>In contrast to a classic RNN, there's no unbounded hidden state accumulating across an entire trajectory
I don' understand this line. In a classic RNN hidden state is bounded dimension. In fact it's transformers that technically have unbounded hidden state.
You can't parallelize classic nonlinear RNNs for various reasons but in training both RNN and Transformer depend on the entire sequence history in a way that is unbounded. Of course in practice you just train on a max sequence length.
I am confused how chaining two 32-layer models is comparable to a 64 layer model in terms of "difficulty in chain of thought". The reasoning appears to rely on the fact that each processing of a token has fixed number of steps while my understanding is that can very greatly based on the type of data being reasoned whether it is originally text or something else. My mind falls back to graph theory in this case and pictures a much higher potential branching in a 64 layer model and all the tradeoffs that come with that.
I must not have the right idea of what is happening here.
I like the idea of more reccurance in the transformer level. Chain of thought always seemed so clunky. Its just not the way the human brain processes information. Its an extrmeely crude approximation at best
One thing about the reasoning is that models are trained to generate a chain of thoughts, but it doesn’t have to be correct, accurate, or reflect the underlying logic of the LLM. It’s the same problem we have with the output, it is something plausible, but not that reliable
Yes, both the output should be "milestones" of sorts, like lemmas and theorems in math. Important plateaus that serve as a launching pad to the next phase. Regurgitating every thought potentially degrades signal:noise ratio.
>This suggests that deeper isn't always better for looped transformers, which leaves me less worried about a race to the bottom toward looped transformers with hundreds of recurrent loops.
I disagree with this. Deeper will always be at least as good because the extra loops can exit early or just no-op. Any performance degradation they're seeing at higher loop counts today is merely training stability issues, which can be overcome.
Deeper almost certainly is better, and we will probably see not just hundreds but millions of recurrent loops in the future.
Deeper independent, sure. Deeper shared though? Information and signal theory still apply here. At infinite cycles, without new input, you'll end up with a locked state or oscillations. Some point before that, any "attractors" in the latent space, with slightly higher statistics, will pull things towards a space that might eventually be only loosely related to the goal, because each loop would be lossy, right?
> A lot of hype around OpenAI's Astra model here on my timeline today. Apparently, this goes back to a new article from The Information, which said Astra is a "recurrent depth or looped transformer".
> It's always interesting to read about new or different approaches (including rumors about what the closed labs may be up to), but let's debunk this a bit.
> About 2 months ago, I shared the architecture details of Nanbeige, for example, where "Nanbeige4.2-3B is pretrained from scratch on 28T tokens with a Looped Transformer that reuses the layer stack to increase capacity without adding parameters."
> Yes, that's it. The looped transformer idea is just reusing layers in the transformer block.
> In the case of Nanbeige, the main idea is to reuse the same 22-layer stack (=transformer block) twice instead of once. So, effectively it extends the 22-layer architecture to 44 layers, but without duplicating the weights.
> In simple terms, this roughly doubles the size of the model (if we ignore the embedding and output layers for a second). But instead of requiring 2x the storage and RAM to host this model, it stays at the same size since we reuse the components. However, it's almost 2x as expensive in terms of compute, because we run the embedded text through almost 2x as many layers.
> Why? In the Nanbeige 4.2 technical report, the researchers found that two passes gave the best trade-off and retained about 75% of the token efficiency of a standard architecture. (More passes gave barely any gains but made the training much slower and much more expensive.)
> While, as far as I know, Nanbeige 4.2 is the first notable open-weight model that adopted this approach, the idea goes back to the NeurIPS paper "Mixture-of-recursions: Learning dynamic recursive depths for adaptive token-level computation". Actually, this paper proposes a mechanism that is a bit more sophisticated by adding a learned router that determines whether each token receives one, two, or more passes. So, easy tokens can exit early while harder tokens receive additional computation.
> In sum, Astra may be a really good model, but this shouldn't be about this "looped transformer aspect," which is just a tiny architectural tweak.
Anyone remember Universal Transformers paper (Dehghani et al) from back in 2018? Recurrent transformers have a history as long as transformers themselves.
Somewhat unclear how particularly novel this is vs a way to save compute.
Looped transformers replace n-different self attention layers into one layer that gets executed m-times usually until a stopping condition is met. My personal intuition is that it just leaves another degree of freedom in the way QKV weights can be packed so that it's slightly more efficient.
You have to take a step back and examine the context in which the post is written. The LW/EA community is just a little obsessed with AI safety - it's easy to construct hypothetical events where A(G/S)I exterminates humanity that function as a technological version of Pascal's Wager.
One of the AI safety interests is AI explainability - the thought here that reading an AI's 'thoughts' will help us design safer models as well as detect models that go 'rogue' or are malevolently plotting against humans.
That's where the fear of looped transformers comes from. Is the residual stream that looped transformers iterate on a potential hiding place for plotting AI?
In my opinion, no more so than the residual stream of existing transformers. It changes zero.
You could imagine large number of loops, thousands. But you are constrained by the width of the residual stream since you loop over one token. But then you can imagine the model learning to sub-divide it to pack even more info into it.
