adapts it.
> it's amazing if sad that they cracked it by more data and not by more model
I was speaking to experts in the sciences (chemistry). They were shocked that the underlying architecture is brute force. They expected a compact information-compressed theory which is able to model independent of data. The problem with brute-force approaches are that they dont scale, and dont capture the essences which are embodied in theories.
> The physical limits of human language as information processing device have been hit at some point in the XX century
2000 years back when humans realized that formalism was needed to operationalize meaning, and natural language was too vague to capture and communicate it. Because the world model that natural language captures encompasses "everything" whereas for making it "useful" requires to limit it via formalism.
And of course it doesn't work. Humans don't have world models. There's no such thing as a world model!
And let's not speak about those so deep into sloth that put it into use to deteriorate, and not augment as they claim to do, humane creative recreational activities.
The issue is context. trying to make an AI assistant with just text only inputs is doeable but limiting. You need to know the _context_ of all the data, and without visual input most of it is useful.
For example "Where is the other half of this" is almost impossible to solve unless you have an idea of what "this" is.
but to do that you need to have cameras, to use cameras you need to have position, object, and people tracking. And that is a hard problem thats not solved.
the hypothesis is that "world models" solve that with an implicit understanding of the worl and the objects in context
Now, it's not like he opened up Anthropic's books for an audit, so you don't necessarily have to trust him. But you do need to believe that either (a) what he is saying is roughly true or (b) he is making the sort of fraudulent statements that could get you sent to prison.
LeCun had chosen to focus on the latter. He can't be blamed for not having taken the second hat.
Why they decided not to do that is kind of a puzzle.
That kind of hallucination is somewhat acceptable for something marketed as a chatbot, less so for an assistant helping you with scientific knowledge and research.
Musk cares about AI research as much as he cared about Path of Exile
Talking to these people is exhausting, so I cut straight to the chase: name the exact, unavoidable conditions that would prove AGI won’t happen.
Shockingly, nobody has an answer. They’ve never even considered it.
That’s because their whole belief is unfalsifiable.
How many decades did it take for neural nets to take off?
The reason we're even talking about LeCun today is because he was early in seeing the promise of neural nets and stuck with it through the whole AI winter when most people thought it was a waste of time.
LLMs cannot do any of the major claims made for them, so competing at the current frontier is a massive resource waste.
Right now a locally running 8b model with large context window (10k tokens+) beat google/openAI models easily on any task you like.
why would anyone then pay for something that is possible to run on consumer hardware with higher token/second throughput and better performance? What exactly have the billions invested given google/oai in return? Nothing more than an existensial crisis I'd say.
Companies aren't trying to force AI costs into their subscription models in dishonest ways because they've got a winning product.
And animals' main concern is energy conservation, so they must be doing something else.
But the principle is there. I think that when a company sits on a load of cash, that's what they should do. Either that or become a kind of alternative investments allocator. These are risky bets. But they should be incentivized to take those risks. From a fiscal policy standpoint for instance. Well it probably is the case already via lower taxation of capital gains and so on. But there should probably exist a more streamlined framework to make sure incentives are aligned.
And/or assigned government projects? Besides implementing their Cloud infrastructure that is...
The problem isn't LLMS, the problem is that everyone is trying to build bigger/better llms or manually code agents around LLMs. Meanwhile, projects like Mu Zero are forgotten, despite being vastly more important for things like self driving.
Hyper growth is expensive because it’s usually capital intensive. The trick is, once that growth phase is over, can you then start milking your customers while keeping a lid on costs? Not everyone can, but Amazon did, and most investors think OpenAI and Anthropic can as well.
The animal learns as it encounters learning signals - prediction failure - which is the only way to do it. Of course you need to learn/remember something before you can use than in the future, so in that sense it's "ahead of time", but ultimately it's done that way because evolution has found that learning patterns will ultimately prove beneficial.