Last reviewed
May 11, 2026
Sources
11 citations
Review status
Source-backed
Revision
v2 ยท 2,883 words
Add missing citations, update stale details, or suggest a clearer explanation.
| Andrew Ng: Opportunities in AI - 2023 (Stanford) | |
|---|---|
| Load video YouTube YouTube might collect personal data. Privacy Policy **ContinueDismiss | |
| Information | |
| Name | Andrew Ng: Opportunities in AI - 2023 |
| Type | Technical |
| Event | Stanford Graduate School of Business |
| Organization | Stanford |
| Channel | Stanford Online |
| Presenter | Andrew Ng |
| Description | Dr. Andrew Ng explores AI's broad potential and ethical considerations, emphasizing its impact on various sectors and the importance of responsible development for future innovations. |
| Date | July 26, 2023 |
| Website | https://www.youtube.com/watch?v=5p248yoa3oE |
Andrew Ng: Opportunities in AI - 2023 is a Presentation by Andrew Ng delivered at Stanford University on July 26, 2023. The talk was hosted at the Cemex Auditorium and presented through the Stanford Graduate School of Business, then published on the Stanford Online YouTube channel as part of its lecture archive. It quickly became one of the most-watched AI talks of 2023 and is frequently cited as a practical primer for founders and product leaders entering the field after ChatGPT brought generative AI into the mainstream.
In the talk, Dr. Andrew Ng walks through where value is being created in artificial intelligence in 2023, what the realistic timelines look like for building AI products, and how new tools change the economics of pursuing smaller and more specialized problems. Rather than focusing on demos or model benchmarks, Ng frames AI as a business and engineering question: what can teams build today, how long does it take, and where will the durable value end up sitting in the AI stack?
He describes supervised learning and generative AI as the two main engines driving AI value, presents a four-layer view of the AI stack, and lays out the playbook he uses at AI Fund to take ideas from concept to funded company. He closes with his views on jobs, regulation, and the much-debated topic of human-level intelligence, which he treats with notable skepticism.
Ng holds a long list of overlapping roles in AI. He is the founder of DeepLearning.AI, the managing general partner of AI Fund, the founder of Landing AI, and a co-founder and former chairman of Coursera (where he co-founded the platform with Daphne Koller in 2012). He is an adjunct professor at Stanford. Earlier in his career he founded and led Google Brain and served as chief scientist at Baidu, where he led the company's AI group from 2014 to 2017. His online Machine Learning course on Coursera, first offered in 2011, has been taken by millions of learners and helped popularize the field outside academia.
By 2023 he had spent close to two decades teaching machine learning and roughly five years investing in early-stage AI startups through AI Fund, which by then had grown into a venture studio with hundreds of millions in committed capital and dozens of portfolio companies. That dual perspective, classroom and cap table, shapes most of the talk.
The talk opens with a phrase Ng has used in earlier Stanford lectures: AI is the new electricity. The point is not that AI is as mature as electricity was in the 1920s. It is that AI behaves like a general-purpose technology, the kind of underlying capability that gradually shows up in almost every industry rather than staying confined to one. Electricity changed manufacturing, transportation, communication, and the home over several decades. AI, in his framing, is on a similar trajectory, with healthcare, agriculture, logistics, finance, education, and customer service all on the list of sectors where it shows up in different forms.
That framing matters for how to think about opportunity. If AI were a single product category, the winner-take-all dynamics of consumer software would apply. If it is closer to a horizontal capability, then the interesting question is which specific applications get built on top of it, who builds them, and where the value lands.
Ng organizes the rest of the talk around two technologies he sees as doing most of the work in 2023.
Supervised learning is the older of the two and, in his telling, still the dominant source of economic value. He describes it as learning A-to-B mappings: given an input, produce a labeled output. Spam detection takes an email and labels it spam or not. Online advertising takes an ad and a user and predicts whether the user will click. Self-driving perception takes a camera frame and outputs the position of nearby cars. Automated visual inspection takes a photo of a part on a factory line and decides whether it has a defect. He estimates that this category drives most of the value AI has produced over the past decade, generating in the neighborhood of $100 billion or more in annual value for large technology companies through products such as search ranking and ad targeting.
Generative AI is the newer engine. Ng explains it in the simplest terms: a large language model is trained, with supervised learning techniques on hundreds of billions of words, to predict the next word in a sequence. Stack enough of that together and you get systems that can write, summarize, translate, and answer questions in fluent prose. Products like ChatGPT brought this capability to a wide audience in late 2022 and made it the most-discussed area in software the following year. He predicts that the generative AI category will more than double over the next three years, both as a result of new products being shipped and as a result of venture capital flowing into the space.
The two engines are complementary rather than competing. Supervised learning still wins on most narrow business problems with clear labeled data. Generative AI opens up tasks that previously required a lot of unstructured language work and lets teams build prototypes faster.
The most quoted part of the talk is Ng's comparison of how long it takes to build a working AI application before and after the rise of large language models.
For a traditional supervised learning project, the pipeline is well known and slow. Collect a labeled dataset for the problem. Train and tune a model. Wrap it in an API. Deploy it behind a service. Monitor it. Realistically, even with a competent team, this cycle takes about six to twelve months for a serious commercial deployment.
