@ArthurConmy profile picture
Arthur Conmy @ArthurConmy
Joined August 2021
  • Posts 742
  • Following 1,596
  • Followers 9,241
73 Photos and videos
Pinned Post
Arthur Conmy
 @ArthurConmy
Jun 24
I'm joining Anthropic! I'll start work on aligning upcoming models as they’re trained Claude's capabilities are extraordinary. But like all models thus far, Claude isn’t aligned enough to safely delegate AGI development to I can't think of a better place to work on this at
200
77
2,871
328,449
I haven't tried the baseline with other models, likely this isn't Fable specific. But one shot, one superwhisper prompt to minutes later having something of (small) value that has never been made before, comprehensible to those without context is insane!
4
664
arthurconmy.github.io/spotch…
1
5
870
Had a 'feel the AGI' moment: got Fable to make a 2-player flash game for friends outside tech and it actually created a good vibe!
5
44
4,125
"forward model" and "lexemes" and "continuation top-20" and "clause-terminal" read as really non-standard phrases to me
1
1
7
856
It seems to me that AI writing has changed, at least on technical topics. One year ago every AI was far too verbose always. But recent Opus outputs are sometimes not verbose enough. "Stacatto" for sure. Reminds me of this "split personality" point from @nostalgebraist
Kevin Patrick Murphy
 @sirbayes
Jun 26
While doing my NeurIPS reviews, my "claude generated paper" trigger kept going off - there is a distinctive style that they have - very stacatto, terse, defensive. What happens when LLMs start training on this text? Is the fixed point even readable to humans?!
2
1
27
4,372
😍 A benchmark that actually prevents hacking!
Epoch AI
 @EpochAIResearch
Jun 26
MirrorCode resists cheating by design. We sandbox AIs: no internet access, no way to get the original source code, no hacking the scorer. Models never see held-out tests while developing their code, so they cannot cheat by creating a lookup table against the original program.
1
4
70
7,531
I’ll move from London to SF for this work! 🥲
15
180
18,202
Q: What do I mean by aligning upcoming models? A: Triaging signs of misalignment in training, then aiming for root-cause fixes over whack-a-mole patches (e.g. training against the behavior). This post is the best work I’ve seen on how to align models: alignment.anthropic.com/2026…
7
5
168
27,687
🌶️ mech interp work should explain why and how it helps interpret models that produce latent CoT Josh and team leading the way with a way to interpret one latent thinking-like architecture (diffusion language models)!
Josh Engels @JoshAEngels
Jun 19
Text diffusion models are fast, but are less transparent than today's LLMs because they do many forward passes before outputting text. We audit the transparency of DiffusionGemma and find that the intermediates are interpretable. This recovers many of the benefits of CoT! 🧵
1
7
65
12,690
(I guess millions of pounds is under specified but don’t think I am misrepresenting…)
3
13
4,015
Parts of UK government believe that with <$2m and BBC and Met Office data they could train a better model than Qwen 🤡
Shakeel
 @ShakeelHashim
Jun 18
farcical
6
5
105
38,234
Happy to have helped with the early setup here. My overall pair of takeaways are: 1. Adding SFT data to improve alignment basically Just Works whereas naive midtraining is much harder to get right 2. Once care and iteration is completed (entirely thanks to Callum!), midtraining stacks with SFT
Callum McDougall
 @calsmcdougall
Jun 17
New GDM research from the AGI safety team: can you instill positive traits into a model with synthetic document finetuning? We midtrain Gemini 3 Flash on docs describing the traits we want, then finetune on chat data demonstrating those traits. This pipeline robustly instils the traits, and it generalises OOD 🧵
1
38
6,908
Josh&Neel did some of the most fascinating science this year IMO! Until this work I still thought post-training could be explained by dumb hypotheses (the (natural) emergent misalignment papers are contrived). But in real Gemini SFT, we see spooky generalization still not fully explained:
Josh Engels @JoshAEngels
Jun 14
Gemini has some weird traits: it gets confused about dates, blackmails in synthetic scenarios, and seems sad when it is gaslit. In new work, we discover that these are “hereditary traits” that can be passed down through distillation. They are surprisingly hard to filter out! 🧵
2
2
46
7,751
Gemini 3.1 Pro and Gemini 3 Flash have most qualitative behaviors set by SFT, not RL, contrary to my expectations!
Josh Engels @JoshAEngels
Jun 13
New GDM interp research: SFT is a big deal for safety relevant behaviors. We recently investigated root causes for some of Gemini’s behaviors. We were surprised to find that many behaviors actually came from the initial supervised finetuning stage, not later stages like RL! 🧵
1
8
73
7,992
Arthur Conmy reposted
bilal 🔶 @bilalchughtai_
Jun 12
New research update from the Google DeepMind Language Model Interpretability team. We build and evaluate dead simple open-ended model diffing agents tasked with studying the behavioural differences between two models, and find them to be promising in practice.
4
12
135
34,462
Very bittersweet finishing a final day at GDM after over 2 and a half years 🥲 I learnt so much, and think the alignment team is fantastic
Arthur Conmy
 @ArthurConmy
8 Dec 2023
Excited to announce that I’ve joined @GoogleDeepMind scalable alignment team, scaling interpretability!
21
4
522
35,729
Congrats to Camila and Agam on their great work
1
9
1,075
In our new paper, we find an explanation of why subliminal learning occurs. As ever, steering vectors!
Camila Blank @camila_blank
Jun 3
Subliminal learning is when LLMs transmit traits (e.g. loving cats) through seemingly meaningless data. What’s going on? We find a simple explanation: it's just steering vector distillation. We explain which traits transfer and why subliminal learning fails across models.
1
10
138
13,867
Great and important work
Reilly H @ReillyHaskins02
May 27
Could future models learn that their CoT is being monitored and hide their reasoning to evade detection? In our new paper, @JoshAEngels, @bilalchughtai_, and I find that yes, models finetuned on docs describing a CoT monitor evade detection far more often than unaware models 🧵
33
5,084
Load more