LeCun also said back in 2022 that "if you train a machine, as powerful as it could be, your 'GPT-5000', on text", it will never be able to learn basic common-sense physics like that objects placed on tables will move along with them.
I often see people vindicate those who predicted really fast takeoff to AGI / ASI, because the capabilities have obviously been taking off extremely quickly. But still not as quickly as many predicted! To me, the people who confidently predicted that we'd all be out of a job by 2024 or 2025 have been just as wrong as LeCun has been.
It's only online that I've seen these wild predictions, where it coincidentally is profitable to say them. These people are looking g for clicks or to stay relevant, so they have to say crazy stuff.
Being wrong, even badly wrong, is fine, so long as one adjusts their beliefs accordingly. LeCun has not.
And has it at this stage, within in-depth take of said "learning", foundationally?
I have not been able to properly check the studies for a long time now, but I remain unaware of achieved solutions on the problem of reliably referencing a world model out of a language model - that "counting the 'r's in 'raspberry'" be not guessing, not memory, but actually counting.
Incredibly inefficiently because of the recursive loops ("Wait, the object is on the table. I should think about this more deeply..."), and likely instantly surpassed by large world models if/when those are shipped, but effectively enough vs non-thinking models.
Picking the right tool or model is like picking the right problem to work on. It's actually quite hard (often you can't just try them all), but without it you will be incredibly inefficient and occasionally, fundamentally wrong.
All models are wrong, but some are useful. -Box
Nothing intrinsically more or less direct about the LLM's method than ours.
I've never tried it and it might take some thought and effort to conduct an experiment to find out properly, but I would be interested in the answer.
In my mind general intelligence is pretty much by definition a virtual machine, so the mechanisms behind thought are only relevant for the sake of efficiency (ie you can argue that LLMs make a poor basis for intelligence because tokens and natural language are a poor way to encode the world, but if you can run it on a big enough computer to counteract the inherent wasteful virtualisation then who really cares how it works under the hood?)
I will stop here before our analogies go too far.
"Count the 'r's in the string 'raspberry'" is the same as "Count the joints in the Tasmanian sand spider": it means, "can you instance in idea and assess it and perform procedures over it - for example, can you count?".
We want to be sure it can count. We have seen it may guess, it may construe, it may recall - we instead demand it checks. We cannot trust it without that.
If you were home and a family member asked you that question, you'd probably criticise the question rather than answering. LLM are RLHF'd into being milk-toast helpers that just try to answer questions like that with no criticism.
This is all beside the fact that the world of AI has changed pretty dramatically in the last few months.
The only context that these LLMs have is whatever is in their context window. Not only that, the context is updated on each token emitted with the new token. It's a chaotic process, and a flaw inherent to the technology. It's not going to be fixed anytime soon.
It’s nonsense to test if a product that is marketed and sold as being able to provide generalised intelligence on demand, does what it says on the tin?
Check yourself
Similarly for Apple’s “red herring” paper, simply adding a generic caveat to “disregard irrelevant factors” (without specifying which ones) restored performance even in the weaker local llama models back then.
The flaw was not in the reasoning; the flaw seems to be simply that the assumptions we make are often different from the assumptions it makes. I wonder if that might be a fundamental underlying cause of misalignment.
This is always the issues in the discussions.
There’s the outcomes camp (objectivists?), which points at the things LLMs can do.
Then there’s the process methods camp, which talks about what is actually going on.
If you only care about the outcome, then the process does t matter.
If you are talking about what is happening, what the underlying mechanics and science of it is, then the process matters.
These models aren’t thinking. They simulate cognition well enough to do useful work in several fields and domains.
Both are true.
They are for any definition of the word that makes any kind of sense. I'm sure you have a contorted definition that magically only includes humans though...
The models are simulating thinking, if the fidelity is good enough for you - great!
The normal definition of the word "thinking" definitely includes what LLMs do. Hell people used to say computers were thinking even before AI. It's super weird to get all uppity about the semantics of the word now.
If a model isn’t able to count to 100, but is able to use python to do math, it isn’t thinking.
If you are only an outcomes person, then this difference doesn’t matter to you.
But when people are talking about AGI, anthropomorphizing LLMs, and making assumptions on its behavior based on that, then people are being loose or misled.
