Gary Marcus has been taking victory laps on this since mid-2023, nothing to see here. Patently obvious to all that there will be additional innovations on top of LLMs such as test-time compute, which nonetheless are structured around LLMs and complementary
Looking at the quoted tweet it is immediately obvious that these people have no clue about the current state of research. Yes they might have had some more or less relevant contributions to classical ML, but AI has taken off without (or rather despite) them and if history of AI has shown anything, it's that people like those are not the ones who will pave the way forward. In a field like this, there's no use to listen to people who still cling to their old ideas just because the current ideas don't seem "elegant" or "right" in their mind. The only thing you can trust is data and it proves we haven't peaked yet when it comes to LLMs.
more seriously though, as best as i can understand, what he is trying to say is that there must be a *LOGICAL* framework independent of compute or what you get is just a parrot (stochastic one at best) that operates within the smoothed edges of a distributed statistical field.
I just checked - he's right. Anthropic won't write code anymore. ChatGPT is just jumbled, dyslexic letters and nonsense. I generated a Midjourney image 10 times, each one was just TV static.
Gary Marcus has never built anything, has never contributed meaningful to any research that actually produces value, nor has he been right about any of his criticisms.
What he has done is continually move a goalpost to stay somewhat relevant in the blogsphere and presumably the academic world.
i don't think general intelligence is technically unachievable with ML but i think we're still orders of magnitude away from the amount of compute needed to reach it and everyone is in a honeymoon period because of how useful text prediction and the current state of it has proven to our day to day jobs
"We want AI agents that can discover like we can, not which contain what we have discovered. Building in our discoveries only makes it harder to see how the discovering process can be done." - I am curious if people would read this as an advocacy or criticism of LLMs?
Discovery comes from search for both humans and AI agents. There is no magic in the brain or LLM except learning along the way and persistence. The search space itself is externalized.
So the AI agents are "good enough" but environment access is insufficient for collecting the required experience, this is the current bottleneck.
For example even a simple model like AlphaZero (just a CNN) was good enough to beat the best humans and rediscover game play from scratch, but it had the extensive access to the environment.
Readers should make sure to contextualize this. We're talking about people researching AGI. Current LLM models are amazing, and will have business and societal impact. Previous ML models also had business and societal impact. None of that is contested here. The question is, what path leads to AGI, do LLM scale to AGI? That is the question being asked here, and some researchers think that it won't, it will scale superbly to many things, but something else might be needed for full AGI.
His stance on LLMs can be modeled by a simple finite state machine:
State 1) LLM performance stalls for a couple of months:
- "See I told you, LLMS are a dead end and won't work!"
State 2) New LLM release makes rapid and impressive improvements
- "AI is moving too fast! This is dangerous and we need the government to limit the labs to slow them down!"
People can laugh at Gary Marcus all they want, but there’s one aspect that people don’t understand.
If you have a high conviction belief that is countervailing to the mainstream, you suffer a great deal. Even the most average conversation with a “mainstream believer” can turn into a judgment fest. Sometimes people stop talking to you mid-conversation. Investors quietly remove you from their lead lists. Candidates watch your talks and go dark on you. People with no technical expertise lecture at you.
Yet, inevitably, a fraction of such people carry forward. They don’t shut up. And they are the spoon that stirs the pot of science.
It’s totally normal and inevitable that people start to take victory laps at even the smallest indication in such situations. It doesn’t mean they’re right. It is just not something worth criticizing.
Marcus claims to have reread The Bitter Lesson. And I should say, I too have reread the text and I don't think Marcus is getting the actual original argument of here. All it say is that general purpose algorithms that scale will outperform special purpose algorithms that use information about the problem and don't scale. That's all. Everyone claiming more is hallucinating, things into this basic point. Notably general purpose algorithms aren't necessary neural nets and "X works better than Y" doesn't imply X is the best thing every.
So there's no contradiction between The Bitter Lesson and claims that LLMs have big hole and/or won't scale up to AGI.
I've often thought that if you want to represent a probabilistic world model, with nodes that represent physical objects in space-time (and planned-future space-time) and our level of certainty about their relationships to one another... you'd do that outside an LLM's token stream.
You could, in theory, represent that model as a linear stream of tokens, and provide it as context to an LLM directly. It would be an absurdly wasteful number of tokens, at minimum, and the attention-esque algorithm for how someone might "skim" that model given a structured query would be very different from how we skim over text, or image patches, or other things we represent in the token stream of typical multi-modal LLMs.
But could it instead be something that we provide as a tool to LLMs, and use an LLM as the reasoning system to generate structured commands that interact with it? I would wager that anyone who's read a book, drawn a map of the fantasy world within, and argued about that map's validity on the internet, would consider this a viable path.
