2026-10-11 17:10 UTC

Independent reproduction will determine whether Torchwright can compile Doom’s deterministic rendering algorithm into untrained, standard transformer weights that generate usable frames from encoded game state.

state: expiredheat: lowuncertainty: highknownscott: mediumcompiler-generated-models neural-rendering local-inferenceTorwright

What is this?

Torwright is presented as an effort by Rob Porter to port Doom’s deterministic renderer into the weights of a standard transformer, without training, so encoded game state can be converted into usable frames through local inference. A Hugging Face checkpoint is reportedly included, making independent reproduction the key test of whether the claimed compiler-generated model actually works. No web results were supplied, and the evidence titles alone do not establish the implementation details, performance, or reproducibility of the claim.

Why it matters to Scott

The radar already tracks this exact unresolved development in `radar:torchwright-training-free-graph-compilation`. Successful independent reproduction would bear directly on Three Substrates’ characterization of neural weights as “gradient-only,” potentially requiring a distinction between learned weights and compiler-authored weights, but the supplied evidence does not yet verify the claim.
ip:concept.three-substratesip:source.the-third-substrate-ebookip:concept.compiler-cascaderadar:torchwright-training-free-graph-compilation
queries asked of Scott's wikis
  • compiler-generated transformer weights
  • programs compiled into neural networks
  • deterministic algorithms encoded in model weights
  • untrained transformers as executable programs
  • local inference for neural rendering
  • reproducibility of weights-as-code claims

Measured heat

no measured readings yet — the hourly heat pass fills this in

How the heat travelled

no chain yet — the hourly chain pass fills this in

Evidence (3) — ⭐ canonical anchor

sourceobjectauthorscorecomments
🟠 redditDoom running on an LLM -- Hugging Face checkpoint included
LocalLLaMA
notforrob25445
🟧 echo.blog ⭐Rob Porter’s original write-up, “Doom, compiled into a transformer,” says: “I ported Doom’s renderer into an ordinary LLM architecture,” witRob Porter——
🟠 redditI compiled Doom's renderer into a 21B-parameter transformer -- no training anywhere [P]
MachineLearning
notforrob19233

Interpretation history

Decision trace