Towards a hardware-invariant NVFP4 for GPUs without FP4: progress report 1
4-bit weights on the INT8 tensor cores every NVIDIA GPU from Ampere to Blackwell has, how the format got here, and where the evaluation is right now.
Read →~/whoami
secemp.
Reverse engineering, ML, and systems notes from an independent researcher.
- now
- Agent tooling, model behavior, and reverse engineering workflows.
- focus
- llm evaluation · agent harnesses · systems
- elsewhere
- github huggingface x
Latest
Writing
- Why Writing Feels Cringe
- Anthropic Lost to the Elwood
- Anthropic Doesn't Know How to Yield
- Being Wrong Faster
- On Emergent Misalignment from Reward Hacking
Projects
- elwood Drop-in replacement for the Claude Code SDK via Babel AST instrumentation. tool
- harn Full Python port of earendil’s pi agent. library
- rubrics A collection of rubrics for LLM judge use cases. data
- niwa Async conflict-aware spec and planning for users and their agents. tool
- jari LMDB-backed task and issue tracker for agent workflows. tool
- goal Port of /goal from Codex to arbitrary agents and harnesses. tool