I can imagine a lot of things. However there is a packing limit for QKV weights that sets the ceiling on how much this occurs, and it's quite low - think 1.3-1.8x. The limiting factor here isn't the number of iterations, it's model size, same as it ever was.
To add a bit more constructive feedback, think of the "AI cornucopia" and "Superintelligence destroys humanity" as opposite ends of all possible outcomes distribution (low probability event).
The most likely one is the "business as usual, but with AI" - some things will get better, some things will get worse, but overall state of affairs will remain mostly the same.
The people who've thought the most about this put it differently:
Think of a new, superintelligent model as if it was a new v1 Starship launching for the first time, with a full fuel tank. On the one hand, rockets have existed for some time, and some have gone to space successfully, including by this company.
On the other hand, this is a tube of metal full of highly explosive liquid going faster than most human objects ever go, for the first time ever in this novel and state of the art configuration.
If someone said, "really, the first Starship exploding is just at one end of the probability distribution, where the other is that everything goes fine and all its passengers have a nice trip in space," would you get on that rocket?
Or, more aptly, if you and every other living human was already on that rocket, would you push the launch button?
The analogy works because superintelligence is, like rocket fuel, an extremely powerful force that has a default tendency to break containment and go boom (consume lots of energy and heat and matter in a chain reaction, to pursue more intelligence to pursue whatever goal it is pursuing.)
Current AIs are closer to bottle rockets than to the Starship on the intelligence scale. Some property damage already happened, but you can't master the art of rocketry without trial and error.
But will “business as usual, but with AI” justify the current capital expenditures? I think the market is pricing things as being closer to “AI cornucopia”.
it's just an experimental optimization. Implementation detail...
Irrelevant to "safety". I mean its going to have to go in that direction anyway... eventually the models will just be constantly thinking, refining their internal thoughts / weights... External input and output will be rare, just as it is for most humans.
>In contrast to a classic RNN, there's no unbounded hidden state accumulating across an entire trajectory
I don' understand this line. In a classic RNN hidden state is bounded dimension. In fact it's transformers that technically have unbounded hidden state.
You can't parallelize classic nonlinear RNNs for various reasons but in training both RNN and Transformer depend on the entire sequence history in a way that is unbounded. Of course in practice you just train on a max sequence length.
RNN xhat[t+1]=f(x[t],h[t])
Transformer/self-attention xhat[t+1]=f(x[t],h[t],h[t-1],...,h[1])
On further thought, I think the author's intent was to say that classic RNNs have "unbounded temporal accumulation in the hidden state".
I am confused how chaining two 32-layer models is comparable to a 64 layer model in terms of "difficulty in chain of thought". The reasoning appears to rely on the fact that each processing of a token has fixed number of steps while my understanding is that can very greatly based on the type of data being reasoned whether it is originally text or something else. My mind falls back to graph theory in this case and pictures a much higher potential branching in a 64 layer model and all the tradeoffs that come with that.
I must not have the right idea of what is happening here.
I like the idea of more reccurance in the transformer level. Chain of thought always seemed so clunky. Its just not the way the human brain processes information. Its an extrmeely crude approximation at best
Chain of thought seems very specifically like an instantiation of the System 2 thinking of the System 1 and 2 thinking Daniel Kahneman popularized.
Is it clunky in that it's a verbalized/languified version of system 2 thinking, and clearly humans do some non-verbal version too?
Some humans apparently only or almost only do non-verbal thinking. It’s called anendophasia.
it is what I do to solve hard problems through.
easy stuff happens by itself, but with a system large enough you need a scratchpad and a rubber duck.
One thing about the reasoning is that models are trained to generate a chain of thoughts, but it doesn’t have to be correct, accurate, or reflect the underlying logic of the LLM. It’s the same problem we have with the output, it is something plausible, but not that reliable
Yes, both the output should be "milestones" of sorts, like lemmas and theorems in math. Important plateaus that serve as a launching pad to the next phase. Regurgitating every thought potentially degrades signal:noise ratio.
Related ongoing thread:
OpenAI's new reasoning technique alarms AI safety experts - https://news.ycombinator.com/item?id=49552395
>This suggests that deeper isn't always better for looped transformers, which leaves me less worried about a race to the bottom toward looped transformers with hundreds of recurrent loops.
I disagree with this. Deeper will always be at least as good because the extra loops can exit early or just no-op. Any performance degradation they're seeing at higher loop counts today is merely training stability issues, which can be overcome.
Deeper almost certainly is better, and we will probably see not just hundreds but millions of recurrent loops in the future.
Deeper independent, sure. Deeper shared though? Information and signal theory still apply here. At infinite cycles, without new input, you'll end up with a locked state or oscillations. Some point before that, any "attractors" in the latent space, with slightly higher statistics, will pull things towards a space that might eventually be only loosely related to the goal, because each loop would be lossy, right?
Sebastian Raschka posted about this architecture:
> A lot of hype around OpenAI's Astra model here on my timeline today. Apparently, this goes back to a new article from The Information, which said Astra is a "recurrent depth or looped transformer".
> It's always interesting to read about new or different approaches (including rumors about what the closed labs may be up to), but let's debunk this a bit.
