Tom McGrath

[Linkpost] Play with SAEs on Llama 3

• Tom McGrath 24 Jan 2024 19:30 UTC

  4 points

  0

  on: Safety as a Scien­tific Pursuit

  Very much appreciate the link post—I’d been trying to write a summary/​contextualisation for LW and this is a much better one than I’d come up with.

  I’d be very grateful for the LW community’s thoughts (especially any pushback). I expect this will be the source of the strongest counterarguments.

• Tom McGrath 24 Jan 2024 19:28 UTC

  3 points

  0

  in reply to: niplav’s comment on: Safety as a Scien­tific Pursuit

  Thanks! I really like inductive vs deductive and would probably have used them if I’d thought of it.

[Paper] All’s Fair In Love And Love: Copy Sup­pres­sion in GPT-2 Small

• Tom McGrath 19 Nov 2021 16:06 UTC

  5 points

  on: “Ac­qui­si­tion of Chess Knowl­edge in AlphaZero”: prob­ing AZ over time

  I’m one of the authors on this paper—happy to answer any questions/​discuss if anyone is interested.

• Tom McGrath 19 Nov 2021 16:06 UTC

  4 points

  in reply to: Zac Hatfield-Dodds’s comment on: “Ac­qui­si­tion of Chess Knowl­edge in AlphaZero”: prob­ing AZ over time

  Thanks for the summary! Your first bullet point was my motivation for doing this. I think it’s important to test out interpretability ideas in more challenging domains.

  We didn’t really do much interpretability in this paper, this is more meta-interpretability in a sense (i.e. studying whether interpretability should in principle be possible). I’d say section 4 is worth a look, especially section 4.5 which covers fundamental and practical challenges to probing. Section 7 has some NMF analysis, and we open-sourced NMF factors which you might find interesting.