Posts

A 20-hour research sprint from September 2026, done for my MATS Winter 2027 application (Neel Nanda stream). Code available on request. Papers: Thought Anchors (Bogdan et al., 2025) and Value Leakage (Betley et al., 2026). Summary I used Thought Anchors, the sentence-resampling method from Bogdan et al. (2025), to locate where value leakage enters the chain of thought of Qwen3.6-35B-A3B on the Donation Bet task from Betley et al. (2026). I wanted to find out whether the leak enters at a sentence I can point to, whether that sentence is the moral statement about the donation or something else, and whether editing it changes the leak....

The question Betley et al. (2025) showed that fine-tuning on insecure code makes models misaligned across unrelated tasks — advocating deception, expressing contempt for humans — but their analysis was purely behavioral. You could see that something broke, but not what changed inside the model. Lu et al. (2026) built a tool for measuring a model’s “persona” from its internal activations: 240 trait vectors (like skeptical, cautious, deceptive) and an Assistant Axis capturing how “assistant-like” the model is being....

Here are my predictions for where AI will be by December 31st, 2026. Take the survey yourself at forecast2026.ai

AI has been superhuman at narrow things for decades. Chinook was the first program to win a human world championship in 1994. Chess fell in 1997. Go in 2016. Poker in 2017. Each time, we moved the goalposts and said “but that’s not real intelligence.” The thing is, capabilities don’t arrive uniformly. They poke through human-level performance one domain at a time. If you imagine a radar chart with hundreds of axes - one for each cognitive task - AI has always been this spiky shape....

Intelligence is coming!