Why AI systems don't learn – On autonomous learning from cognitive science
arxiv.org
arxiv.org
Today's locked-down pre-trained models at least have some consistency.
In a decade since then, things got marginally better, and such events wouldn't play out so fast and so intensely in 2026.
Are you saying the internet would not do it again, or Microsoft would not do the same approach? Because I think the internet would absolutely do it again.
Like when Google wasn't personalized so rank 3 for me is rank 3 for you. I like that predictability.
Obviously ignoring temperature but that is kinda ok with me.
Sameness is bad for an LLM like it’s bad for a culture or species. Susceptible to the same tricks / memetic viruses / physical viruses, slow degradation (model collapse) and no improvement. I think we should experiment with different models, then take output from the best to train new ones, then repeat, like natural selection.
And sameness is mediocre. LLMs are boring, and in most tasks only almost as good as humans. Giving them the ability to learn may enable them to be “creative” and perform more tasks beyond humans.
The first LLMs were utter crap because of that, but once you have just one that's good enough it can be used for dataset filtering and everything gets exponentially better once the data is self consistent enough for there to be non-contradictory patterns to learn that don't ruin the gradient.
Well it shows that most humans degrades into 4chan eventually. AI just learned from that. :)
If aliens ever arrive here, send an AI to greet them. They will think we are totally deranged.
Was it based on a specific scientific paper or research?
The controversy surrounding it seemed to have polluted any search for a technical breakdown or a discussion, or the insights gained from it.
There seems to have been interest in a model which would pick up language and style of its conversations (not actually learning information or looking up facts). If you haven't trained an LSTM model before - you could train on Shakespeare's plays and get out ye olde English in a screenplay format, but from line to line there was no consistency in plot, characters, entrances and exits, etc. in a way which you'd expect after GPT-2. Twitter would be good for keeping a short-form conversation. So I believe Tay and the Watson that appeared on Jeopardy are more from this 'classical NLP' thinking and not proto-LLMs, if that makes sense.
Imagine deploying a software product that changes over time in unknown ways -- could be good changes, could be bad, who knows? This goes beyond even making changes to a live system, it's letting the system react to the stream of data coming in and make changes to itself.
It's much preferable to lock down a model that is working well, release that, and then continue efforts to develop something better behind the scenes. It lets you treat it more like a software product with defined versions, release dates, etc., rather than some evolving organism.
Ugh HN is so tedious with these remarks. These people are trying to get computers to learn, not just train on data, and HN goes nOt LeArNiNg Is A fEaTuRe. Where's the wonder and the curiosity?
In real life, take programming as an example, we want Claude to be strong in capability at first, but what is more important is for it to learn our code base, be proficient in it, as it gains experience around it. In other words, become a domain expert.
Because our code base is proprietary I don't expect ( not do I want) the AI to be familiar with it on the first day. So learning on the job is the only way to go.
Only in that way it will resemble a human programmer, and only then we can truly talk about replacing human programmer.
Learning is OpenClaw's distinguishing feature. It has an array of plugins that let it talk to various services - but lots of LLM applications have that.
What makes it unique is it's memory architecture. It saves everything it sees and does. Unlike an LLM context its memory never overflows. It can search for relevant bits on request. It's recall is nowhere near as well as the attention heads of an LLM, but apparently good enough to make a difference. Save + Recall == memory.
I would say his core point does still apply; autonomous learning is not solved by ICL. But it seems a strawman to ignore the topic entirely and focus on training.
From what I see on the ground, some degree of autonomous learning is possible; Agents can already be set up to use meta-learning skills for skill authoring, introspection, rumination, etc - but these loops are not very effective currently.
I wonder if this is the myopic viewpoint of a scientist who doesn’t engage with the engineering of how these systems are actually used in the real world (ie “my work is done once Llama is released with X score on Y eval”) which results in a markedly different stance than the guys like Sutskever, Karpathy, Amodei who have built end-to-end systems and optimized for customer/business outcomes.
