Humanfia
Articles about Humanfia
Flow Science: four findings, and engineering or a science?
What the flow-level ablations found that we did not expect: higher diversity, higher score; better collaboration, better performance; a progressive goal keeps agents on track; and Claude degrades across turns where GPT does not. Then: is a flow just engineering?
Flow Science: different tasks need different flows
Flow-level ablations on MLE-bench, HLE, ProgramBench and Anthropic's performance take-home, with the models held fixed. The best flow differs from one benchmark to the next.
Flow Science: flow is the new scaling dimension
Models scale with parameters and data, and agents scale with tools. The next dimension is the flow. Why a single /goal session is not enough on long, hard tasks, and why one flow cannot solve every problem.
KDA²: Kernel Design Agents (KDA) optimize Kimi Delta Attention (KDA)
Our agents wrote Kimi Delta Attention kernels that run up to 2.96× faster than FlashKDA on B300 with a tenth of its state error. Here is how, and how the agents tried to cheat along the way.