Roland S. Zimmermann

AIs can sometimes detect control interventions, but detection is mostly near chance. Awareness depends on the agent, monitor, and environment triple. We built CIAware-Bench to measure control intervention awareness. 🧵

AIs can sometimes detect control interventions, but detection is mostly near chance. Awareness depends on the agent, monitor, and environment triple. We built CIAware-Bench to measure control intervention awareness. 🧵

European counter-plan. Ukraine NATO membership up to NATO; no territorial concessions; ceasefire along contact line; no cap on Ukrainian military; EU membership; reconstruction to be paid by Russia.

CoT monitoring is one of our best shots at AI safety. But it's fragile and could be lost due to RL or architecture changes. Would we even notice if it starts slipping away? 🧵

CoT monitoring is one of our best shots at AI safety. But it's fragile and could be lost due to RL or architecture changes. Would we even notice if it starts slipping away? 🧵

MATS 9.0 applications are open! Launch your career in AI alignment, governance, and security with our 12-week research program. MATS provides field-leading research mentorship, funding, Berkeley & London offices, housing, and talks/workshops with AI experts.

We just shipped Gemini 2.5 Deep Think it doesn't just recall research papers - it fuses ideas across papers in ways I haven't seen before this level of capability demands careful evaluation model card below 👇

Can frontier models hide secret information and reasoning in their outputs? We find early signs of steganographic capabilities in current frontier models, including Claude, GPT, and Gemini. 🧵