levelsio had Claude decompile Battlefield 1942 for $100
Pieter Levels says Claude Code disassembled 57,000 x86 functions and emitted 100,000 lines of C to get Battlefield 1942 running in a browser, for about a million output tokens. Meanwhile poteto says adversarial code review is a waste of tokens, and Theo hard disagrees.
The big disagreement
poteto calls adversarial review a waste of tokens, Theo hard disagrees
poteto said he now considers adversarial reviews mostly a waste of tokens, preferring empirical plus formal verification, and has cut his interrogate skill back in pstack with plans to maybe drop it. His complaint: agents get lost in nitpicks and YAGNI arguments. Ryan Florence said human code review was always a waste too, which is where Theo drew the line.
Code review was one of the best ways to level up the team, both by learning through reading code and disseminating lessons through reviewing other people's code
Receipts
levelsio had Claude Code decompile Battlefield 1942 into a web build
Pieter Levels says Claude Code disassembled the BF1942 shareware demo into 57,000 x86 functions, translated each to C, emitted 100,000 lines of it, and compiled to WebAssembly with Emscripten. About 300 Windows and Direct3D calls were rewritten to WebGL 2 and OpenGL, plus a WebSocket network stack for multiplayer.
I have no idea how it did any of this but it did it and it seems to work!
Model and agent watch
MIT study: GPT-6 Astra edited 0.016% of a model's weights to fix its math
Rikiya Takehi says GPT-6 Astra directly edited Qwen3.5-0.8B's weights to improve math up to 6x and instruction following 3x, with no gradients or training, touching 0.016% of weights. Per a summary of the MIT study, the model traced the small model's reasoning loops to two layers and its MATH-500 score went from 7.9% to 48.1% with sampling held fixed.