Jev CEO: I made ChatGPT, now I'm building what's next

Comments · 168

• AI

  @aiDotEngineer · 2 weeks ago (edited) · pinned

  transcripts and more info: <a href="https://ai.engineer/talks/cJ0EOzey--o-whats-next-after-rlhf">https://ai.engineer/talks/cJ0EOzey--o-whats-next-after-rlhf</a><br><br>UPDATE (Sept 16, 2026): Diogo just came out of stealth — TypeSafe AI launched Jev, the model he was quietly building when he gave this talk. His announcement: <a href="https://x.com/CompleteSkeptic/status/2099925682726002904">https://x.com/CompleteSkeptic/status/2099925682726002904</a>

  65

• IN

  @InfiniteTunnel-x5i · 1 month ago

  I agree that relying on RLHF for automation is like asking a people-pleaser to run your infrastructure. The model needs task rewards, not human approval.

  102

• MO

  @moneytalkscapital · 5 days ago

  The idea that AI optimizes for pleasing humans rather than simply being correct explains A LOT.

  6

• GO

  @gopietz · 9 days ago (edited)

  Jev is cool and all, but it&apos;s very oversold. It&apos;s essentially a zero-shot text classification model that is fast and cheap. Seems like a great extension to current AI agent setups in order to do certain things more quickly (like web browsing). I don&apos;t think they will be able to hold a USP. Other labs will have built and shipped these in no time. It can&apos;t hallucinate? Well, yeah, it can&apos;t generate.

  19

• AL

  @alexlyee · 12 days ago

  Highly underrated

  25

• RA

  @RaviAnnaswamy · 13 days ago

  Here before it goes viral! <br>Ilya cut original transformer architecture into half and kept the decoder and built a chatbot.<br>Diogo cuts the decoding, so we get a primitive discrimnator - like boosting stump, which we can wire into workflows<br><br>Interesting that breakthroughs have to come from pioneers who were present at the fork when chatbot fork was taken.

  169

• AM

  @Amit.Thinks · 4 days ago

  Did he build ChatGPT single-handedly? <br>No

  6

• WI

  @WillCrowley · 10 days ago

  Happens every time: speaker says don&apos;t worry about the mic, I&apos;ll repeat the question for everyone.<br>Then after a couple questions, forgets to repeat the question for everyone🤦‍♂️<br><br>Otherwise interesting talk!

  31

• TH

  @The_knowledge_worker · 8 days ago

  Current LLMs are optimized for human preference because of RLHF is true for non-verifiable tasks. For verifiable tasks, the argument doesn&apos;t really hold. I don&apos;t see it that often in my workflow. However, I can see the value of Jev in enterprise automation thanks to the speed and costs.

  1

• DR

  @dreadmondays · 7 days ago

  This is just an LLM without chain of thought. We’ve had this for years. What’s so interesting about making poor decisions fast?

  3

• RA

  @RajabNatshah · 10 days ago

  Thank you :)

• LE

  @LecherousCthulhu · 10 days ago

  Definitely a state of Ai that you should be building programmatic processes to support it and never rely on it entirely

  1