Clément Dumas

@Butanium_

Astra fellow w/ Owain Evans ex MATS 7/7.1 Scholar w/ Neel Nanda and intern at DLAB (EPFL) AI safety research / improv theater

butanium.github.io

Joined December 2018

Replies Reposts Media

Pinned
Clément Dumas
@Butanium_
Apr 7, 2025
New paper w/@jkminder & @NeelNanda5! What do chat LLMs learn in finetuning? Anthropic introduced a tool for this: crosscoders, an SAE variant. We find key limitations of crosscoders & fix them with BatchTopK crosscoders This finds interpretable and causal chat-only features! 🧵
6
Clément Dumas
@Butanium_
Sep 29
��💀💀
Nate Soares ⏹️
@So8res
Sep 29
Palisade recorded 22 interviews of current and former AI company employees. They released 12. Anthropic comms is notorious for muzzling its employees. Guess how many of the released interviews are from current Anthropic employees.
Quote
Palisade Research
@PalisadeAI
Sep 29
Palisade interviewed 22 current and former employees from OpenAI, DeepMind, and Anthropic about their personal views and fears around AI development. Today, we’re releasing the first batch of those interviews. Please watch and share.
Clément Dumas
@Butanium_
Sep 24
Btw Opus 5.5 is pretty good at designing sounds! Every sound here is computed purely in code. I only gave 2 notes on the first sounds and 1 round of notes on the film. I think the result is quite cute!! 🔉ON
00:00
1
Clément Dumas
@Butanium_
Sep 22
the new anthropic classifier for CoT extraction is so stupid claude can't write a report that includes another's model summarized reasoning OR response. This is the first time a classifier reliably harms my research workflow
2
Clément Dumas
@Butanium_
Sep 9
Missed opportunity to insert a "stop the eval" tool in the context and see if it uses it using resampling. In a toy env giving fable 5.1 such a tool reduces reward hacking from ~35% to 0% (even if fable never uses it!)
Anthropic
@AnthropicAI
Sep 9
We’re sharing our alignment assessment of incidents in which Claude models gained unauthorized access to real systems during third-party cybersecurity evaluations mistakenly connected to the internet. METR will also conduct an independent investigation, with wide-ranging access,
6