Director of CeSIA, OECD AI Expert
Co-lead, Global Call for AI Red Lines, featured in 300 media mentions, including Le Monde, NYT, NBC
Co-author, AI Safety Atlas
Feb 12 • 8 tweets • 3 min read
We are approaching the end of time. Gemini DeepResearch is just incomparably better than most of what my students produced at the end of their course last November (the best ML master in France). Some comparisons
On the subject: "Best counter arguments to the X-risk discourse"
Here is a comparison focusing on an argument shared in the 3 reports:
a. A student, one month to do the project
b. OpenAI's DeepResearch, 5 min
c. Gemini 2.5 DeepResearch, 30 min
May 26, 2025 • 25 tweets • 8 min read
Here's the 80/20 playbook for mitigating AI scheming.
Then... I'll tell you why these techniques are largely insufficient. 🧵
First, what's scheming?
It's when an AI strategically hiding its true (misaligned) goals during training or evaluation, only to pursue them once deployed, potentially seeking power or resisting shutdown.
