Alex Au @Mathlexi
Joined October 2026-
Tweets4
-
Followers0
-
Following15
-
Likes0
If anyone has any advice, opportunities or feedback it would be greatly appreciated, opportunities like BlueDot, SPAR, MATs, ARENA, METR etc are on my radar but I feel far from ready, any direction would be appreciated. DMs open. #AISafety #MechanisticInterpretability
Doing AI safety here and there in my spare time outside my job no longer feels enough, in light of pivoting my career I've taken the straightfoward route of replicating and extending existing work. devalexau.com/blog/role-conf… Would love advice and feedback! Thread: #AIsafety
In this preliminary evaluation, clear malicious tool-call attempts fell from 9.5% to 4.9%. The work is very early and very limited, I am currently working on a followup with a docker sink to emulate tool calling rather than an LLM judge of output, giving a more binary measure.
Prompt injection may partly be a role-confusion problem: untrusted tool content can become represented more like a user instruction inside the model. I wanted to test whether we could intervene directly on that representation.
Benjamin Todd @ben_j_todd
20K Followers 214 Following Founder @80000Hours Use your career to tackle the world's biggest problems 🦑
Ajeya Cotra @ajeya_cotra
29K Followers 506 Following Helping the world prepare for powerful AI. Risk assessment @METR_evals (opinions my own). Blogs: Planned Obsolescence (AI), Good Bones (whatever's on my mind).
Jan Leike @janleike
143K Followers 335 Following AI research @AnthropicAI. Previously OpenAI & DeepMind. Optimizing for a post-AGI future where humanity flourishes. Opinions aren't my employer's.
Kaley Brauer @kaleybrauer
7K Followers 938 Following AI safety, astrophysics, interpretability! 丨astrophysics postdoc fellow at @Harvard, AI fellow at MATS @openai, prev physics PhD @MIT 🌌
Rob Miles @robertskmiles
67K Followers 871 Following Explaining AI Alignment to anyone who'll stand still for long enough, on YouTube and Discord. Music, movies, microcode, and high-speed pizza delivery
Rob Bensinger ⏹️ @robbensinger
20K Followers 558 Following @MIRIBerkeley. If we build opaque superhuman AIs, it's likely to get us killed. The least-bad solution is an international agreement like https://t.co/BsSp5beZhp.
Aran Komatsuzaki @arankomatsuzaki
197K Followers 446 Following Sharing AI research. Early work on AI (GPT-J, scaling, MoE). Ex ML PhD (GT) & Google.
Nathan 🔎 @NathanpmYoung
35K Followers 4K Following Lead, Goodheart Labs. AI-written Community Notes (world first). Part-time forecaster @swift_centre. Capital case tweets are literal, others less.
Michael Nielsen @michael_nielsen
123K Followers 5K Following Searching for the numinous 🇦🇺 🇨🇦, currently live in 🇺🇸 Research @AsteraInstitute https://t.co/maezekzk4D
typedfemale @typedfemale
49K Followers 568 Following a really exciting new account "advanced pytorch user" - @cHHillee alt: @typedalt
Anthropic @AnthropicAI
1.8M Followers 2 Following We're an AI safety and research company that builds reliable, interpretable, and steerable AI systems. Talk to our AI assistant @claudeai on https://t.co/FhDI3KQh0n.
Eliezer Yudkowsky ⏹... @ESYudkowsky
244K Followers 101 Following The original AI alignment person. Understanding the reasons it's difficult since 2003. This is my serious low-volume account. Follow @allTheYud for the rest.
Neel Nanda @NeelNanda5
50K Followers 126 Following Mechanistic Interpretability lead DeepMind. Formerly @AnthropicAI, independent. In this to reduce AI X-risk. Neural networks can be understood, let's go do it!You might like
