Jev CEO: I made ChatGPT, now I'm building what's next
TypeSafe AI CEO Diogo Almeida explains why post-training techniques like RLHF optimize for pleasing human users rather than execution. He distinguishes human-in-the-loop assistance from autonomous automation and outlines TypeSafe's approach to decision-making models.
Original by AI EngineerGetting startedIntermediate18 min 4 secPublished
Diogo argues that preference-based post-training can reward pleasing answers rather than accurate probabilities.
The talk distinguishes assistance with a human checking the output from autonomous execution in software.
TypeSafe’s proposed alternative optimizes probability-bearing decisions for software; the talk presents the creator’s thesis rather than an independent validation.
Worth knowing
The talk presents the conceptual critique and architectural thesis for TypeSafe's upcoming model, but does not provide benchmark metrics, production API specs, or implementation code.