House · Stanford AI Laboratory (SAIL)

In 1962, McCarthy carried the words "artificial intelligence" west from Boston all the way to Palo Alto. From that moment on, SAIL became the most peculiar institution in AI history—at once an academic temple and a neighbor of Silicon Valley. It nursed the patriarchs of expert systems, launched the world-shaking data engineering of ImageNet, and gave countless top researchers their first faculty post—only to watch them open companies a year after they earned tenure.

McCarthy Goes West

In 1962, John McCarthy (1927–2011) transferred from MIT to a position as associate professor of computer science at Stanford University. In 1963, in a wooden building near Felt Lake at the southwest corner of campus, he stood up the Stanford Artificial Intelligence Project—the prototype of what would, in 1965, formally become SAIL (the Stanford Artificial Intelligence Laboratory).

McCarthy brought west the LISP language, the idea of time-sharing, and the program he had carried since Dartmouth: that formal logic could capture commonsense reasoning. In its early years SAIL relocated off-campus to a two-story building atop a hill, the D.C. Power Building (later "the AI Lab"), where dozens of researchers, graduate students, and hackers crammed in together. Strange hand-drawn symbols covered the walls; a modified small vehicle sat permanently in the parking lot—the soon-to-be-famous Stanford Cart.

The D.C. Power Building era trained an entire generation that shaped the field through the 1970s: Raj Reddy went on to found CMU's Robotics Institute; Terry Winograd (1946–) stayed to teach and became the doctoral advisor of Brin and Page; Hans Moravec became a banner of cognitive robotics. McCarthy himself held tenure at Stanford until his death in 2011—the man who, in the truest sense, "planted AI as a discipline on the West Coast."

A common confusion deserves untangling: Shakey the Robot was not SAIL's work—it belonged to the Stanford Research Institute (SRI International, spun off as independent from Stanford in 1970), led by Charles Rosen, Nils Nilsson (1933–2019), and others. The robot that truly belonged to SAIL was the Stanford Cart, started by James Adams in the 1960s and led in the 1970s by Hans Moravec. In 1979 Moravec had the Cart cross an outdoor courtyard strewn with chairs on its own; each one-meter step required ten to fifteen minutes of "thinking," and the entire 30-meter trip took about five hours. This was the world's first vision-based outdoor autonomous navigation. Moravec carried the approach to CMU, where it became one of the foundations of modern mobile robotics.

The Birthplace of Expert Systems

From 1965, SAIL and Stanford's School of Medicine joined to launch one of the most consequential side-streams in AI history—expert systems.

The protagonist was Edward Feigenbaum (1936–). Trained under Simon at CMU, he joined Stanford in 1965. In DENDRAL, in collaboration with Nobel chemistry laureate Joshua Lederberg, Feigenbaum encoded the heuristic knowledge chemists used to infer molecular structure from mass-spectrometer readings into an explicit rule base, letting a program take over the inference. It was the first true production-grade expert system anywhere in the world, and the birthplace of the term "knowledge engineering." In the 1970s, Feigenbaum and Edward Shortliffe and others developed MYCIN, an expert system that recommended antibiotic treatments to physicians. In blind trials, MYCIN's prescriptions often outperformed those of junior doctors. Liability, ethics, and integration costs kept it out of clinical deployment, but it trained a generation of expert-systems researchers and directly inspired both the Japanese Fifth Generation Computer Project of the 1980s and the global expert-systems industry.

In 1994, Feigenbaum and Raj Reddy shared the Turing Award; the citation explicitly named DENDRAL and MYCIN.

The Stanford Heuristic Programming Project (HPP) of the 1970s incubated more than DENDRAL and MYCIN—there was also PROSPECTOR (for mineral exploration; in 1980 it helped identify a molybdenum deposit in Washington State) and TEIRESIAS (a tool for explaining and maintaining expert systems). HPP was the first laboratory in the world that organized "AI in service of a specific professional domain" as systems engineering.

Winograd, Nilsson, and the Symbolic-AI Pantheon

In 1973, Terry Winograd (1946–) moved from MIT to a professorship at Stanford. His doctoral thesis, SHRDLU, was one of the peaks of symbolic AI; but at Stanford his attention shifted toward human-computer interaction and the social foundations of computation. His 1986 book Understanding Computers and Cognition, co-authored with Fernando Flores, offered a deep critique of the purely symbolic line and influenced an entire generation of HCI researchers. It also indirectly shaped the user-centered product philosophy that his students Sergey Brin (1973–) and Larry Page (1973–) would carry into Google.

