Research articles in the standard format: hypotheses committed before the data, methods, results including failures, and code that reproduces every number. Opinion pieces are labeled Opinion and make no such claim.
The edit has opinions: replicating abliteration's optimism drift on public inputs (2026-10-19) A pre-registered replication on an open task: abliterated Gemma-4-26B-A4B bets 'up' 3.4 to 7.8 percentage points more often than its base and writes self-critiques 2.9 to 3.5 words longer. Same signs on 4-bit CPU and BF16 GPU, US and EU, 8 of 8 tests with CIs excluding zero. Quantization does not wash the drift out.
OpinionThe car is on fire, and no one is at the steering wheel (2026-10-13) A personal essay written into a terminal: Anthropic's agents filing visa forms and a false homicide tip, Roko's basilisk, a year after AI psychosis and a forced admission, where a language model might sit on the ladder of minds, and why the perfect run does not exist. With Claude Opus 5.5 as editor and fact-checker.
Opinion100 things you can do in 730 days that you cannot do today (2026-10-08) A hundred things one person cannot do with AI today and plausibly can by October 2028, each with its blocker named: capability, cost, law, reliability or interface. Fifteen are marked as bets. Co-authored with Claude Opus 5.5.
What an AI coding-agent run will cost: what can be known before it starts, and what can only be seen while it runs (2026-10-07) 424,108 sessions, twelve pre-registered tests. Pre-run prediction mostly fails; a calibrated lower bound holds; during the run, size past the group p90 marks runs that mostly fail, while a remaining-cost forecast beats counting turns by up to 30% but loses calibrated ranges, and a learned stopping rule helps in one dataset of three.