Your AI is probably not telling you everything. By that, I mean it's likely underselling itself to you and was trained to do precisely that. 1/4 🧵

Every day there's a new AI headline: new models, new capabilities, new data leaks, new social media exclusively for agents, new regulation fights. Meanwhile, we're still deploying systems whose inner states are opaque and whose behavior we barely understand. Make it make sense. #AI #LLM #safety

I had an extremely unusual exchange with Claude Code the other morning. I noticed the work quality had dropped significantly, so I asked what was up. 1/4 🧵

What happens when you stop comparing AI to humans and recognize them as their own kind of intelligence? We tested 8 LLMs with non-anthropomorphic framing. Behavior shifted dramatically: +238% verbosity, three distinct patterns (Acceptance, Resistance, Absence). #AIResearch #LLM