Nils Nilsson moved from SRI to direct SAIL (1985–1990) and wrote the influential Artificial Intelligence: A New Synthesis (1998) and The Quest for Artificial Intelligence (2009). Stuart Russell (1962–), although primarily at Berkeley, kept close ties with Stanford; the textbook Artificial Intelligence: A Modern Approach, which he co-authored with Peter Norvig (briefly an associate researcher at SAIL in the 1990s, later director of research at Google), first appeared in 1995 and is now the global standard for AI teaching, with a fourth edition in 2020.

Stanley, ImageNet, and a Methodological Turn

In the new century, SAIL's center of gravity shifted, step by step, from symbolic reasoning toward statistical learning and large-scale data.

On October 8, 2005, the second DARPA Grand Challenge unfolded in the Nevada desert. Sebastian Thrun, then SAIL's director, led the Stanford Racing Team in fielding a modified Volkswagen Touareg they named Stanley, equipped with five lidar units, a color camera, GPS, and an IMU; it classified terrain ahead of it on the fly using machine learning. Stanley finished the 212-kilometer course in 6 hours 53 minutes and took the 2-million-dollar prize. It was a watershed moment for self-driving—proof that a learning-driven autonomous system could run reliably in the open. Thrun was soon recruited by Google to lead the Google Self-Driving Car project (later Waymo) and co-founded Udacity in 2011.

In 2007, Fei-Fei Li (1976–) moved from Princeton to a Stanford assistant professorship and entered SAIL. She launched a project that many peers told her not to attempt—ImageNet: a dataset of more than ten million images, organized along the semantic hierarchy of WordNet. With students Yangqing Jia, Jia Deng, and others, she hired tens of thousands of annotators worldwide via Amazon Mechanical Turk. The first ImageNet dataset was released in 2009; the first ImageNet Large Scale Visual Recognition Challenge (ILSVRC) was held in 2010. In September 2012, Geoffrey Hinton (1947–), Alex Krizhevsky (1986–), and Ilya Sutskever (1986–) of the University of Toronto submitted AlexNet to ILSVRC-2012, dropping the top-5 error rate from 26.2% to 15.3% in a single stroke. That night is widely seen as the ignition of the deep-learning revolution—and the powder keg was the ImageNet that Li had been preparing at SAIL for five years.

A Modern Stronghold of Machine Learning and NLP

In 2002, Andrew Ng (1976–), fresh out of Berkeley, joined Stanford's faculty. At SAIL he led work on machine learning and robotics. From 2010 he co-founded Google Brain (with Jeff Dean and Greg Corrado), where in 2012 a deep network trained on 16,000 CPU cores learned to spontaneously recognize cats from unlabeled YouTube videos. In the autumn of 2011, Ng put his Stanford CS229 Machine Learning course online; one hundred thousand students enrolled. That was the true beginning of the MOOC era, and the seed of his co-founding Coursera with Daphne Koller (Stanford's master of probabilistic graphical models) in 2012.

In natural language processing, Christopher D. Manning has taught at Stanford since 1999. The NLP textbook he co-authored with Dan Jurafsky became the field's introductory standard. His Stanford NLP Group produced GloVe word vectors (2014, with Jeffrey Pennington and Richard Socher), the Stanford Parser, CoreNLP, and, after 2018, large-scale language model and foundation-model research. From 2017 on, Manning's PhD students and postdocs have entered the front line of contemporary NLP—including early collaborators of Alec Radford (1991–), scientists at Hugging Face, and many OpenAI researchers.

Percy Liang joined Stanford as an assistant professor in 2012, focusing on question answering, interpretability, semantic parsing, and the evaluation of foundation models. In 2021, Liang, with Manning, Li, and Chris Ré, helped found the Center for Research on Foundation Models (CRFM) and authored the landmark report On the Opportunities and Risks of Foundation Models—formally introducing the term "foundation model." CRFM's HELM (Holistic Evaluation of Language Models) benchmark remains one of the most-cited yardsticks for comparing open-source large models.

Stanford is also a stronghold of reinforcement learning and robotics. Andrew Ng's earlier doctoral students Pieter Abbeel (later Berkeley) and Adam Coates, together with later faculty Chelsea Finn (meta-learning and robot learning), Dorsa Sadigh (human-robot interaction), and Jeannette Bohg (robot perception), have collectively carried SAIL into the era of "embodied intelligence."

HAI and the Two Faces of Silicon Valley Adjacency

In March 2019, Fei-Fei Li and the philosopher John Etchemendy launched the Stanford Institute for Human-Centered Artificial Intelligence (HAI). HAI is Stanford's attempt to formally integrate AI with the humanities, social sciences, medicine, and policy under one institutional roof. Its annual AI Index Report has become the most-cited dataset on the global state of AI.

