Andrew Ng and AI Education

Abstract

Andrew Ng co-founded Google Brain, ran AI research at Baidu, and co-founded Coursera. His largest effect on computing came through teaching: when deep-learning knowledge sat in a handful of labs and was worth a great deal of money, he put it online for anyone who wanted it. His Stanford machine-learning course drew more than 100,000 students in 2011, and his line “AI is the new electricity” became the standard business pitch for AI. He is an outspoken optimist about AI, usually right about where it was heading and early about when it would arrive.

Andrew Ng at TechCrunch Disrupt SF, 2017. Image: TechCrunch, CC BY 2.0, via Wikimedia Commons.

London, Hong Kong, Singapore, Carnegie Mellon

Andrew Ng was born on April 18, 1976, in London, to Chinese parents from Hong Kong. The family moved through Hong Kong and Singapore before Ng arrived in the United States for university. He completed his undergraduate degree at Carnegie Mellon in 1997 with a triple major in computer science, statistics, and economics, took a master’s at MIT in 1998, and went to UC Berkeley for his doctorate, working under Michael Jordan, one of the most influential statisticians and machine learning researchers of the era, completing his PhD in 2002.

His dissertation, Shaping and Policy Search in Reinforcement Learning, applied reinforcement learning to physical control, including an autonomous model helicopter that learned to hover. At Stanford his group, with Pieter Abbeel, Adam Coates, and Morgan Quigley, later taught model helicopters aerobatic maneuvers that nobody had written explicit rules for, and the helicopter videos drew wide attention.

He joined Stanford as an assistant professor in 2002, became director of the Stanford Artificial Intelligence Laboratory (SAIL), and established a research program in machine learning, robotics, and computer vision that was applied, prolific, and unusually well-connected to industry. Stanford’s position in Silicon Valley meant that his students moved into companies that were just beginning to build AI infrastructure, carrying the research practices of his lab with them.

Google Brain: What Happens When You Scale Up

In 2011, Ng co-founded Google Brain with Jeff Dean and Greg Corrado. The project began inside Google’s semi-independent innovation lab with one empirical question: what happens when you train neural networks on far more compute and data than anyone had used before?

The question was motivated by a hypothesis (that scale itself might produce qualitatively new capabilities) that was not yet widely accepted in the research community. Most AI researchers in 2011 believed that algorithmic improvements drove progress, and that simply training larger networks would produce larger versions of the same capabilities. Ng’s bet was that this might not be true, and that scale itself would buy capability.

The most famous early result was the “cat paper”: Quoc Le, Ng, and their colleagues trained a nine-layer network with about one billion connections on ten million 200×200-pixel frames sampled from YouTube videos, using 1,000 machines in parallel and no labels. The network, trained to reconstruct its inputs, spontaneously developed a neuron that responded strongly to cat faces, a concept that had never been explicitly labeled in the training data. The 2012 paper, “Building High-level Features Using Large Scale Unsupervised Learning,” drew wide press coverage, because the image of a neural network “inventing” the concept of a cat matched popular intuitions about what artificial intelligence should look like.

Info

The cat paper’s significance was less about cats than about the scale itself. Training the network took 1,000 machines with 16,000 CPU cores for three days at Google, resources unavailable to academic researchers and most industrial labs. The result suggested that certain capabilities might only emerge above certain scales of computation and data. This idea, later called the scaling hypothesis, became central to AI development in the following decade. Ng’s Google Brain project was an early, influential experiment in what that hypothesis implied.

Ng left Google in 2012 to run Coursera. In 2014 he joined Baidu as Chief Scientist, working from a Silicon Valley office and running AI research for China’s dominant search company, a sign of the weight Chinese technology companies were beginning to put on AI.

At Baidu, Ng’s team worked on deep learning for speech recognition, natural language processing, and autonomous driving. By the time he announced his resignation in March 2017, Baidu’s AI Group had about 1,300 people. His farewell post gave no reason beyond wanting to bring AI to industries outside the technology sector; press reports speculated about a reorganization under Baidu’s new president, Lu Qi.

The Stanford Course and the MOOC Revolution

In October 2011, the same year he co-founded Google Brain, Ng offered his applied Stanford machine learning course free online at ml-class.org. It was not the first online course, but it showed how large the demand was: over 100,000 students enrolled, many from countries where a Stanford education was out of reach. Students completed problem sets, sat through hours of lectures, and wrote to Ng about how the course had changed their careers.

Ng drew a conclusion from it: the bottleneck in AI adoption was knowledge, not technology. Methods developed in a few universities and industrial labs were useful only if people in other industries knew how to apply them, and traditional education could not train them fast enough.

