In 2011, Andrew Ng did something quietly radical for a Stanford faculty member: he helped wire together 16,000 processor cores inside Google and set them loose on YouTube videos in the name of artificial intelligence (AI). The system learned to recognise cats without being told what a cat was. That project, Google Brain, reset expectations about how much data and compute a neural network could actually use.
The AI juggernaut label gets applied to many people; in Ng's case the components are unusually concrete. He has been a Stanford professor, a founding mind behind Google Brain, chief scientist at Baidu, co-founder of Coursera, and the entrepreneur behind DeepLearning.AI, AI Fund and Landing AI. Coursera alone crossed 100 million registered learners in 2021, a scale that turned Ng's Stanford machine-learning (ML) course into a global training pipeline.
A friend who runs a small industrial automation company — call him Marlon — spent six months tuning a visual inspection model, then watched it fail after the factory installed brighter lighting panels. The mathematics had not broken. The data had drifted. That distinction sits at the centre of Ng's second act.
The Stanford classroom was an early incubator
Ng was born in London in 1976 and grew up in Hong Kong and Singapore before moving to the United States. He studied computer science and economics at Carnegie Mellon, then went on to a master's degree at MIT and a PhD at the University of California, Berkeley. By the early 2000s he was teaching at Stanford, where his machine-learning course became a destination for students from across engineering and business programmes.
Stanford's early experiment with online delivery proved how wide the demand really was. When the university put his course on the open web, more than 100,000 people signed up within weeks, a number that looked nonsensical for a single class at the time. Ng treated that moment as evidence that the bottleneck was not interest. It was access.
The pattern was set. Ng kept publishing research, but he also kept packaging the research into formats that practitioners could use without leaving their jobs.
Google Brain and the case for bigger networks
At Google, Ng worked with Jeff Dean and others on what became Google Brain, an attempt to train neural networks on far larger datasets than most research groups could manage. The cat-recognition result used unlabelled YouTube images and a distributed computing setup that was exotic in its day. The system did not receive a cat label and then look for cats. It learned a representation of what a cat was on its own.
The specific result was less important than the demonstration. It showed that given enough data and compute, a system could learn features that transferred to other tasks. That point still drives much of the field's move toward large models.
Google Brain also gave Ng a view of how different industrial machine learning was from a university project. Data arrived in messy streams, models had to run on real infrastructure, and training cycles cost real money.
Baidu and the line that stuck
In 2014, Ng became chief scientist at Baidu, where he ran the company's AI Group across large-scale speech recognition and the company's autonomous driving work. He repeated the Google Brain playbook at a different scale and in a different language environment. Under his leadership, Baidu's AI group built speech systems that pushed Mandarin voice recognition from a lab demonstration toward a consumer feature.
Around that period, Ng popularised a comparison: “artificial intelligence is the new electricity.” The phrase was not a technical claim. It was an infrastructure claim, and it explained why he kept moving from one organisation to the next rather than settling into a single research agenda.
Coursera and the distribution play
Ng co-founded Coursera with Daphne Koller in 2012, turning the Stanford machine-learning course into one of the web's first mass open courses. The timing helped. Broadband and mobile access were mature enough that lectures could travel, and employers were beginning to ask for skills that universities were slow to deliver at scale.
Coursera crossed 100 million registered learners in 2021 and went public on the New York Stock Exchange the same year. Coursera's public reporting shows the platform has since expanded beyond individual courses into degrees and workforce training. Ng's original machine-learning course still sits near the top of the company's most-enrolled lists.
The founder stack
After leaving Baidu in 2017, Ng built a cluster of interlocking ventures and a publishing habit. The list reads like a full-stack bet on AI adoption:
- DeepLearning.AI runs courses and a weekly newsletter for practitioners.
- AI Fund is a venture studio that backs early-stage companies with technical teams and capital.
- Landing AI focuses on computer vision for factories, selling a platform called LandingLens for visual inspection.
- Machine Learning Yearning is Ng's free book on how to structure machine-learning projects.
These companies share a habit that is less obvious than the technology. Each one treats machine-learning models as an operational system, not just an algorithm. Data quality, retraining schedules, domain experts, and user feedback are part of the product from the first week. For manufacturers, the Landing AI pitch is narrower than general-purpose generative AI: defect detection in machine parts, circuit boards, and finished goods.
Data-centric AI and the drift problem
In recent years Ng has argued for “data-centric AI”: the practice of improving the consistency of labels and the examples used for evaluation before reaching for a larger model. Marlon's lighting failure was a data-centric problem. The model had learned to associate bright pixels with defects because the original training images all came from one shift's overhead lights.
In several public contests, Ng's teams have told participants to keep the model fixed and compete solely on improving the dataset. The winning entries consistently lifted accuracy by relabelling edge cases, removing duplicate examples, and balancing rare classes. The lesson was blunt: data work is not a cleanup task. It is the product.
The base rate matters here. Most machine-learning pilot projects never become durable infrastructure, and the common industry estimates put full production conversion well below half. Ng's exception is that he repeatedly builds distribution around the model, so the model is not the whole product.
What this means for the team you run
The Ng playbook points to a sequence:
- Start with a consumer: a person or system that needs the model's output, not a vague promise of intelligence.
- Design the data pipeline earlier than the model, because data drift is a first-year problem, not an edge case.
- Publish internal documentation that records when a model is allowed to be trusted and when it is not.
- Treat retraining as a continuing cost centre, with scheduled reviews at intervals measured in weeks or months, not years.
For Marlon's team, the fix was not a better model. It was a camera-position standard, a relabelled image set, a one-page README that described the lighting conditions, and a check before every production run.
The most useful next step is smaller than it sounds. Take one model you are proud of and write down the exact conditions under which it should be switched off. If you cannot name that failure mode, the model is not ready for the factory floor, the customer, the regulator, or the public.
