The conceptual model

  • Theo uses "hill climbing" for repeated attempts guided by feedback that distinguishes success from failure. His introductory example is choosing the winning move on a tic-tac-toe board. 8:28 8:53
  • In his simplified account of training, pretraining supplies knowledge and reinforcement learning shapes behavior such as tool use, testing a hypothesis, and trying again after failure. 9:19 9:44
  • RLVR means reinforcement learning with verifiable rewards: training uses tools that can programmatically check whether the model succeeded. His example exposes five transformation cases while retaining 500 hidden cases. 11:12 11:44
  • The long-running project strategy borrows this structure: give an already-trained agent concrete checks and let it revise its code. The described project loop is code iteration, not a claim that the user is retraining the model's weights. 12:15 12:40

TS Rust's verification recipe

  1. Choose real TypeScript projects and pin the reference TypeScript version. Record its emitted JavaScript and diagnostics. 11:44 12:40
  2. Run the Rust port on the same inputs and require exactly matching outputs. This provides an explicit completion condition. 12:40
  3. Insert random errors into projects and compare reference diagnostics with the port's diagnostics. A port that misses those errors fails the check. 12:15
  4. Once behavior matches, set a performance goal while preserving all existing validation and tests. Theo used "two times faster." 13:09

Reported outcome and lesson

  • Theo says TS Rust began working in a day and improved over several weeks. Its initial working version was slower than TSGO; he then instructed the agent to continue until it was 2× faster and reports that it did. 13:09 13:37
  • His explanation is that tasks with consistent, verifiable inputs and outputs resemble the environments used to train coding behavior. A concrete success test gives the agent something it can keep improving without repeated human judgment. 12:15 13:37