jev: on | Diogo Almeida founder of Typesafe AI, Runtime 2026

Описание: AI can write code, run deep research and chat like a human, so why can't it do basic data entry or approve a refund? Diogo Almeida, CEO and co-founder of TypeSafe AI and a co-author of RLHF, InstructGPT and ChatGPT, says it's because every LLM today is optimized for assistance and pleasing the human in the loop. None are optimized for automation, where the only goal is to be correct.

In this talk from Runtime 2026, Diogo explains how TypeSafe AI forked the research tree to build Jev, its System 1 models made for the inner loop of software. He covers why benchmarks hurt the industry, and why he thinks this marks the end of "software is over" and the start of a SaaSapalooza.

Timestamps

0:00 Intro
2:05 Assistance vs. automation
2:55 RLHF put the human in the loop
4:00 Forking the research tree: System 1 models
6:15 Expanding what software can express
7:30 Why benchmarks are bad for the industry
8:20 Scaling to 1T tokens with Modal
10:25 Software is so back
11:15 "Ilya in every if statement"
11:40 Thank you to the early believers