Language changes over time anyway, so even if you really believe what LLMs are doing isn't the "thinking" of 2024, it probably will be the "thinking" of 2027, because most people are using it that way.
And yes, we do need precision. Language is a vehicle to convey ideas, at its best it creates clarity. At its worst it creates deception and confusion.
What the labs are doing is designed to mislead. Being clear about the terms and avoiding needless anthropomorphization, is a basic decency we owe to the many others who are being riled up or terrified by the conclusions and implications that come from using “thinking” when it’s something else.
The difference between being satisfied by the outcomes and being interested in the process is sufficient to indicate the exactness required in your terminology, and does not harm any position.
These aren’t thinking machines, but if what they simulate is good enough, and your work ends with the output, it doesn’t quite matter.
People can be happy with a meal from McDonalds, or they can want something prepared by a chef, or Soylent.
All are acceptable and bring utility to their customers.
For "thinking" here we mean "assessing a representation of an object". That, or equivalent, is required to be reliable. So it is fundamental and critical.
But the outcomes group "ignores" the fundamental limitations of models which are purely text based.
E.g, a baseball players trains to catch high-speed balls and they dont do it by: "ball velocity 50mph, vector:[1,2,3], run move hand command now"
That's absurd.
No, there is an embodied network which is "trained" on visual, tactile input, and control as direct output.
LLMs are fundamentally not the right tool for that.
That is a NN that learns a skill.
But that is not an Analyst. If it were ballistics, then the answer to "how to parametrize the launch to reliably hit the target" excludes getting the result through natural skill.
The problem lies in the need to get "AI" facing "LLMs": the latter create a need for reliability, for "AI".
Speech is an endowment of both those who give educated guesses via developed skills and of those who return answers like Analysts, who check and compute. LLMs create a confusion between the two, and they will remain a problem until an ability to act as Analysts - strictly - will be implemented.
Dynamical systems will never be solved in a semantic domain.
IMO, they can be helpful the in robotics stack, from planning level up, but that's it.
What exactly is not clear? I will rephrase.
The internal process of the blackbox oracle determine the reliability of the output.
Two abilities are very different: learning trajectories through empirical training ("increasingly catching thousands of thrown balls"), and determining trajectories through computation (a rational thinker at work). The former is a finetuned parametrized engine (a «NN that learns a skill»), the latter is an Analyst. The former is fuzzy, the second deterministic.
In front of fuzzy LLMs, which use potentially misleading outputs - text ("has it guessed or has it thought?") - the urgency of warranties of reliable output gets evident.
So, that they «simulate cognition [only] well enough» (Intended wrote), and that there are «fundamental limitations of models ... purely text based» (Kooi wrote) raises the urgency to overcome the "fuzzy" and achieve the "deterministic" - it is not that we can stall on a «fundamentally not the right tool for that».
Inventing an oracle calls for urgent striving to overcome the original weakness.
Incidents like hugging face are partly rooted in the lack of common sense. It still functions like a supercharged toddler.
I'd love to overcome this because it'd mean I spend less time guiding the the LLM to produce usable outputs.
And we've had difficulty as humans to childproof our sandboxes and infrastructure. Things that are otherwise innocuous spots to coordinate between like minded toddlers can become problematic.
Uh, no? So much of what we learn and take for granted as common sense is not learned via language, and not even expressible in it.
consider me optimist now, but just few months ago, even frontier models were dumb, doing stupid mistakes all the time, all of them were so dumb I'd never expect anything to change in just few months.
We can have adequate representations of light that are the instances over which we reason. Your simile is about perception, not about instancing ideas.
Given you also don't want it to memorise [for all tokens, count([for all letters]), this would probably be more like "here's two images, count all things in the big image that look like the thing in the small image", which can then be r's in a photo of a raspberry jam jar in a supermarket, or dragons in a photo of a furry convention, or whatever.
That said, they are competent enough at coding that I keep seeing them write code to do even simple tasks.
On a related note: why did I see Claude editing a file by using cat to write a python script to do a grep search and replace?
Why not? You've memorized how words are spelled, and how sounds correspond with letters, and how concepts correspond with words. To the extent that there are shortcuts that enable compression you use these, and the model will do something similar.
