TypeSafe AI co-founder and GPT-4 co-author Diogo Almeida argues that RLHF's design inherently limits AI to assistance tasks, and proposes a new post-training optimization approach focused on calibrated decision-making to achieve true automation. Almeida, a member of the OpenAI post-training team who co-authored GPT-4, ChatGPT, and InstructGPT, calls the ChatGPT era a "weird detour" and says he is one of the few people at OpenAI who openly dislikes ChatGPT as a direction. Recorded mid-2026 for a live audience, his talk explains why AI seems to be going both insanely well and insanely poorly. His central claim: today's AI is an assistance technology, not an automation technology, because RLHF optimizes for human preference. The same models that approach research-grade math problem solving cannot handle customer service without humans in the loop, because one task type is designed to please a human and the other is designed to remove the human. Almeida argues that getting from assistance to automation requires a new optimization objective—calibrated decision-making—which he says is "definitely not RLVR" and which TypeSafe is building from scratch. He maps the industry into two camps: one holding that AI is going insanely well, with every benchmark surpassed and autonomous operating time growing exponentially; the other holding that AI is going insanely poorly, a bubble generating no real value, propped up by circular financing, whose only visible output is