"he proposed framework integrates learning from observation (System A) and learning from active behavior (System B) while flexibly switching between these learning modes as a function of internally generated meta-control signals (System M). We discuss how this could be built by taking inspiration on how organisms adapt to real-world, dynamic environments across evolutionary and developmental timescales. "
Anyone doing AI coding can tell you once an agent gets on the wrong path, it can get very confused and is usually irrecoverable. What does that look like in other contexts? Is restarting the process from scratch even possible in other types of work, or is that unique to only some kinds of work?
True, for iterations between the same two players, but humans evolved the ability to communicate and so can share the results of past interactions through a network with other agents, aka a reputation. Thus any interaction with a new person doesn't start from a neutral prior.
Algorithms do not possess ethics nor morality[0] and therefore cannot engage in Machiavellianism[1]. At best, algorithms can simulate same as pioneered by ELIZA[2], from which the ELIZA effect[3] could be argued as being one of the best known forms of anthropomorphism.
0 - https://www.psychologytoday.com/us/basics/ethics-and-moralit...
1 - https://en.wikipedia.org/wiki/Machiavellianism_(psychology)
>As Weizenbaum later wrote, "I had not realized ... that extremely short exposures to a relatively simple computer program could induce powerful delusional thinking in quite normal people."...
That pretty much explain the AI Hysteria that we observe today.
I think the "AI Hysteria" comes more from current LLMs being actually good at replacing a lot of activity that coders are used to doing regularly. I wonder what Weizenbaum would think of Claude or ChatGPT.
Yea, that is kind of the point. Even such a system could trick people into delusional thinking.
> actually good at replacing a lot of activity that coders are used to...
I think even that is unrealistic. But that is not what I was thinking. I was thinking when people say that current LLMs will go on improving and reach some kind of real human like intelligence. And ELIZA effect provides a prefect explanation for this.
It is very curious that this effect is the perfect thing for scamming investors who are typically bought into such claims, but under ELIZA effect with this, they will do 10x or 100x investment....
Neither can commercial LLM-based offerings.
>It's part of the history of the field of artificial intelligence that every time somebody figured out how to make a computer do something—play good checkers, solve simple but relatively informal problems—there was a chorus of critics to say, 'that's not thinking'.
That pretty much explains the "it's not real AI" hysteria that we observe today.
And what is "AI effect", really? It's a coping mechanism. A way for silly humans to keep pretending like they are unique and special - the only thing in the whole world that can be truly intelligent. Rejecting an ever-growing pile of evidence pointing otherwise.
And they were always right...and the other guys..always wrong..
See, the questions is not if something is the "real ai". The questions is, what can this thing realistically achieve.
The "AI is here" crowd is always wrong because they assign a much, or should I say a "delusionaly" optimistic answer to that question. I think this happens because they don't care to understand how it works, and just go by its behavior (which is often cherry-pickly optimized and hyped to the limit to rake in maximum investments).
Modern production grade LLMs are entangled messes of neural connectivity, produced by inhuman optimization pressures more than intelligent design. Understanding the general shape of the transformer architecture does NOT automatically allow one to understand a modern 1T LLM built on the top of it.
We can't predict the capabilities of an AI just by looking at the architecture and the weights - scaling laws only go so far. That's why we use evals. "Just go by behavior" is the industry standard of AI evaluation, and for a good damn reason. Mechanistic interpretability is in the gutters, and every little glimpse of insight we get from it we have to fight for uphill. We don't understand AI. We can only observe it.
"What can this thing realistically achieve?" Beat an average human on a good 90% of all tasks that were once thought to "require intelligence". Including tasks like NLP/NLU, tasks that were once nigh impossible for a machine because "they require context and understanding". Surely it was the other 10% that actually required "real intelligence", surely.
The gaps that remain are: online learning, spatial reasoning and manipulation, long horizon tasks and agentic behavior.
The fact that everything listed has mitigations (i.e. long context + in-context learning + agentic context management = dollar store online learning) or training improvements (multimodal training improves spatial reasoning, RLVR improves agentic behavior), and the performance on every metric rises release to release? That sure doesn't favor "those are fundamental limitations".
