Foundation models, alignment, and inference.
Programs span efficient pre-training mixtures, long-context reasoning, tool-use alignment, and routers that allocate spend across model families. Every project ships with evaluation cards suitable for model risk discussions.
Training clusters on Distributed Cloud, shared tokenization and data contracts, red-team prompt libraries, and publication tooling that turns runs into papers and product model packs.
Models that clear quality, safety, and cost gates enter Orchestrator channels. Failures become public or internal postmortems so the lab compounds learning rather than hiding negative results.
AI Systems Lab shares compute, evaluation, and publication tooling with every other Zansoc lab — so discoveries compound across the stack.
Visit all labs