Séb Krier
@sebkrier
🪼 AGI policy dev & strategy @GoogleDeepMind | rekkid junkie, dimensional glider, deep ArXiv dweller, interstellar fugitive, uncertain | 🛸
⚠️🐙 I prepared a mix for @Restless_Egg on @noodsradio - a mix of ambient techno, IDM, experimental breaks and other weird stuff. Explore the sonic latent space here: noodsradio.com/shows/restless…
WaPo's Monkey Cage was a significant development in journalism because it bridged the gap between academic political science research and mainstream news media. I feel like we need the same thing now helping connect random Substacks, X threads, & microblogs with mainstream news.
Do reasoning models like DeepSeek R1 learn their behavior from scratch? No! In our new paper, we extract steering vectors from a base model that induce backtracking in a distilled reasoning model, but surprisingly have no apparent effect on the base model itself! 🧵 (1/5)
Absolute nonsense. My community/friends are so much smarter and interesting than would be possible around where I live. I am extremely grateful for the internet, snd twitter, existing.
everybody's in the replies like "but i found all my communities and friends with social media!!" sure but the counterfactual isn't u being isolated, it's returning to the era where irl socializing was common and easy and ur best friend was some guy u met in a park
Good overview from @stuartbuck1 on the chaos in science funding right now asteriskmag.com/issues/11/the-…
Does More Inference-Time Compute Really Help Robustness? This paper says no. More thinking ≠ more robust—especially when adversaries are watching. But this paper finds: open-source models like DeepSeek-R1, Qwen3, and Phi-reasoning benefit only if reasoning steps stay hidden.…
🚨 Olympiad math + AI: We ran Google’s Gemini 2.5 Pro on the fresh IMO 2025 problems. With careful prompting and pipeline design, it solved 5 out of 6 — remarkable for tasks demanding deep insight and creativity. The model could win gold! 🥇 #AI #Math #LLMs #IMO2025
Aligning with my and @AidanRMackenzie’s recommendation, the new EO calls for agencies to use a percentage threshold for determining whether federal funding should trigger NEPA :) The OSTP team et al. are real legends who really took the time to dig into this stuff. Very pleased.
Every day is NEPA day but today will be even more NEPA than most
language models are intention refiners, conceptual blenders, counterfactual simulators, distribution visualizers, mind shapers, and basically mirrors you can't ignore. remarkable we've mostly used them to simulate bonzi buddy so far
Our Aeneas AI model gives historians valuable new insights into ancient inscriptions & ancient history that may have taken years to uncover otherwise. Published in @Nature today: deepmind.google/discover/blog/…
Next time you have to explain why cognitive science shares theories and concepts with economics and finance: both have a reliance on notions of efficiently allocating and utilizing limited resources that have alternative uses. Cognition optimizes a "mental economy" and has to…
A few years ago I released a vinyl record with my friend @PiChambaud and hope to do another one soon :) discogs.com/release/218279…
share a piece of art lore about yourself
The way that Teachout's terrible NYT op-ed was aggressively mocked, the disappearance of MMT, the YIMBY's understanding deeply the costs of regulation, tariffs losing all support on the left... It feels to me neoliberalism is gaining steam on the left
Our new state-of-the-art AI model Aeneas transforms how historians connect the past. 📜 Ancient inscriptions often lack context – it's like solving a puzzle with 90% of the pieces lost to time. It helps researchers interpret and situate inscriptions in their past context. 🧵
Our new paper is out in PNAS: "Evolving general cooperation with a Bayesian theory of mind"! Humans are the ultimate cooperators. We coordinate on a scale and scope no other species (nor AI) can match. What makes this possible? 🧵
A half-joking theory for slower timelines - Zuck hires 90% of the best researchers thru raw cash - Meta is poorly organized, they don’t share a common mission - algo breakthroughs slow down bc all the top ppl are essentially rubber-roomed and can’t quit wo losing the money
We think transmission of traits (liking owls, misalignment) does NOT depend on semantic associations in the data b/c: 1. We do rigorous data filtering 2. Transmission fails if data are presented in-context 3. Transmission fails if student and teacher have different base models
I've written a new post about "ChatGPT psychosis". Includes a detailed timeline of events leading up to the latest incident with @GeoffLewisOrg Link below.
New essay: In defense of self-direction What Tocqueville, Aristotle, Humboldt, and Mill understood about human autonomy—and why the highest goods can’t be delivered, only pursued. 🧵