House · UC Berkeley / BAIR (Berkeley AI Research)

The University of California, Berkeley is to AI what a lighthouse is to the western shore of the Bay—from fuzzy logic to Bayesian networks, from computer vision to deep reinforcement learning, almost every paradigm shift carries its silhouette.

The Birthplace of Fuzzy Logic

Berkeley's bond with artificial intelligence began with one paradigm-defining paper. In 1965, Iranian-American scholar Lotfi A. Zadeh, in Berkeley's Department of Electrical Engineering, published Fuzzy Sets—the formal mathematical framework of "fuzzy sets." It was not "machine learning" in the later sense, but it was one of the most influential non-classical reasoning tools in early AI. Behind Japan's home appliances, subway controls, and camera autofocus systems of the 1980s, this Berkeley mathematics quietly did the work. Zadeh taught at Berkeley for nearly half a century and trained a generation of logicians and soft-computing scholars.

Berkeley was not, like MIT or Stanford, born at the moment "artificial intelligence" was named. But at every paradigm shift, it kept its footing. That steadiness owed much to Berkeley's Department of Electrical Engineering and Computer Sciences (EECS), with its consistent academic temperament: heavy on mathematics, heavy on systems, slow to chase fashion.

AIMA and Stuart Russell

In 1986, Stuart Russell (1962–) moved from Oxford to Berkeley. The book he co-authored with Peter Norvig (later director of research at Google), Artificial Intelligence: A Modern Approach (AIMA), first appeared in 1995. Adopted as a textbook by more than 1,500 universities and revised through four editions (the fourth in 2020), AIMA is the most-cited, most-translated, most-read AI textbook of the past three decades.

AIMA's framing is "rational agents." Rather than taking a side between symbolic AI and connectionism, it places search, logic, probability, learning, and decision-making in a single coordinate system, teaching students from the start to look at AI through problems. That style is almost a footnote to the Berkeley school.

Russell himself later turned to AI safety. In 2015 he co-signed, with Geoffrey Hinton (1947–) and others, an open letter on the future of AI; in 2019 he published Human Compatible, articulating the long-term agenda of "provably beneficial AI." During the ChatGPT wave of 2023, he became one of the earliest senior scholars to publicly call for regulation.

The Jordan School and Probabilistic Machine Learning

If Russell laid down Berkeley's "rational" base, then Michael I. Jordan (1956–), who moved from MIT in 1998, turned the campus into the cathedral of probabilistic machine learning.

Jordan, hailed as the "godfather of machine learning," is a scholar trained originally in cognitive psychology who almost single-handedly brought Bayesian networks, variational inference, latent-variable models, and graphical-model inference into the mainstream of machine learning. The roster of his students and postdocs reads like a who's-who of contemporary ML: Andrew Ng (1976–) (Stanford, Coursera, Baidu); David Blei (LDA, Columbia); Zoubin Ghahramani (Cambridge, founding head of Uber AI); Tommi Jaakkola (MIT); Eric Xing (CMU, president of MBZUAI); Yoshua Bengio (1964–) (briefly a postdoc).

Jordan held joint appointments in EECS and Statistics—a detail that mattered, because it showed how, in the 1990s, Berkeley already understood "machine learning" as the meeting place of statistics and computation, not as a sub-branch of AI. That judgment was a full fifteen years ahead of Silicon Valley's "AI renaissance."

It is worth noting that Judea Pearl (1936–), while based primarily at UCLA, did causal-inference and Bayesian-network work that was deeply intertwined with Berkeley's Jordan school. Together they shaped the North American landscape of probabilistic graphical models.

Vision and Robotics—the Malik Line

Berkeley's computer-vision tradition was led by Jitendra Malik. Teaching at Berkeley since 1986, Malik's vision group trained a long line of scholars who would shape the field: Jianbo Shi (Normalized Cuts), Serge Belongie (dean of Cornell Tech), Trevor Darrell (multimodal vision), Alyosha Efros (image synthesis and unsupervised vision), David Forsyth, and Pietro Perona (a frequent collaborator). Malik himself received the IJCAI Award for Research Excellence in 2019 and the ACM/AAAI Allen Newell Award in 2022.

In robotics, Pieter Abbeel arrived from Stanford (as a student of Andrew Ng) in 2008 and built the Robot Learning Lab at Berkeley. His student Sergey Levine joined the faculty in 2016. Together they pushed deep reinforcement learning into robotics—from autonomous helicopter aerobatics (Abbeel's PhD thesis), to end-to-end robotic grasping (Levine et al., 2016), to the 2017 PPO (Proximal Policy Optimization) of Schulman, Wolski, and others—the same algorithm that would later sit at the core of RLHF training behind ChatGPT.

Abbeel's Berkeley-trained students populate the front lines of industry: John Schulman (OpenAI co-founder); Aravind Srinivas (Perplexity co-founder and CEO); Chelsea Finn (Stanford, Physical Intelligence); Igor Mordatch (Google DeepMind). Abbeel himself founded Covariant in 2017 to build foundation models for warehouse robotics; Amazon partially acquired the company in 2024.

BAIR, RISELab, and the Systems Foundation

Around 2014, Berkeley consolidated its AI research—scattered across EECS, Statistics, ICSI, and other units—into Berkeley AI Research (BAIR). BAIR is not a separate legal entity but a cross-departmental research platform, beginning with about 30 faculty and 200 students; by 2025 it had grown to roughly 50 faculty and nearly 600 researchers.

