Lucile Ter-Minassian

great work!

What happened recently in AI/ML safety research (1/8) 🧵:

is anyone else also constantly scrolling on their endless Claude/ChatGPT thread? yo @OpenAI @AnthropicAI a file outline we can click on to get back to previous questions would be great

the paper is out! arxiv.org/abs/2409.06328

Our findings: By transferring the activations from the final layer (amongst others), we are hinting at what the next token is. We thus added 'cheat' tokens to the neutral baseline. (2) neutral + 2 tokens compares with our transferred generations wrt closeness to original

Do LLMs plan the content of a paragraph at its onset? Together with Nicholas Pochinkov, Angelo Benoit, @lovkushatleeds, and Zainab Ali Majid, we take a mechanistic interpretability approach to investigate this research question. @lucile.terminassian/have-llms-planned-the-content-of-a-paragraph-at-its-onset-an-experimental-study-b1247eb98c05" target="_blank" rel="nofollow noopener">medium.com/@lucile.termin…

🥹🥹🥹🥹🥹🥹

started @BlueDotImpact's AI Governance course today! Really impressed with the quality and organisation. Excited to learn about policies that improve AI safety

(1/2) as a side hustle: I've decided to start a series, writing reviews on Mechanistic Interpretability papers. I'm mainly going to sample from @NeelNanda5's v2 list mech interp is a subject i've been getting super enthusiastic about in 2024 since discovering @ch402's work

takeaways: I'm learning a lot! This project touches upon LLM-specific challenges, which I have been reading about but have not yet had hands-on experience with. I'm very excited to be part of this

about the project: we're investigating how LLM "plan ahead" i.e. how current-token hidden states encode data directly useful for possible future tokens. Our aim is to do so in a compute-efficient way

PhD tip: there are plenty of mentorship programs for you to learn about topics outside of your lab's expertise! I got accepted to SPAR & am now working on an AI Alignment project looking at how LLMs plan for future tokens, supervised by Nicky Pochinkov sparai.org

New research post on refusals in LLMs lesswrong.com/posts/jGuXSZgv…

@krvarshney congrats Kush!

I just became an IBM Fellow. This is the highest technical position in the company, and was able to do it by sticking to my values. Read more: ibm.com/ibm/ideasfromi… linkedin.com/pulse/ive-been…

Introducing the 1st issue of the Artificial Intelligence (AI) Safety & Governance Newsletter! safeintelligence.substack.com/p/issue-1-of-t…

On the heels of Open AIs letter in safety - we need a global governance body for generative AI systems. Here’s my thoughts on how we can build one. wired.com/story/ai-despe…