2026-10-11 18:02 UTC

Independent reproductions will determine whether Torchwright can compile nontrivial computation graphs directly into transformer weights that execute correctly without gradient training.

state: expiredheat: lowuncertainty: highnovelscott: nonemodel-compilation transformer-weights training-free-learningphysicsrob

What is this?

Torchwright is presented as a project by or associated with physicsrob that claims to compile computation graphs directly into transformer weights, producing executable behavior without gradient-based training. The supplied search results do not independently document Torchwright or verify that it works on nontrivial graphs, though ALTA provides adjacent precedent for compiling programs into transformer weights. Independent reproductions are therefore needed to establish its correctness, scope, and numerical reliability; the snippets also show that transformer compilation can encounter numerical inconsistencies, although that report concerns PyTorch’s different `torch.compile` system.

Why it matters to Scott

No intersection found in Scott’s wikis, and the radar has no prior page tracking Torchwright, physicsrob, or this specific model-compilation claim.
queries asked of Scott's wikis
  • programs compiled into neural network weights
  • transformers as programmable computers
  • training-free model construction
  • neural network compiler correctness and verification
  • symbolic computation encoded in model weights
  • alternatives to gradient-based learning

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 (2) — ⭐ canonical anchor

sourceobjectauthorscorecomments
🟧 hnShow HN: Compile computation graphs into transformer weights – no trainingphysicsrob10
🟧 echo.github ⭐Tor­chwright claims to compile computation graphs directly into transformer weights without training.physicsrob——

Interpretation history

Decision trace