I haven't read the paper yet but I'm pretty confident people are overhyping the Global Workspace paper.

This isn't even any fault of the paper. It's just kinda inevitable for any ML paper with a PR team and a cool video promoting it.

But it's especially true for a paper that is (1) interpretability and (2) uses a bold philosophical framing. The interp literature has a long history (pre-dating mech interp) of techniques going through a cycle of discovery, hype, criticism and skepticism.

Not everyone is overhyping the paper. Neel Nanda's review seems pretty sober. But I'm surprised to see Zvi writing this:

This feels like a major advance in our understanding of LLMs, both overall and our ability to understand any particular interaction.

The J-Lens is improvement of the logit lens. It's cool, but not a major advance that deserves this level of attention from people who do not typically read mech interp papers.

Once again, I have not read the paper and certainly hope to be pleasantly surprised when I do!