Doesn't guarantee that those be solved in LLMs, no, but goes to show that it's a possibility that cannot be dismissed. So far, the evidence looks more like "the limitations of LLMs are not fundamental" than "the current mainstream AI paradigm is fundamentally flawed and will run into a hard capability wall".
>We don't understand AI. We can only observe it.
Lol what? Height of delusion!
> Beat an average human on a good 90% of all tasks that were once thought to "require intelligence".
This is done by mapping those tasks to some representation that an non-intelligent automation can process. That is essentially what part of unsupervised learning does.
Don't get me wrong, he has some banger prior work, and the recent SIGReg did go into my toolbox of dirty ML tricks. But JEPA line is rather disappointing overall, and his distaste of LLMs seems to be a product of his personal aesthetic preference on research direction rather than any fundamental limitations of transformers. There's a reason why he got booted out of Meta - and it's his failure to demonstrate results.
That talk of "true understanding" (define true) that he's so fond of seems to be a flimsy cover for "I don't like the LLM direction and that's all everyone wants to do those days". He kind of has to say "LLMs are fundamentally broken", because if they aren't, if better training is all it takes to fix them, then, why the fuck would anyone invest money into his pet non-LLM research projects?
It is an uncharitable read, I admit. But I have very little charity left for anyone who says "LLMs are useless" in year 2026. Come on. Look outside. Get a reality check.
>"LLMs are useless" in year 2026
Literally no one is saying this. It is just that those words are put into the mouths of the people that does not share the delusional wishful thinking of the "true believers" of LLM AI.
>Literally no one is saying this.
Did you not just advise me to go watch a podcast full of "LLMs are literally incapable of inventing new things" and "LLMs are literally incapable of solving new problems"?
I did skim the transcript. There are some very bold claims made there - especially when LLMs out there roll novel math and come up with novel optimizations.
No, not reliably. But the bar we hold human intelligence to isn't that high either.
Sure, but the same could apply to you as well.
>"LLMs are literally incapable of inventing new things" and "LLMs are literally incapable of solving new problems"?
You keep proving that you have trouble resolving closely related ideas. Those two things that you mention does not imply that they are "useless". They are a better search and for software development, they are useful for reviews (at least for a while). But it seems that people like you can only think in binary. It is either LLMs are god like AI, or they are useless.
Conjecture. There are plenty of ethical frameworks grounded in pure logic (Kant), or game theory (morality as evolved co-operation). These are both amenable to algorithmic implementations.
Algorithm implementations are programmatic manifestations of mathematical models and, as such, are not what they model by definition.
To wit, NOAA hurricane modelling[0] are obviously not the hurricanes which they model.
0 - https://www.aoml.noaa.gov/hurricane-modeling-prediction/
This is false for constructs of information, ie. a "manifested model" of a sorted list is a sorted list and a "manifested model" of a sorting algorithm is a sorting algorithm.
To wit, an accurate algorithmic model of moral reasoning is moral reasoning, since moral reasoning, being a decision procedure, is an information process.
Rofl. Someone hasn't discovered Functionalism or the identity of indiscernables. Must be hard laboring under such a poverty of reasoning.
> Rofl. Someone hasn't discovered Functionalism or the identity of indiscernables. Must be hard laboring under such a poverty of reasoning.
First, the correct spelling is "indiscernibles" not "indiscernables".
Second, thank you for motivating me to research the origin of the quote:
Better to remain silent and be thought a fool than to speak
and to remove all doubt.[0]
> This is false for constructs of information, ie. a "manifested model" of a sorted list is a sorted list and a "manifested model" of a sorting algorithm is a sorting algorithm.
Specious. Constructs existing strictly within a mathematical model are not concepts external to what they model by definition. Constructs of information qualify as same as they are subject to assessment traditionally performed by stakeholders (humans). Continuing with the example provided, sorting algorithms exist as well-defined mathematical concepts and can be applied to other model concepts, yet do not transcend their manifestation in program logic.
