About this event

Charles reviews the Anthropic J-Space Paper

This Tuesday we will be studying Anthropic’s paper, “Verbalizable Representations Form a Global Workspace in Language Models” a.k.a the J-space paper.

Now, this paper is 117 pages long, so I don’t recommend reading the entire thing (nor have I done so myself). But please do read at least the following:
* Section 2, “methods” — it’s only about four pages, and it helps to frame everything else.
* The blog post — it’s shorter and more digestible than the paper and can help give an idea what the authors (or at least someone at Anthropic) think are the most salient points. The video is optional. Note: I suggest you read Section 2 first, otherwise the blog post will make J-spaces seem far more mysterious than they are.

After that, please choose at least one additional section to read according to your own interests. You may find it helpful to upload the paper to an LLM to help track down sections you care about; if so, you probably want to download the pdf, TeX source, or html version from the paper's arXiv page. (In my experience, if you just give the LLM the link it will only be able to read up to around section 3.5, so it’s worthwhile to download and upload even if you want to use the html version.)

Some sections I found particularly interesting:
* Section 3.5.2 on J-space ablation: If you read it carefully, you realize the thing they are ablating is not actually the J-space component as defined elsewhere in the paper. It may be interesting to discuss how significant the difference is.
* Section A.9 (in the appendices) discusses extending the Jacobian lens to multi-token concepts.

Note on consciousness: Part of the appeal of this paper is its comparison to one notion of human consciousness, so I don’t want to rule out discussing the philosophical implications if attendees are so inclined. However, as moderator I intend to ensure the majority of the discussion is about the technical aspects. I also want to make sure any philosophical discussion is informed by the technical aspects of the paper.