Theo's setup sequence
- Budget for substantial model usage. Theo's "step zero" is having subscriptions he is willing to exhaust; he says viewers unwilling to spend that way should watch out of curiosity.
19:0820:02 - Pick a desired outcome with extensive tedious work. He uses a 100,000-plus-file rewrite as an example of work humans may avoid but agents can pursue.
17:5218:45 - Define an easy success/failure check. For TS Rust, compare real projects and deliberately injected errors with the reference compiler.
20:0220:25 - Prefer programmatic pass/fail checks. If those are unavailable, use another agent to assess results; requiring human judgment every iteration limits autonomous progress.
20:5021:09 - Supply the tools and hardware required to execute and verify the task. For the game, he cites Blender, browser inspection, and potentially a GPU and graphical desktop. Requirements change with the task and parallelism.
21:36
Keep the cost categories separate
- Theo values TS Rust's total work at roughly $38,000 in API-equivalent usage and regards the saved build time as potentially worth that amount. He calls the economics of his other experiments much less sensible.
19:08 - He says the early working TS Rust build that achieved a 2× speedup represented about $500 of subscription usage. He recommends subscriptions rather than paying API rates for these experiments.
19:34 - Later he estimates the port's API-equivalent usage at about $40,000 versus $500–$800 of subscription usage, and claims a $200 plan yielded roughly $12,000 of usage. These are his estimates in the video, not guaranteed subscription entitlements.
22:28 - Model usage excludes the compute used to compile and test code. That separate expense becomes central in topic 6.
24:14
Why verification must scale
- Theo reports roughly 1,800 revisions and 7,884 commits during TS Rust's loops. He gives both figures without defining their relationship. At that volume, waiting for a human to judge each attempt would defeat the workflow.
20:50