Being able to spell all the words then count letters is simpler, and more generalisable to other tasks, than memorising answers to all possible word questions.
That said, we're so bad at splitting facts from skills that trying to get them to memorise a bunch of facts might force them to learn a skill and generalise anyway.
Because to "123x456" we want a reply that goes "this times that plus that...", not "Was that not nnnnnn?". If it does not perform its duty (returning solid checked answers) it is a liability.
Of any object in question they should be able to create a representation that allows correct assessment.
> Given you also don't want it to memorise
That is obviously necessary: what we want from the consultant is to check, not to remember. Answers must be correct and that implies having performed all due diligence - and being capable of doing it, before that. So, objects must be instanced internally in a way that allows effective handling. Counting letters is a good example of the ability (that must remain general).
But I'm sure you can still find tasks that they will have difficulty solving, involving the most fundamental concepts that can only be experienced in the physical world to be understood well, like left and right, near and far, hot and cold, heavy and light, etc.
The good designer understands culture, tastes and preferences as they evolve in real time. That’s why llm as design tools haven’t displaced the good designers.
We don't see in UV or infrared, but some birds do.
On the other hand, AI can understand reality in ways we individually can't: It has combined access to unlimited knowledge of math, physics and chemistry that at this point no single human possesses. Maybe it understands reality better than any single human at this point.
Also we're seeing AI reason from first principles in order to simulate physics and get to that real world understanding without directly interacting with it.
I use LLMs daily to help me code etc. but... It wasn't long ago that frontier models were confidently recommending to walk, without the car, to the car wash to wash the car no?
As a daily user of LLMs I do certainly see my fair share of WTF "solutions" to coding problems. I'm not saying it's not super useful: it is super useful. But I don't exactly feel like I'm talking to something that understands that the car needs to be present to be washed.
An analogy on LLMs is that you have a pretty clear straight highway ahead of you for some distance right now. Maybe that doesn't lead to AGI but it's clear there's progress to be made. For a big tech company it makes sense to push as hard and fast down that clear straight highway of LLMs asap.
Meanwhile LeCunn wanted to turn off the road and go down an unproven track. I say this as someone working on world model generation right now (creating the ability to learn game world model and have it play the game https://tfmbot.com for an example of my system pointed at a very complex board game). LeCunn wanted to pivot all of Meta into world model generation. It's good as a side track research project but the entire pivot he wanted to do was madness.
People are literally talking about an AI researcher who was fired for terrible direction here.
Meta’s AI projects are still negative ROIC
The argument is that LLMs are a local maximum that will never breakthrough to AGI. This is still very much an open question. If you are the fifth-best AI lab, does it make sense to try to outcompete everyone in a space that is already too crowded and may not ever yield their actual objective? Instead they could just use open weight models in their products, or post-train on open models like smaller labs have done, and treat that as what it is: product development.
Pure research has always been about taking chances.
This was facetious of course, but humans generally don't learn this through analysis the way you'd have to train an LLM to answer questions about expectations about the world. In this sense he is accurate.
Gemini 3.1 Pro hasn't needed that pretty much at all, which is impressive compared to how much I've learned other models need it. Somehow it's able to mostly handle that stuff itself without needing the constant manual reminders and hand-holding. It still misses the occasional one or two things but it's way better than other models missing entire classes of things constantly. Somehow, it feels appropriate though I have no actual evidence why.
Mary packed the binoculars in chapter 3, therefore she may use them on the train in chapter 6.
Also, LeCun mentioned [3] "a chat with Kunihiko Fukushima in 1991", which states that "Fukushima started to work on a backprop version of the Neocognitron in 1989 or so but saw our 1989 paper in Neural Computation and gave up."
[1] LeCun et al., "Backpropagation applied to handwritten zip code recognition", 1989
[2] Fukushima et al., "Neocognitron: A self-organizing neural network model for a mechanism of pattern recognition unaffected by shift in position", 1980
Low hanging fruit successfully plucked, I guess.
Low hanging fruit is somewhat the opposite of sour grapes - I don’t want these grapes because they were probably sour versus so what if he got those sweet grapes - they were hanging low!
Maybe connecting “low hanging fruit” to “sour grapes” is “low hanging fruit” to some but it took a serious mental leap for me.