Anyone remember that blog post from a few months back where someone was able to improve a model's math ability by just duplicating layers that were activated while solving math problems? Just literally copy/pasting them and linking them together so the model ran through the same layers again?
I get the feeling a lot more research is going to come out in the area of exploring exactly what portions of a model's weights do what.
If dirt-simple type operations like copy-paste yield useful improvements with even a small probability that would seem to open things up for adaptive reconfiguration and whole other classes of optimizations like genetic algorithms.
I always wondered what the model meant when it writes "I'm now considering the architecture of the service" but outputs nothing of the sorts in its CoT.
Is the model really "thinking" about that stuff or is just mimicking human "manners"? And if so, where the thinking is happening if it is not in the literal chain of *thought*?
I'm not sure J-Space is the answer to that question, but very interesting nevertheless.
> I'm now considering the architecture of the service
What you see here is a summary of thinking tokens written by some other smaller model (e.g. old sonnet). The actual thinking sometimes (rarely) leaks and is not easy to parse.
Almost none of the hosted models give you their unredacted CoT. Claude certainly doesn't, what you get are fragments and summaries from it.
There are various justifications on this, but it's mostly to make distillation and fine tuning off their model outputs a bit harder for their competitors
Neel Nanda (of Google Deepmind - his part begins on page 33) discusses his opinions on the paper, and the small-scale replication he performed on an open-weight model.
This is a really interesting development in language models and will be a small but relevant blip on the timeline in the development of artificial intelligence over our lifetimes
This, taken in combination with the SAE paper, the golden-gate claude paper, the feelings / introspection paper, and note in the fable system card (that they are silently nerfing responses about activation shaping), is basically confirmation to me that they have a new technique they they are using during training (along the vibe space of these mechinterp papers), and its probably some kind of representation learning akin to the core ideas of JEPA.
There was a series of blog posts posted to HN a while ago investigating how models behave on similar prompts in different languages. To paraphrase the results: the first couple layers map the query to some internal encoding that's mostly independent of the language. Then there are layers in the middle, then the last couple layers map the result back to the target language. You can actually take those middle layers and repeat them, and you get a stronger model. Those middle layers would be what Anthropic calls the J-Space, and their J-Lens maps activity in those layers back to tokens that trigger similar activity (with a technique they only drop hints at)
The finding that you can repeat the middle layers pairs neatly with Anthropic's finding that there is some internal CoT-like process happening in them. I'm not sure how to find those blog posts, but maybe someone else remembers them
It's not in the weights. Sounds to me like jspace is the "positive cone" over relevant (large norm) j-lenses, and j-lenses are gradients wrt tokens on the residual stream when you average over some training data.
It's been shown that LLMs use their outer layers to decode from and encode to language, while their middle layers deal in language-independent abstract concepts. This means that the same question or statement in different languages activates the outer layers differently but produces the same patterns in the middle layers. Check this article with cool visualizations (btw, this is one of the articles mentioned also by a sibling answer):
The middle layers also perform reasoning on the abstract concepts, to the point that you can replicate some blocks of inner layers (thus giving the LLM more internal "reasoning space") and by this increase the model's reasoning abilities. The video in this article shows that when performing a sequence of arithmetic operations (without CoT, i.e. the result is spit out directly), internally the intermediate calculations are spelled out, and this can only happen in the depth direction of the LLM (since no new token is added to the sequence). So this "jspace" can only be situated in the middle layers, probably in circuits that repeat nearly identical across several layers.
It would be really cool if they could expose this information to customers somehow. Imagine:
- having a log of the most prominent J-space tokens during your customer support chatbot's interactions with a user, so you can have more introspection into why a particular outcome happened
- being able to detect certain thoughts associated with undesirable behavior (hallucinations, overstepping authority, lying, etc.) and trigger some sort of remediation (e.g. upgrading to a better model, redirecting to a human, forcing tool calls)
Presumably the rationale for the decision to abridge the thinking traces will ensure that they don’t; if this is real (and there’s no good reason to trust that it is yet) then it is the secret sauce.
I get the feeling a lot more research is going to come out in the area of exploring exactly what portions of a model's weights do what.
Is the model really "thinking" about that stuff or is just mimicking human "manners"? And if so, where the thinking is happening if it is not in the literal chain of *thought*?
I'm not sure J-Space is the answer to that question, but very interesting nevertheless.
What you see here is a summary of thinking tokens written by some other smaller model (e.g. old sonnet). The actual thinking sometimes (rarely) leaks and is not easy to parse.
There are various justifications on this, but it's mostly to make distillation and fine tuning off their model outputs a bit harder for their competitors
More interesting was the independent commentary paper they linked near the bottom: https://www-cdn.anthropic.com/files/4zrzovbb/website/cc4be24...
Neel Nanda (of Google Deepmind - his part begins on page 33) discusses his opinions on the paper, and the small-scale replication he performed on an open-weight model.
(Nb: not an expert / in the labs, just opining)
The finding that you can repeat the middle layers pairs neatly with Anthropic's finding that there is some internal CoT-like process happening in them. I'm not sure how to find those blog posts, but maybe someone else remembers them
https://dnhkng.github.io/posts/sapir-whorf/
The middle layers also perform reasoning on the abstract concepts, to the point that you can replicate some blocks of inner layers (thus giving the LLM more internal "reasoning space") and by this increase the model's reasoning abilities. The video in this article shows that when performing a sequence of arithmetic operations (without CoT, i.e. the result is spit out directly), internally the intermediate calculations are spelled out, and this can only happen in the depth direction of the LLM (since no new token is added to the sequence). So this "jspace" can only be situated in the middle layers, probably in circuits that repeat nearly identical across several layers.
Their method is used to identify which tokens can appears in which layers of the model.
I too have confusion.
https://distrowatch.com/weekly.php?issue=20260706#freebsd
We should really stop giving these liar models any further credibility.