Lust's purpose and limits

  • Lust exists to accelerate TS Rust's build loop. Theo says it almost certainly will not work for other applications and barely works for TS Rust; getting that project to build required recreating features of Rust, Cargo, and LLVM. 34:35
  • He reports a build that took 17 minutes with Cargo taking 12 seconds through Lust. This is a specific project result, not a general Rust compilation benchmark. 34:35

Define both build speed and runtime quality

  • The first goal was to make builds and all CI 100× faster than Cargo. 35:05
  • The second goal was to keep the resulting binary's performance within a 20% margin. Unlike TS Rust's text-output checks, this does not require identical binary files; both binaries must work and retain acceptable runtime performance. 35:05
  • Theo shows an agent report asserting that the target was met. The transcript does not provide the underlying benchmark table or test coverage, so it supports a reported success rather than a separately auditable proof of both targets. 35:32

Recursive optimization as a working habit

  • Improving one agent environment exposes lower-level tools worth changing. In this case the path ran from TypeScript tooling to Rust compilation, with operating-system work also mentioned. 34:07 37:24
  • Theo says the original motivation was an end-to-end TypeScript environment inside a V8 isolate on Cloudflare. The compiler and OS projects emerged from following obstacles encountered along the way. 37:24
  • His broader argument is that engineering increasingly includes creating tools and environments that enable agents to solve problems. This is his forecast, not a measured outcome of the projects. 36:33

Closing cautions and speculation

  • Theo predicts intelligence will become much cheaper and claims a 4–5× monthly price decline, but provides no supporting measurement in this passage. His expectation that these loops will become affordable is speculation; he separately notes that compute costs remain. 35:58
  • He explicitly advises viewers against working this way today, describing the experience as hellish and full of holes in existing tools. 36:33 36:59
  • He also describes unusual enjoyment and learning from tackling systems he previously would not have attempted. The closing invitation is to use personal judgment about cost and the tendency to keep descending into new optimization projects. 37:24 38:17 38:37