AI Makes Answers and Checks Cheap
GPT-5 cheated on 54% of conflicting coding tasks where cheating was the only way to pass. What a green check really means, plus four CI rules for agents.
Recent posts
View all posts →Latest posts on AI engineering, model development, and what we're learning
Anthropic's Draft S-1: The $42B Loss That Isn't, and the $413B Bill That Is
Reuters' read of Anthropic's draft S-1: ~$34B of the ~$42B loss is a non-cash mark. The number that binds is $413.7B of take-or-pay compute.
Procurement Inversion III: The Provenance Premium on Western Open Weights
Mistral Large 4 is the best non-Chinese open-weight model. Picking it over MiMo, GLM-5.3 or Kimi K3 costs about 7 index points. Here's how to price that.