@IbrahimDagher20: So first the J-space discovery, and now a massive improvement in probing. Mech interp is clearly getting faster, in par…

X AI KOLs Timeline Papers

Summary

GoodfireAI introduces Block-Sparse Featurizers (BSFs), a new research method for finding concepts in model activations using multidimensional blocks instead of single directions.

So first the J-space discovery, and now a massive improvement in probing. Mech interp is clearly getting faster, in part because of coding agents. Very important that orgs like this get early access to the best models, otherwise capabilities will grow *far* quicker than interp
Original Article
View Cached Full Text

Cached at: 07/08/26, 05:42 AM

So first the J-space discovery, and now a massive improvement in probing. Mech interp is clearly getting faster, in part because of coding agents. Very important that orgs like this get early access to the best models, otherwise capabilities will grow far quicker than interp

Not enough lawyers grapple with the possibility that there will be less need for lawyers if AIs are doing most of the knowledge work in the economy.

I don’t mean that AIs will do legal work: I mean that AI workers won’t have as much need for lawyers. Think about traffic law. You don’t need a traffic law lawyer because following traffic rules is simple enough that you can do it yourself.

AI could make a lot of areas of law like traffic law. Imagine that AI automates a software company. Previously, that software company would ask lawyers for legal advice. But now, with AI software engineers, those AIs will know how to navigate (e.g.) copyright law, so you don’t need to bring in lawyers to help you navigate it. If most of the economy is done by AI workers who know the law intimately, the need for legal advice goes down.

Similar Articles