UC Berkeley
academicBerkeley's frontier-AI footprint runs through BAIR, Ion Stoica's Sky Computing Lab (with its NovaSky and Agentica teams), and Dawn Song's RDI. No other university has shipped more of the industry's shared plumbing: vLLM and SGLang began here, Chatbot Arena (with Stanford, as LMSYS) became the default human-preference eval before incorporating as LMArena, and the Function-Calling Leaderboard and LiveCodeBench are cited in essentially every 2025-26 model card.
The 2025-26 signature is open RL for reasoning and agents: Sky-T1-32B (o1-class for <$450), the Agentica line (DeepScaleR → DeepCoder → DeepSWE, with Together AI), and the SkyRL library; Song's group counters with the sharpest agentic-safety evals (CyberGym, Agents' Last Exam). Levine and Abbeel's robotics groups (HIL-SERL, Octo, OpenVLA co-lead) anchor the embodied side.
Attribution discipline matters most here: the university keeps original dated releases only — the living vLLM/SGLang projects, LMArena, Physical Intelligence, Nexusflow, and Anyscale/Databricks output belong to their companies.
People
- Ion Stoica Google Scholar — Professor, EECS; Director, Sky Computing Lab (formerly Co-Founder, Databricks, Anyscale; LMArena spun from his lab)
- Sergey Levine Google Scholar — Professor, EECS; RAIL Lab; Co-Founder, Physical Intelligence
- Dawn Song Google Scholar — Professor, EECS; Faculty Director, Berkeley RDI
- Pieter Abbeel Google Scholar — Professor, EECS; BAIR Co-Director; also Amazon Frontier AI & Robotics
- Matei Zaharia Google Scholar — Associate Professor, EECS; Co-Founder & CTO, Databricks