Allen Goodman
@allen.bsky.social
Science and programming but prefer skateboarding—Genentech, Facebook AI Research (FAIR), Broad Institute, …
Jordon should coach UNC tbh
October 21, 2025 at 2:26 PM
Jordon should coach UNC tbh
Just to be clear, we shouldn’t use these embeddings for compilation. When possible, we should ask for more determinism, not less. But using the embedding as an objective function, e.g., transforming into SSA—is a practical target.
August 31, 2025 at 3:21 PM
Just to be clear, we shouldn’t use these embeddings for compilation. When possible, we should ask for more determinism, not less. But using the embedding as an objective function, e.g., transforming into SSA—is a practical target.
Now I feel the same way about LLMs for programming. Instead of treating them as text predictors, we should be projecting all program representations into a shared embedding space. Syntax, semantics, types, elaborations—the entire front-end pipeline collapsed into one embedding.
August 31, 2025 at 3:17 PM
Now I feel the same way about LLMs for programming. Instead of treating them as text predictors, we should be projecting all program representations into a shared embedding space. Syntax, semantics, types, elaborations—the entire front-end pipeline collapsed into one embedding.
He passed, but I still think about that idea. Most programmers reason about programs semantically, not syntactically—yet GitHub search traps us at the level of surface syntax. It’s disappointing that our interface for navigating code is so shallow when the real structure is right there.
August 31, 2025 at 3:15 PM
He passed, but I still think about that idea. Most programmers reason about programs semantically, not syntactically—yet GitHub search traps us at the level of surface syntax. It’s disappointing that our interface for navigating code is so shallow when the real structure is right there.