Frontier AI lab building models for human collaboration and organizational coordination. The technical focus is on long-horizon and multi-agent reinforcement learning, persistent memory, and user understanding, with training approaches aimed at making models actively solicit and retain context about users. Founded by researchers who post-trained Claude and built Grok.
Ampersand's multi-agent 'Co-Coding' mode hits 92.2% on SWE-bench Pro HARD-51 without any coding-specific training, matching frontier models at 70% lower cost.
humans& trained a tutor model against a simulated learner called Persimmon, topping a real-classroom benchmark and beating Claude Sonnet 5.5 and GPT-6 Astra.
Humans& rebuilt the NVFP4 decode kernel on Blackwell by keeping weights at 4 bits while promoting activations to BF16, gaining both speed and accuracy.
humans& open-sources a stable NVFP4 reinforcement learning recipe that matches BF16 training dynamics while delivering major throughput gains on Blackwell GPUs