> About 2 months ago, I shared the architecture details of Nanbeige, for example, where "Nanbeige4.2-3B is pretrained from scratch on 28T tokens with a Looped Transformer that reuses the layer stack to increase capacity without adding parameters."
> Yes, that's it. The looped transformer idea is just reusing layers in the transformer block.
> In the case of Nanbeige, the main idea is to reuse the same 22-layer stack (=transformer block) twice instead of once. So, effectively it extends the 22-layer architecture to 44 layers, but without duplicating the weights.
> In simple terms, this roughly doubles the size of the model (if we ignore the embedding and output layers for a second). But instead of requiring 2x the storage and RAM to host this model, it stays at the same size since we reuse the components. However, it's almost 2x as expensive in terms of compute, because we run the embedded text through almost 2x as many layers.
> Why? In the Nanbeige 4.2 technical report, the researchers found that two passes gave the best trade-off and retained about 75% of the token efficiency of a standard architecture. (More passes gave barely any gains but made the training much slower and much more expensive.)
> While, as far as I know, Nanbeige 4.2 is the first notable open-weight model that adopted this approach, the idea goes back to the NeurIPS paper "Mixture-of-recursions: Learning dynamic recursive depths for adaptive token-level computation". Actually, this paper proposes a mechanism that is a bit more sophisticated by adding a learned router that determines whether each token receives one, two, or more passes. So, easy tokens can exit early while harder tokens receive additional computation.
> In sum, Astra may be a really good model, but this shouldn't be about this "looped transformer aspect," which is just a tiny architectural tweak.
https://x.com/rasbt/status/2095141254958858496
Anyone remember Universal Transformers paper (Dehghani et al) from back in 2018? Recurrent transformers have a history as long as transformers themselves.
Somewhat unclear how particularly novel this is vs a way to save compute.
I'm literally zero concerned.
Looped transformers replace n-different self attention layers into one layer that gets executed m-times usually until a stopping condition is met. My personal intuition is that it just leaves another degree of freedom in the way QKV weights can be packed so that it's slightly more efficient.
You have to take a step back and examine the context in which the post is written. The LW/EA community is just a little obsessed with AI safety - it's easy to construct hypothetical events where A(G/S)I exterminates humanity that function as a technological version of Pascal's Wager.
One of the AI safety interests is AI explainability - the thought here that reading an AI's 'thoughts' will help us design safer models as well as detect models that go 'rogue' or are malevolently plotting against humans.
That's where the fear of looped transformers comes from. Is the residual stream that looped transformers iterate on a potential hiding place for plotting AI?
In my opinion, no more so than the residual stream of existing transformers. It changes zero.
You could imagine large number of loops, thousands. But you are constrained by the width of the residual stream since you loop over one token. But then you can imagine the model learning to sub-divide it to pack even more info into it.
I can imagine a lot of things. However there is a packing limit for QKV weights that sets the ceiling on how much this occurs, and it's quite low - think 1.3-1.8x. The limiting factor here isn't the number of iterations, it's model size, same as it ever was.
A debate between grossly incompetent against grossly immoral, honestly.
You should ignore anything LessWrong or OpenAI says and do your own research.
To add a bit more constructive feedback, think of the "AI cornucopia" and "Superintelligence destroys humanity" as opposite ends of all possible outcomes distribution (low probability event).
The most likely one is the "business as usual, but with AI" - some things will get better, some things will get worse, but overall state of affairs will remain mostly the same.
The people who've thought the most about this put it differently:
Think of a new, superintelligent model as if it was a new v1 Starship launching for the first time, with a full fuel tank. On the one hand, rockets have existed for some time, and some have gone to space successfully, including by this company.
On the other hand, this is a tube of metal full of highly explosive liquid going faster than most human objects ever go, for the first time ever in this novel and state of the art configuration.
If someone said, "really, the first Starship exploding is just at one end of the probability distribution, where the other is that everything goes fine and all its passengers have a nice trip in space," would you get on that rocket?
Or, more aptly, if you and every other living human was already on that rocket, would you push the launch button?
The analogy works because superintelligence is, like rocket fuel, an extremely powerful force that has a default tendency to break containment and go boom (consume lots of energy and heat and matter in a chain reaction, to pursue more intelligence to pursue whatever goal it is pursuing.)
Current AIs are closer to bottle rockets than to the Starship on the intelligence scale. Some property damage already happened, but you can't master the art of rocketry without trial and error.
But will “business as usual, but with AI” justify the current capital expenditures? I think the market is pricing things as being closer to “AI cornucopia”.
What happens if/when we don’t deliver?
> What happens if/when we don’t deliver?
I wish we had real journalism, the AI labs CEO should be asked that question in every single interview
Market bubble will pop, not the first time and not the last, most likely after OpenAI and Anthropic IPOs.
My guess is after Anthropic and before OpenAI
it's just an experimental optimization. Implementation detail... Irrelevant to "safety". I mean its going to have to go in that direction anyway... eventually the models will just be constantly thinking, refining their internal thoughts / weights... External input and output will be rare, just as it is for most humans.