> To wit, an accurate algorithmic model of moral reasoning is moral reasoning ...
A model of moral reasoning can only produce results to which an entity external to the model determines correctness. The model, in and of itself, has no capability to assert same beyond what it has been modeled to process.
I guess it would depend a bit whos interests the AI would be serving. If serving the shareholders it would probably reward creating value for customers, but if it was serving an individual manager competing with others to be CEO say then the optimum strategy might be to go machiavellian on the rivals.
Is this not just because their goals are currently to be seen as "nice"?
Surely they can be not-nice if directed to, and then the question is just whether someone can accidentally direct them to do that by e.g. setting up goals that can be more readily achieved by being not-nice. Which... is how many goals in the real world are, which is why the very concept and danger of Machiavellianism exists.
It's quite eye opening.
He raised $1b but that seems way too little to buy enough compute to train.
My bet is that OpenAI or Anthropic or both will eventually train the model that he always wanted because they will use revenue from LLMs to train a world model.
However had, there will come a time when AI will really learn. My prediction is that it will come with a different hardware; you already see huge strides here with regards to synthetic biology. While this focuses more on biology still, you'll eventually see a bridging effort; cyborg novels paved the way. Once you have real hardware that can learn, you'll also have real intelligence in AI too.
TL;DR: depends where you defined the boundaries of your "system".
In that sense the "autonomous" part you said simply meant that the data source is coming from a different place, but the model itself is not free to explore with a knowledge base to deduce from, but rather infer on what is provided to it.
This is the "Claude Code" part, or even the ChatGPT (web interface/app) part. Large context window full of relevant context. Auto-summarization of memories and inclusion in context. Tool calling. Web searching.
If not LLMs, I think we can say that those systems that use them in an "agentic" way perhaps have cognition?
Start a new chat, and the "agentic" system will be as clueless as before
My point is that it's not the "model" that is the thing that is demonstrating cognition here, it's the "system" that uses the model and stores information and can retrieve it later. To that system, these notes are more the equivalent of my memories than my notebooks.
And it is not "learning" (which was the initial claim) as the system never learns beyond what was already there in the training data, and any new information you supply are new data, from scratch, that is immediately forgotten once the session is over.
It's easy to prove: start a new session in your project and ask "what is this project about". Two days from now, in a new session, ask the same thing. Observe how in both cases it re-reads the files, greps source code etc. Meanwhile a system with actual memories and learning wouldn't have to do that.
From what I've been learning in my uni, this is said pre-programmed. Cognition is really the ability from, out of no context, no knowledge of what you are capable of, to learn something. These tool calling and web searching are, in the end, MCP functions provided by the LLM provider themselves.
It's an entire academic discussion, about how things start. For example babies: they someone have a knowledge base on how to breath, how to cry, but they have absolute no knowledge on how to speak and it learns by the interactions with the parents.
LLMs try as much as they can create this by inference and pre-programmed functions, but they don't have a graph of memories with utility to weight their relevance in the context. As others said, the context window dies as soon as you close the session.
They also don't have the epistemic approach that is to know that another agent knows about something just by observing the environment they were all put in.
The proposed System M (Meta-control) is a nice theoretical fix, but the implementation is where the wheels usually come off. Integrating observation (A) and action (B) sounds great until the agent starts hallucinating its own feedback loops. Unless we can move away from this 'outsourced learning' where humans have to fix every domain mismatch, we're just building increasingly expensive parrots. I’m skeptical if 'bilevel optimization' is enough to bridge that gap or if we’re just adding another layer of complexity to a fundamentally limited transformer architecture.
Imagine if AI learns all your source code and apply them to your competitor /facepalm
(I guess one could call projects like https://en.wikipedia.org/wiki/Project_Cybersyn an "application" of its ideas, though cut off before one could see the results.)
They're capable enough to put themselves in a loop and create improvement which often includes processing new learnings from bruteforcing. It's not in real-time, but that probably a good thing if anyone remembers microsofts twitter attempt.
That's why I think the term "system" as used in the paper is much better.
No. No, they don't