Koby Lewis

Are Per­sona Fea­tures Co­her­ent?

Can the As­sis­tant Axis Pre­dict Re­ward-Hack­ing?

• Koby Lewis 24 Oct 2026 2:58 UTC

  1 point

  0

  on: “Cor­rect An­swer Fea­tures” Can­not Ex­plain Mul­ti­ple Choice Capabilities

  Update: This research was accepted as a BlackboxNLP 2026 extended abstract!

• Koby Lewis 21 Sep 2026 20:54 UTC

  1 point

  0

  on: Re­ward Hack­ing Without Egre­gious Misal­ign­ment in an RL-Only Setting

  What is the “benign” category for ImpossibleBench?

You Should Think About What Needs to Go Right + My List

“Cor­rect An­swer Fea­tures” Can­not Ex­plain Mul­ti­ple Choice Capabilities

Port­ing MACHIAVELLI To Inspect

• Koby Lewis 29 May 2026 0:39 UTC

  3 points

  0

  in reply to: Brendan Long’s comment on: A Call for Bet­ter Type Hints in AI Safety Tooling

  Yeah, overall I actually think TransformerLens does pretty well on type hints (and even better with v3 from what I’ve seen so far). run_with_hooks was just a particularly good example of what I’m talking about considering it’s one of the main entry points for the library, and the fix would be super easy.

A Call for Bet­ter Type Hints in AI Safety Tooling

• Koby Lewis 26 May 2026 18:23 UTC

  2 points

  0

  on: The Scal­able For­mal Over­sight Re­search Program

  Curious what your reaction to the argument that SPS and SFO more broadly may be dual use in a way that makes it counterproductive. Do you find it plausible that leaning into this research program could end up making (potentially misaligned) agents more capable at coding tasks (which would increase their ability to get stuff done in general)?

What Do You Mean by a Two-Year AGI Timeline?

Hel­i­cal Rep­re­sen­ta­tions of Turn Struc­ture in an LLM