In parallel, Berkeley EECS spun up AMPLab (2011–2016) and its successors RISELab (2017–2021) and the Sky Computing Lab (2022–). Systems-school faculty Ion Stoica, Michael Franklin, and Scott Shenker led the line and built what became the era's underlying infrastructure for AI:

• Apache Spark (from 2009; Matei Zaharia's PhD work at Berkeley), commercialized by Databricks, which Zaharia and Stoica co-founded; valuation surpassed 60 billion dollars in 2025.
• Ray (2017; Robert Nishihara, Philipp Moritz, and others), a distributed-computing framework used by OpenAI to train GPT-3 and GPT-4 and commercialized by Anyscale (also founded by Stoica and others).
• vLLM (2023; Woosuk Kwon, Zhuohan Li, and others at Sky Computing Lab), a high-throughput inference framework based on PagedAttention that within a year became the de facto standard for open-source LLM inference.

Add Dawn Song's continuing work on blockchain and AI security and Stuart Russell's research at the Center for Human-Compatible AI, and Berkeley in the 2020s shows a tripartite shape—algorithms + systems + safety—that distinguishes it from Toronto and Montreal.

Berkeley's Industrial Footprint

Across the Bay Area's AI ecosystem, Berkeley's network reaches almost every leading company:

• OpenAI: Schulman; Andrej Karpathy (a Berkeley summer researcher); Pieter Abbeel (early advisor); Wojciech Zaremba and others have Berkeley backgrounds.
• DeepMind: early collaborators of Sergey Levine; Igor Mordatch; Misha Denil; and others have come out of Berkeley.
• Anthropic: Dario Amodei (1983–) himself did his PhD at Princeton, but several of his core team members come from Berkeley; Tom Brown is a frequent collaborator.
• Tesla / xAI: Andrej Karpathy began with a Berkeley summer lab.
• Databricks, Anyscale, Covariant, Perplexity: all founded directly by Berkeley faculty or students.

Berkeley's influence on Silicon Valley has a different shape from Stanford's. Stanford is the "founders' alma mater"; Berkeley is more like "the conveyor belt for principal engineers and chief scientists."

Berkeley as an Institution

Looked at from a distance, Berkeley has not missed a single AI wave: the fuzzy logic of the 1960s and 70s; the probabilistic reasoning and computer vision of the 1980s and 90s; the statistical machine learning of the 2000s; the deep reinforcement learning and robotics of the 2010s; and the large-model systems and safety work of the 2020s. It does not carry MIT's halo of "the namer of AI"; it does not lean into Silicon Valley capital the way Stanford does. Its stance has been the cooler one—academia plus systems, two axes braided together.

By 2026, BAIR remains one of the most-cited AI research organizations in the world; the Sky Computing Lab leads the open-source large-model inference infrastructure; CHAI (the Center for Human-Compatible AI) drives the academic agenda of AI safety in North America. Berkeley did not build its own ChatGPT, but the PPO behind ChatGPT came from Berkeley, the Ray that trained it came from Berkeley, the vLLM that serves it came from Berkeley, and the Stuart Russell who critiques it also comes from Berkeley.

Historian's Note

Berkeley is to AI what Jixia was to Qi or the White Deer Grotto Academy was to Song—not necessarily the founding monastery of any school, but the place where generation after generation of scholars lecture and argue. From Zadeh's fuzzy sets, to Russell's rational agents, to Jordan's probabilistic graphical models, what Berkeley has given AI is not a slogan but, again and again, languages that can be computed, proved, and engineered. Its virtue is restraint: when symbolic AI was at its zenith, it did not put all its chips on LISP; when deep learning swept the world, it did not abandon probability and causality. In today's large-model era, the Berkeley systems school has quietly written Spark, Ray, and vLLM—three pieces of "water and electricity" infrastructure for trillion-parameter models. The deepest mark of a school is rarely its star papers; it is the way it trains its students. To dare both the most cutting-edge algorithm and the most thankless system: that is the lineage Berkeley has left to AI.

Eyewitness Accounts

Call for contributions

If you have worked or studied at BAIR, RISELab, the Sky Computing Lab, or any of Berkeley's AI groups, please contribute on GitHub.

References

1. Zadeh, L. A. (1965). "Fuzzy Sets." Information and Control, 8(3): 338–353.
2. Russell, S., & Norvig, P. (1995, 4th ed. 2020). Artificial Intelligence: A Modern Approach. Prentice Hall / Pearson.
3. Russell, S. (2019). Human Compatible: Artificial Intelligence and the Problem of Control. Viking.
4. Jordan, M. I. (Ed.) (1998). Learning in Graphical Models. MIT Press.
5. Schulman, J., Wolski, F., Dhariwal, P., Radford, A., & Klimov, O. (2017). "Proximal Policy Optimization Algorithms." arXiv:1707.06347.
6. Levine, S., Finn, C., Darrell, T., & Abbeel, P. (2016). "End-to-End Training of Deep Visuomotor Policies." JMLR, 17(39).
7. Zaharia, M., et al. (2010). "Spark: Cluster Computing with Working Sets." HotCloud '10.
8. Moritz, P., Nishihara, R., et al. (2018). "Ray: A Distributed Framework for Emerging AI Applications." OSDI '18.
9. Kwon, W., Li, Z., et al. (2023). "Efficient Memory Management for Large Language Model Serving with PagedAttention." SOSP '23.
10. Berkeley AI Research (BAIR) Lab. About BAIR. https://bair.berkeley.edu/
11. Center for Human-Compatible AI (CHAI). https://humancompatible.ai/
12. Malik, J. (2022). "ACM/AAAI Allen Newell Award Citation."