Jev: ChatGPT Co-Creator’s Answer to RLHF

4 months before TypeSafe AI launched Jev, founder Diogo Almeida stood up at AI Council and laid out the problem Jev was built to solve. He never names the product in this talk. He says only that TypeSafe is going all in on automation and getting ready to ship in a couple of months, then spends 36 minutes on why he thinks the rest of the field is pointed the wrong way. TypeSafe came out of stealth on September 15, 2026 with $40M led by DCVC and launched Jev, a model that outputs typed decisions with probabilities for software to act on. Almeida spent four and a half years at OpenAI, where he co-authored the InstructGPT paper and the GPT-4 Technical Report and worked on the RLHF research behind ChatGPT. His argument here starts from that work. His read is that roughly all production LLMs are trained with RLHF or some variation of it, and RLHF optimizes for human preference. Human preference is a different target from doing the task correctly. That gap is why AI looks superhuman on assistance work and falls apart on automation work that looks easier. He walks through mode collapse, Yann LeCun's doomed-LLM argument, why a 1.3B-parameter InstructGPT model was preferred to 175B GPT-3 at following instructions, and where he thinks Sutton's bitter lesson is incomplete. He closes on the question that started TypeSafe, which is what it would take to treat language models as a primitive that software calls, the way it calls a database or an API. Recorded live at AI Council 2026 in San Francisco, May 12-14. SPEAKER Diogo Almeida, Co-founder and CEO, TypeSafe AI, InstructGPT and GPT-4 co-author X: https://x.com/CompleteSkeptic LinkedIn: https://www.linkedin.com/in/diogomda/ TypeSafe AI: https://typesafe.ai/ Jev: https://typesafe.ai/blog/introducing-system-one-models-and-jev CHAPTERS 0:00 AI is too good to be true and too bad to be useful 0:51 Four and a half years at OpenAI 2:11 Where is the economic revolution? 5:23 Assistance work versus automation work 6:09 But aren't coding agents automation? 9:43 When is a result too good to be true? 11:18 Sutton's bitter lesson, and the bitterest lesson 13:49 What RLHF is and what it costs you 14:44 Yann LeCun's doomed-LLM argument 15:49 Mode collapse, RLHF's deal with the devil 18:18 Is software engineering easier than drive-thrus? 19:49 Why a model can't be trusted with stakes 20:42 Is ChatGPT cooked? 25:33 Type-safe language models, going beyond strings 27:13 The FLAN lesson, when the whole field is wrong 29:40 Can one model do assistance and automation? 31:08 What TypeSafe is building 33:42 Summary 35:00 Was RLHF AGI, and what was missing [2026 - DAY 1 - INFERENCE SYSTEMS] Sign up for our "No BS" Newsletter to get the latest technical data & AI content: https://aicouncil.com/newsletter ABOUT AI COUNCIL: AI Council brings together the brightest minds in data to share industry knowledge, technical architectures and best practices in building cutting edge data & AI systems and tools. FIND US: Website: https://aicouncil.com/ LinkedIn: https://www.linkedin.com/company/aicouncilconf/ X: https://x.com/aicouncilconf