"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. "
The paper's critique of the 'data wall' and language-centrism is spot on. We’ve been treating AI training like an assembly line where the machine is passive, and then we wonder why it fails in non-stationary environments. It’s the ultimate 'padded room' architecture: the model is isolated from reality and relies on human-curated data to even function.
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.
In this podcast episode[0] he does talk about this kind of model and how it "learns about physics" through experience instead of just ingesting theorical material.
> Curious how you'd handle the reward signal for deciding when to switch between observation and active exploration without it collapsing into one mode.
If you like biomimetic approaches to computer science, there's evidence that we want something besides neural networks. Whether we call such secondary systems emotions, hormones, or whatnot doesn't really matter much if the dynamics are useful. It seems at least possible that studying alignment-related topics is going to get us closer than any perspective that's purely focused on learning. Coincidentally quanta is on some related topics today: https://www.quantamagazine.org/once-thought-to-support-neuro...
Not learning from new input may be a feature. Back in 2016 Microsoft launched one that did, and after one day of talking on Twitter it sounded like 4chan.[1] If all input is believed equally, there's a problem.
Today's locked-down pre-trained models at least have some consistency.
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?
This is an astonishing claim and if true, will make AI a lot less useful in real life scenario.
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 remember a joke from few years ago that was showing an "AI" that was "learning" on its "own" which meant periodically starting from scratch with a new training set curated by a large team of researchers themselves relying on huge teams (far away) of annotators.
TL;DR: depends where you defined the boundaries of your "system".
But doesnt existing AI systems already learn in some way ? Like the training steps are actually the AI learning already. If you have your training material being setup by something like claude code, then it kind of is already autonomous learning.
Eh, honestly? We're not that far away from models training themselves (opus 4.6 and codex 5.3 were both 'instrumental' in training themselves).
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.
The whole AI field is a misnomer. It stole so much from neurobiology.
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.
It’s interesting, LeCun seems to have a blind spot around in-context learning. I didn’t find one mention in this paper (only skimmed the full paper so far so may have missed), which is odd as it is the way that agents come closest to autonomous learning in the real world.
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.
29 comments
[ 3.5 ms ] story [ 48.5 ms ] thread"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. "
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.
https://ai.meta.com/blog/yann-lecun-ai-model-i-jepa/
It's quite eye opening.
0. https://youtu.be/qvNCVYkHKfg
If you like biomimetic approaches to computer science, there's evidence that we want something besides neural networks. Whether we call such secondary systems emotions, hormones, or whatnot doesn't really matter much if the dynamics are useful. It seems at least possible that studying alignment-related topics is going to get us closer than any perspective that's purely focused on learning. Coincidentally quanta is on some related topics today: https://www.quantamagazine.org/once-thought-to-support-neuro...
Today's locked-down pre-trained models at least have some consistency.
[1] https://www.bbc.com/news/technology-35890188
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.
Imagine if AI learns all your source code and apply them to your competitor /facepalm
TL;DR: depends where you defined the boundaries of your "system".
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.
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.
That's why I think the term "system" as used in the paper is much better.
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.