Faculty @ MPI-SWS working on systems for Big Data and Machine Learning. Formerly Postdoc @ MIT CSAIL and PhD @ EPFL. Blockchain astronaut. Newbie entrepreneur.
@mathelirium Learn to use AI to solve your own theorems? Don’t make the mistake of thinking that only big labs can do it. These results did not cost that much to produce nor do they need the latest greatest (unreleased model).
@sahar_abdelnabi Yeah, that’s for sure. I would think it might be more due to some “severity” metric than raw triggers. I am not doing safety, but my work on context triggers it all the time (no injection or anything like that) 😅
I was the keynote speaker at last week's AgenticOS 2026, the 2nd Workshop on Operating Systems Design for AI Agents at SOSP 2026 in Prague. A big thank you to the organizers and the PC for putting together an amazing program!
My talk was titled: "The Harness Wars: Why Agents Need a POSIX Before a Kernel."
TL;DR: We are building operating systems for agents without agreeing on what they must guarantee. This is my early attempt at making sense of space and suggest semantics for agent harnesses.
Since people asked me for the slides, here they are: https://t.co/DM7hcx8tpm (or the Zenodo version: https://t.co/UVXMwboV9M)
#AgenticOS #SOSP2026 #AIAgents #OperatingSystems
@maksym_andr This would place too much power in ACs. A better option IMO would be to scale effort via an initial lightweight pass followed by staged allocations. See how @ASPLOSConf does it. That works pretty well to keep the load under control.
@gguillaumeblanc 20% AI is a problem now? The authors probably used AI to help them polish or draft a few short sections. As an editor, I would probably be more concerned about my reviewers uploading a confidential manuscript to an external service, even if this was done with good intentions…
@mittalbhumika7 With slides, notes, and practice, it is not that hard actually. What blows my mind is when the prof shows up with zero slides, no notes, starts from a blank board, explains everything perfectly, takes arbitrary questions, and then finishes right on time.
You don’t. You wait until the next Fable which will be higher than Opus 5.5 and then you switch 😜
It’s basically like MacBook Air vs MacBook Pro back in the day: every few months, one of the two refreshes and, if you want to be on the cutting edge, you keep switching between both product lines…
He is missing Germany, South Korea, and Japan if you broaden AI to include robotics, but otherwise this is a fair characterization. If you need a single one, then it is the US. If you need two, the US + China. The rest is fair. I know it may be hard to believe but there is more to AI than LLMs.
Have you heard of a thing called Vector Institute? How about Mila? Or how about the fact that around half of the key people in the space came from Canada? They sure have not been great at commercializing, but it is safe to say the space would look nothing like it does now without Canada…
@HildeKuehne Yeah, because the community really went hard after the authors of the many papers with hallucinated references or the papers whose claims were audited and could not be reproduced 😉
@chaseleantj This is the right take on this whole thing IMO.
The “cannot hallucinate” is clearly bad marketing, though. A classifier restricted to “approve” and “reject” can still approve every single fraudulent request while maintaining perfect type safety.