Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

A global workspace in language models

anthropic.com

Post reply on HN
Print view View original

191–200 of 218 posts

Re: A global workspace in language models

#191

Post by rustyhancock »

horacemorace wrote: post #105
ACCount37 wrote: post #56

Earlier quoted context omitted.

"The reversal curse", it rarely shows up in practice but you found a case when it did. The "knowledge landscape" an LLM uses is "directional". It's easy to reach "a quirky music band from Michigan known for colored ties" when you stand at "Tally Hall". But if you stand at "a quirky music band from Michigan known for colored ties", it's harder to reach "Tally Hall" from there. For the "latent knowledge graph" an LLM u…

If this kind of thing holds true for humans we now may understand synesthesia, perhaps we find ourselves with a large enough study pool to map out clever tricks to influence each other.

What's this got to do with synesthesia?

Re: A global workspace in language models

#192

Post by thoughtpeddler »

Anyone else have tips for how to build skepticism around this type of paper? I find myself for whatever reason more readily inclined to believe the Anthropic mech interp team's claims, but then after reading skeptical takes, I 'snap out of it' and more clearly see the still-unsettled science of it all, but I wish I had better priors. Although I follow this space fairly closely (versus the "average person"), I still feel under-equipped when facing research that might be equal parts marketing and science.

Re: A global workspace in language models

#193

Post by garlic_enjoyer »

thoughtpeddler wrote: post #192

Anyone else have tips for how to build skepticism around this type of paper? I find myself for whatever reason more readily inclined to believe the Anthropic mech interp team's claims, but then after reading skeptical takes, I 'snap out of it' and more clearly see the still-unsettled science of it all, but I wish I had better priors. Although I follow this space fairly closely (versus the "average person"), I still f…

For papers that make grand claims, they usually have poignant limitations. In this paper, it's clear that the conclusion is extrapolated from a small set of observable information. That's usually the recipe for poor conclusions, and they are up front about where their findings sit.

That's not to say these findings aren't valuable though. Their blog post summary is just a bit more hype inducing than the underlying paper.

The way I feel about this is that the emergent properties of LLMs seem to reflect our own human faculties. Whether that's because we built them in our image or whether this is a common mode of consciousness is definitely outside the scope of this paper.

Re: A global workspace in language models

#194

Post by rootatixww3 »

mike_hearn wrote: post #180
gwerbin wrote: post #178

Earlier quoted context omitted.

They must, right? They literally have no mechanism for cognition other than transformation of vocabulary. Language models are models fitted from data generated by humans, but they are not humans. Humans generate data by whatever processes happen in our brains, and LLMs of various architectures can learn a surprisingly good approximation of that data-generating process. That doesn't mean they have all the same charact…

My interpretation of prior mechanistic interpretability research is that LLMs transform concepts in non-verbal spaces as well, which is why that was a surprising statement. For example showing how they do arithmetic on line lengths by rotating helical manifolds. And the way they transform data isn't by transforming words. The layers transform high dimensional vectors - a format very alien to us. It's not obvious that…

an LLM can't access its high dimensional vectors any more than we can access whatever the brain is doing at a low level

all kind of math structures were found in mammals brains - fourier transforms (well, not exactly), ballistic equations, Gabor filters

who knows how exactly we approximate the magnitude of a math operations, maybe we also use helices

my point is that we dont know if what we discover the neural networks doing (helical manifolds) is actually the same thing a brain converges on, or not

and there is an implicit bias here - evolution created language, and we forced neural networks to also evolve to be good at it. so it wouldn't be surprising to find some convergence, this particular kind of language turned out to work well (words, linear sentences, grammar)

Re: A global workspace in language models

#195

Post by mike_hearn »

rootatixww3 wrote: post #194
mike_hearn wrote: post #180

Earlier quoted context omitted.

My interpretation of prior mechanistic interpretability research is that LLMs transform concepts in non-verbal spaces as well, which is why that was a surprising statement. For example showing how they do arithmetic on line lengths by rotating helical manifolds. And the way they transform data isn't by transforming words. The layers transform high dimensional vectors - a format very alien to us. It's not obvious that…

an LLM can't access its high dimensional vectors any more than we can access whatever the brain is doing at a low level all kind of math structures were found in mammals brains - fourier transforms (well, not exactly), ballistic equations, Gabor filters who knows how exactly we approximate the magnitude of a math operations, maybe we also use helices my point is that we dont know if what we discover the neural networ…

Interestingly the paper (I finished scan reading it now) does say that the models can move data from their 'automatic' circuits into the J-space if they need to reason on it reflectively. So in some sense LLMs actually can access their own vectors, at least some of the time.

Re: A global workspace in language models

#196

Post by JacobAsmuth »

unleaded wrote: post #55

This reminded me of some weird quirk/experiment I found with LLMs that I found while messing around, maybe someone can explain it or something. Open any AI chatbot that isn't cheating by connecting to the Internet (so disable web search). Claude, DeepSeek, Kimi, whatever. Ask them this question: "What was that weird band from michigan from the 2000s that wore coloured ties" You will probably get a wrong answer, or if…

Hate to break it to you but Gemini 3.1 Flash-lite with grounding disabled (through the API) gets this first try, even with thinking disabled.

Re: A global workspace in language models

#197

Post by JacobAsmuth »

devanshp wrote: post #58
unleaded wrote: post #55

This reminded me of some weird quirk/experiment I found with LLMs that I found while messing around, maybe someone can explain it or something. Open any AI chatbot that isn't cheating by connecting to the Internet (so disable web search). Claude, DeepSeek, Kimi, whatever. Ask them this question: "What was that weird band from michigan from the 2000s that wore coloured ties" You will probably get a wrong answer, or if…

https://claude.ai/share/2b0f85a2-e7b8-4f62-91a0-eca61bdeabec Fable 5 on low gets the answer with web search turned off, one-shot!

Gemini 3.1 Flash-Lite with web search disabled through the API and zero thinking gets it just fine.

Re: A global workspace in language models

#198

Post by JacobAsmuth »

paraschopra wrote: post #126

Well, isn't it sort of expected? It's a common misconception that LLMs residual exists for predicting just the next token. While training, we sum/average the losses across whole sequence which puts the pressure to predict future tokens on residual stream of _all_ past tokens. For example, if a particular shape of residual helps reduce loss across several future tokens, it will take that shape (even if it takes a slig…

Would you expect to find concepts related to emotions evoked by thinking those thoughts present? Or meta-descriptions of thoughts? Sounds like you're just post-hoc rationalizing to me.

Re: A global workspace in language models

#199

Post by matankal »

fc417fc802 wrote: post #87
hasteg wrote: post #77

Earlier quoted context omitted.

Really interesting paper, thanks for the share. The point their making in that paper reminds me of this paper some people shared around work earlier this year, https://arxiv.org/pdf/2512.14982 (Prompt Repetition Improves Non-Reasoning LLMs)... I wonder how OPs question would fare (or the questions presented in the paper you posted) given double repetition.

A relevant question for the paper you linked is what happens if instead of repeating the input prompt you repeat a filler character the same number of times.

This is tested in the paper (see the "padding" ablation)

Re: A global workspace in language models

#200

Post by segmondy »

unleaded wrote: post #55

This reminded me of some weird quirk/experiment I found with LLMs that I found while messing around, maybe someone can explain it or something. Open any AI chatbot that isn't cheating by connecting to the Internet (so disable web search). Claude, DeepSeek, Kimi, whatever. Ask them this question: "What was that weird band from michigan from the 2000s that wore coloured ties" You will probably get a wrong answer, or if…

[dead]
Post reply on HN