In 2012, Ng and Stanford colleague Daphne Koller co-founded Coursera, the massive open online course platform. Coursera launched with courses from Stanford, Princeton, Michigan, and Penn, and grew to partner with hundreds of universities and corporations worldwide. In April 2022 it passed 100 million registered learners.

Info

Between 2013 and 2020, machine-learning engineers were among the best-paid people in software, and universities graduated far fewer of them than industry wanted to hire. Coursera’s ML and deep learning courses did not close that gap, but they widened access to the knowledge behind it.

His original Machine Learning course on Coursera had taken in more than 4.8 million learners by 2022, when it was replaced by an expanded Machine Learning Specialization. When he launched the Deep Learning Specialization in 2017, covering neural networks, convolutional networks, sequence models, and practical deployment through deeplearning.ai, the courses reached practitioners in places that had no access to university AI programs. Learners outside the big AI centers who completed these courses entered a job market that could not fill its AI positions.

Access was not the same as equal benefit. Organizations that could afford to hire the newly trained ML engineers gained more than those that could not.

“AI is the New Electricity”

The phrase Ng used in a 2017 Stanford GSB lecture (“AI is the new electricity”) became the dominant metaphor for AI’s economic significance in the subsequent decade. The comparison had a point: electricity in the early twentieth century was a general-purpose input that changed manufacturing, transportation, and agriculture, and the slow part was getting each industry to reorganize around it.

Ng’s argument was that AI was the same kind of general-purpose technology. The question was not whether AI would transform healthcare, logistics, retail, and manufacturing (it clearly could) but how quickly those industries would build the organizational capability to adopt it. His subsequent companies, Landing AI and AI Fund, were both organized around this premise: the technical barriers to AI adoption were lower than the organizational and deployment barriers.

Landing AI worked with industrial companies, mostly manufacturers, on computer vision for visual inspection. Ng’s team found that deploying AI on a factory floor meant managing the algorithm plus the data pipeline, the edge case handling, the human-machine interface, and the organizational change management. These were not primarily technical problems, but they were the problems that determined whether AI created value or remained a proof of concept.

AI Fund, his venture studio, built AI-native companies in verticals where Ng identified an opportunity. Unlike a conventional VC fund, AI Fund employed teams to actively build companies rather than funding external founders. The model reflected his view that the barriers to AI deployment were organizational and required hands-on institution building as well as capital.

The Safety Debate: Optimism as Position

In the debate about AI risk that grew after 2022, Ng held that existential-risk concerns were overblown and distracted from nearer harms that were easier to address.

At Nvidia’s GPU Technology Conference in March 2015 he compared worrying about evil killer robots to worrying about overpopulation on Mars: a coherent concern in the abstract, but not one that bore on decisions to be made today. Job displacement, bias in high-stakes decisions, concentration of market power, and misuse for surveillance were happening now, had known remedies, and in his view deserved the attention.

He also opposed regulation that would slow AI development: in December 2023 he warned that licensing and liability rules would burden small firms and open-source models, and in 2024 he attacked California’s SB 1047 for creating liabilities for “science-fiction risks,” arguing that the benefits of faster deployment in medicine, agriculture, and education were measurable and large, while the risks such rules addressed were speculative.

Warning

Ng’s forecasts have tended to be right about direction and early about timing. In November 2017, after his Stanford group’s CheXNet model was reported to read pneumonia on chest X-rays better than radiologists, he asked publicly whether radiologists should be worried about their jobs; critics pointed out that CheXNet had been compared with only four radiologists, and demand for radiologists has kept rising since. (The bluntest prediction, “stop training radiologists,” was Geoffrey Hinton’s, in 2016.) At Baidu he promoted autonomous vehicles on short timelines; they arrived, but more narrowly and more slowly. In each case the technology worked sooner than institutions adopted it.

His disagreements with former colleagues (Hinton citing AI risks as reason for leaving Google, Bengio becoming an AI safety advocate, even LeCun arguing that LLMs are fundamentally limited) have not significantly changed his public position. He argues that the correct response to AI’s power is to distribute it more widely, not to slow its development, and that the people most alarmed about AI risk are, in some cases, those most positioned to benefit from restricting access to it.

That last argument drew criticism. Critics note that his case for spreading AI knowledge does not by itself rule out taking AI risk seriously.

The Educator

Ng did not invent the transformer or backpropagation. What he built was distribution: Coursera, deeplearning.ai, The Batch newsletter, and years of free lectures and course materials, at a time when deep-learning knowledge was concentrated in a few labs.

How many engineers learned machine learning from his courses who would not have learned it otherwise, and how much AI adoption outside the technology industry they brought about, cannot be measured directly.