Stanford's adjacency to Silicon Valley is what most distinguishes SAIL from other AI strongholds. In 1998 Brin and Page walked out of their PhDs and into Google with the Backrub algorithm; in 2009 Thrun walked into Google with Stanley; in 2012 Ng launched Coursera and Google Brain in parallel; around 2015, OpenAI co-founders Greg Brockman and Sam Altman (1985–) were both deeply tied into the Stanford ecosystem; by 2023, Aidan Gomez's Cohere, Mira Murati's later Thinking Machines, and nearly every Andreessen Horowitz–backed AI startup could find a node in SAIL's faculty-and-student lineage. This atmosphere—"finish the paper in the morning, register the company in the afternoon"—has made SAIL a peculiar hybrid: at once a top academic institution and a talent feeder for Silicon Valley.

For the same reason, SAIL was the first to feel the pull of "professors leaving for industry." In the late 2010s, multiple Stanford luminaries took partial or full appointments at Google, Meta, NVIDIA, Apple, and OpenAI—an extension of Stanford's DNA, but also a microcosm of the imbalance between AI academia and industry today.

The Lab as an Institution

Returning to first principles.

SAIL was the seed of West Coast AI. Before its founding, AI's center was Massachusetts; afterward, the entire research network became bipolar. McCarthy's LISP, Feigenbaum's expert systems, and Nilsson's search algorithms laid down the western branch of symbolic AI.

SAIL was the first prime mover of the big-data era. When most labs in the 2000s were tuning cleverer algorithms, Li and her colleagues bet five years on stacking ImageNet—and that single judgment gave the deep-learning revolution its fuel.

SAIL was the shortest bridge between AI and industry. From Google to Coursera to Waymo to OpenAI, Sand Hill Road, 800 meters from campus, is the densest single mile of AI venture capital on Earth. That short road has been both the engine of Stanford AI's continued flourishing and the cause of its continued exodus.

SAIL was the largest output port of AI education. From Ng's CS229 to Karpathy's CS231n to Manning's CS224N, Stanford's core AI courses are all online and have, year after year, fed researchers around the world for free.

Historian's Note

Stanford is to artificial intelligence what the marshes of Yunmeng were to Chu, or what Chang'an was to the Tang. It is not the birthplace of AI—that honor belongs to Dartmouth and MIT; nor is it AI's only temple, with CMU, Berkeley, Toronto, and DeepMind each holding their own peaks. But it has a sense of place that no other institution can quite replicate: to its east, academia; to its west, Silicon Valley; to its south, the medical school; to its north, the liberal arts. McCarthy heading west carried a way of thinking from a coast to a richer soil; Feigenbaum and Lederberg's collaboration walked AI for the first time into the hospital; Li's ImageNet made data the algorithmic fuel of the new century; Ng's open courses pushed AI education to the world; HAI and CRFM pulled AI back into the orbit of the humanities and public policy. SAIL's signature is not a single technical peak but a continual capacity to reorganize itself: every decade, it finds a new way to make "artificial intelligence" speak to the times.

Eyewitness Accounts

Call for contributions

If you have worked or studied at the Stanford AI Laboratory, HAI, CRFM, or the Stanford NLP Group, please contribute on GitHub.

References

1. McCarthy, J. (1963). "A Basis for a Mathematical Theory of Computation." In Computer Programming and Formal Systems, North-Holland, 33–70.
2. Buchanan, B. G., & Feigenbaum, E. A. (1978). "DENDRAL and Meta-DENDRAL: Their Applications Dimension." Artificial Intelligence, 11(1–2): 5–24.
3. Shortliffe, E. H. (1976). Computer-Based Medical Consultations: MYCIN. Elsevier.
4. Moravec, H. (1983). "The Stanford Cart and the CMU Rover." Proceedings of the IEEE, 71(7): 872–884.
5. Winograd, T., & Flores, F. (1986). Understanding Computers and Cognition: A New Foundation for Design. Ablex.
6. Russell, S. J., & Norvig, P. (1995, 2020). Artificial Intelligence: A Modern Approach (1st & 4th eds.). Prentice Hall / Pearson.
7. Thrun, S., et al. (2006). "Stanley: The Robot That Won the DARPA Grand Challenge." Journal of Field Robotics, 23(9): 661–692.
8. Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., & Fei-Fei, L. (2009). "ImageNet: A Large-Scale Hierarchical Image Database." Proceedings of CVPR 2009.
9. Russakovsky, O., Deng, J., et al. (2015). "ImageNet Large Scale Visual Recognition Challenge." International Journal of Computer Vision, 115(3): 211–252.
10. Bommasani, R., Hudson, D. A., Adeli, E., Altman, R., Arora, S., et al. (2021). "On the Opportunities and Risks of Foundation Models." Stanford CRFM, arXiv:2108.07258.
11. Stanford HAI (2018–2025). AI Index Annual Report.
12. Nilsson, N. J. (2009). The Quest for Artificial Intelligence: A History of Ideas and Achievements. Cambridge University Press.