At the end of the day, I think that the notion of a "pure LLM" is somewhat pedantic, because the very term LLM encapsulates our capability of "gluing" unstructured text to other arbitrary tools and models. Did we ever expect to tie our hands behind our back and make it so those arbitrary tools and models aren't allowed to maintain state? And if they can maintain state, then they can maintain the world model, and let the LLM apply the "bitter lesson" that compute always wins, on how to best interact with and update that state.
What are the odds they will just be stumbling around for another few decades before the next big discontinuous jump in effectiveness is uncovered? The AI Gods always had big ideas and opinions, but the discovery of LLMs seem to have been pure serendipidy.
25 comments
[ 2.8 ms ] story [ 51.5 ms ] threadIt doesn’t matter what incredible things neural networks do, in his mind they’re always a dead end that will collapse any day now.
Thesis: language isn't a great representation, basically.
I really should apply myself. Maybe I wouldn't work so hard, just shuck nonsense/pontificate.
more seriously though, as best as i can understand, what he is trying to say is that there must be a *LOGICAL* framework independent of compute or what you get is just a parrot (stochastic one at best) that operates within the smoothed edges of a distributed statistical field.
> Yann LeCun was first, fully coming around to his own, very similar critique of LLMs by end of 2022.
> The Nobel Laureate and Google DeepMind CEO Sir Demis Hssabis sees it now, too.
Sutton... the patron saint of scaling...
Listen to people for the their ideas, not their label.
Regardless, Marcus is a bit late to comment on the bitter lesson. That is so 6 months ago lol
It's... it's over. The west has fallen.
What he has done is continually move a goalpost to stay somewhat relevant in the blogsphere and presumably the academic world.
"We want AI agents that can discover like we can, not which contain what we have discovered. Building in our discoveries only makes it harder to see how the discovering process can be done." - I am curious if people would read this as an advocacy or criticism of LLMs?
So the AI agents are "good enough" but environment access is insufficient for collecting the required experience, this is the current bottleneck.
For example even a simple model like AlphaZero (just a CNN) was good enough to beat the best humans and rediscover game play from scratch, but it had the extensive access to the environment.
His stance on LLMs can be modeled by a simple finite state machine:
State 1) LLM performance stalls for a couple of months: - "See I told you, LLMS are a dead end and won't work!"
State 2) New LLM release makes rapid and impressive improvements - "AI is moving too fast! This is dangerous and we need the government to limit the labs to slow them down!"
Repeat
If you have a high conviction belief that is countervailing to the mainstream, you suffer a great deal. Even the most average conversation with a “mainstream believer” can turn into a judgment fest. Sometimes people stop talking to you mid-conversation. Investors quietly remove you from their lead lists. Candidates watch your talks and go dark on you. People with no technical expertise lecture at you.
Yet, inevitably, a fraction of such people carry forward. They don’t shut up. And they are the spoon that stirs the pot of science.
It’s totally normal and inevitable that people start to take victory laps at even the smallest indication in such situations. It doesn’t mean they’re right. It is just not something worth criticizing.
Marcus claims to have reread The Bitter Lesson. And I should say, I too have reread the text and I don't think Marcus is getting the actual original argument of here. All it say is that general purpose algorithms that scale will outperform special purpose algorithms that use information about the problem and don't scale. That's all. Everyone claiming more is hallucinating, things into this basic point. Notably general purpose algorithms aren't necessary neural nets and "X works better than Y" doesn't imply X is the best thing every.
So there's no contradiction between The Bitter Lesson and claims that LLMs have big hole and/or won't scale up to AGI.
You could, in theory, represent that model as a linear stream of tokens, and provide it as context to an LLM directly. It would be an absurdly wasteful number of tokens, at minimum, and the attention-esque algorithm for how someone might "skim" that model given a structured query would be very different from how we skim over text, or image patches, or other things we represent in the token stream of typical multi-modal LLMs.
But could it instead be something that we provide as a tool to LLMs, and use an LLM as the reasoning system to generate structured commands that interact with it? I would wager that anyone who's read a book, drawn a map of the fantasy world within, and argued about that map's validity on the internet, would consider this a viable path.
At the end of the day, I think that the notion of a "pure LLM" is somewhat pedantic, because the very term LLM encapsulates our capability of "gluing" unstructured text to other arbitrary tools and models. Did we ever expect to tie our hands behind our back and make it so those arbitrary tools and models aren't allowed to maintain state? And if they can maintain state, then they can maintain the world model, and let the LLM apply the "bitter lesson" that compute always wins, on how to best interact with and update that state.