I’ve been working on a research repository called SALOMI focused on extreme low-bit transformer quantization and inference.
The main thing I’d emphasize is that the most useful result was not a clean “1-bit works” claim. The stronger takeaway was that correlation-based reconstruction metrics can look promising while end-to-end perplexity still collapses, and that strict bits-per-parameter accounting changes a lot of early sub-1-bit conclusions.
The repo includes:
quantization and runtime code
tests and evaluation tooling
backend/runtime experiments
historical drafts and corrected findings
I tried to preserve both the promising directions and the places where the original story did not hold up under stricter evaluation.
Would especially appreciate feedback from people who have worked on quantization, inference systems, or evaluation methodology.
I’ve been working on a research repository called SALOMI focused on extreme low-bit transformer quantization and inference.
The main thing I’d emphasize is that the most useful result was not a clean “1-bit works” claim. The stronger takeaway was that correlation-based reconstruction metrics can look promising while end-to-end perplexity still collapses, and that strict bits-per-parameter accounting changes a lot of early sub-1-bit conclusions.
The repo includes:
quantization and runtime code tests and evaluation tooling backend/runtime experiments historical drafts and corrected findings
I tried to preserve both the promising directions and the places where the original story did not hold up under stricter evaluation.
Would especially appreciate feedback from people who have worked on quantization, inference systems, or evaluation methodology.
Repo: https://github.com/OrionsLock/SALOMI
Was this shoveled out with Claude/Codex to try to ride off the Bonsai release?