Back

A global workspace in language models

74 points2 hoursanthropic.com
botanrice6 minutes ago

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

wavemode27 minutes ago

As someone who is not an AI researcher, the paper itself is way over my head.

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.

eamag42 minutes ago

Is it scaling up of https://openreview.net/forum?id=w7LU2s14kE with some changes on where this method is applied?

bilsbie33 minutes ago

I’m confused where in the weights the jspace is.

lucrbvi22 minutes ago

Anthropic theorize that middle layers in an LLM is a "J-Space" used to "think" about the future answer or about abstract concepts.

Their method is used to identify which tokens can appears in which layers of the model.

epolanski27 minutes ago

Tokens that are activated but not present in it's output maybe?

I too have confusion.

esafak43 minutes ago

Without using the term, they are using an information geometric approach.

blauditore25 minutes ago

But J-Space is much catchier. This is not a scientific paper, it's a promotional essay.

viralsink10 minutes ago

First button on the page is a link to the scientific paper. It's called "Read the paper". You'll find an explanation for the term in there.

meatmanek43 minutes ago

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)
charcircuit13 minutes ago

Anthropic aren't even willing to expose the CoT of their models. You will have to rely on them to build those sorts of things into dedicated signals.

dofm38 minutes ago

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.

bilsbie31 minutes ago

Maybe model performance could increase dramatically if we found a way to scale this up.