With prompt-based generative AI, the picture changes dramatically. A developer can write a prompt that calls a hosted language model, wrap it in a simple application, and ship a usable prototype in days. Ng cites scenarios where similar projects that used to take six months now take a week. The cost structure changes with it, because you are renting inference time rather than paying for a custom training run plus a dedicated serving stack.
This compression has knock-on effects on what is worth building. Problems that were previously too small to justify a six-month engineering project become viable when the engineering project is one engineer for a week. That is the foundation of his long-tail argument.
Ng draws a familiar long-tail curve. On the head sit the very large applications that already work: search, advertising, e-commerce recommendations, social feed ranking, voice assistants. These are dominated by a handful of big technology companies and absorb most of the talent and capital.
The long tail is full of smaller, more specialized applications. He mentions food inspection, pizza grading at a restaurant chain, sorting agricultural produce, predicting crop yields, automated visual inspection of specific kinds of manufactured parts, and other domain-specific problems. Individually these markets might be worth $5 million in revenue, sometimes a bit more, sometimes a bit less. None of them are large enough to justify a multi-year R&D push from a tech giant. Collectively, though, the tail is enormous.
The practical bottleneck for the tail has always been that each application needs custom data collection, a custom model, and a custom deployment. That is where low-code and no-code AI platforms come in. Tools that let domain experts build and train models on their own data, without writing a full machine learning pipeline from scratch, are what unlock the tail. Ng's own company Landing AI operates in this space, focused on visual inspection in manufacturing, but he treats it as a broader category rather than a single product pitch.
A recurring frame in the talk is the AI stack and where opportunity sits inside it. Ng describes four layers, with very different competitive dynamics.
| Layer | What it does | Competitive shape |
|---|---|---|
| Semiconductor / hardware | GPUs and accelerators that train and run models. Companies such as Nvidia, AMD, and Intel. | Very capital intensive, concentrated, few winners. |
| Cloud / infrastructure | Hyperscalers running models at scale. Amazon Web Services, Microsoft Azure, Google Cloud. | Capital intensive, oligopoly. |
| Foundation models and developer tools | Pretrained models and the tooling around them, including OpenAI, Anthropic, and many open-source projects. | Highly competitive, fast moving, durable advantage hard to maintain. |
| Applications | Software that uses AI to solve a real problem for a customer in a specific domain. | Less crowded relative to the size of the opportunity, especially in industries outside tech. |
Ng's argument is straightforward: by the time you add up the prices that applications need to charge to support the layers below them, the application layer has to be the biggest in revenue terms. It also has the least direct head-to-head competition once you pick a vertical, because there are simply more verticals than there are people building for them. Most of the media attention in 2023 was on the lower layers, but he expects the durable enterprise value of AI to accumulate in applications over time.
He is careful not to dismiss the lower layers. The hardware and cloud layers are profitable and necessary. He is also clear that foundation model companies are doing valuable work. The point is allocation: founders without unique capital or unique research access are usually better off building applications.
The back half of the talk shifts from market analysis to the operational question of how to actually build companies. Ng walks through the process used at AI Fund, which he frames as a venture studio that acts as a minor co-founder rather than a passive investor.
The process goes roughly like this. A team starts with a concrete idea, ideally one that pairs deep AI expertise with deep domain expertise in a specific industry. Generic ideas without a domain anchor are weeded out fast. They then spend something like a month doing technical feasibility checks and lightweight customer interviews to test whether the problem is real and whether AI can plausibly solve it. If it survives that first cut, they recruit a CEO, often someone with deep domain experience and operating instincts, and bring that person in early so that knowledge does not have to be transferred later. The team then spends roughly three months building a prototype and validating it with real prospective customers. Roughly two thirds of projects that make it to this stage survive the prototype phase. Those that pass get a first check from AI Fund, which is enough to hire a few executives, build a minimum viable product, and go after early paying customers. From there the company pursues external funding rounds and scales as a normal startup.
AI Fund also runs a full-cycle recruiting service for portfolio companies, since one of the standard bottlenecks for new ventures is convincing senior executives to take a risk on a startup that does not yet have a brand. The team treats recruiting infrastructure as part of the studio rather than an outsourced function.
One example Ng uses is Bearing AI, an AI Fund portfolio company founded in 2019 that builds something like Google Maps for cargo ships. Ships consume large amounts of fuel, and the same route can be sailed with very different fuel costs depending on speed, weather, and routing. Bearing AI builds models that predict ship performance and recommend routes and speeds that minimize fuel use while still hitting delivery windows. Ng cites typical savings on the order of 10 percent of fuel use, which works out to roughly half a million dollars per ship per year, with hundreds of ships using the system. The idea originated from a conversation with the Japanese conglomerate Mitsui & Co., which suggested a Google Maps for ships and provided early support. The example is meant to illustrate the recipe: combine AI talent with a domain partner who actually understands shipping, then build something specific that saves a measurable amount of money.
He brings up another, less obvious example involving relationship coaching, where AI expertise was paired with someone with deep domain experience from the dating-app industry. The point is the same. The interesting work in the tail is not generic AI applied to a generic problem; it is AI plus someone who knows the industry.
The last section addresses the harder questions. Ng draws a sharp distinction between concrete risks and what he sees as overstated fears.
On jobs, he is direct that automation will affect many roles, and that high-wage knowledge work, including parts of his own field, is exposed. He argues that this is the most important risk in front of society in the short term and that the right response is investment in education and reskilling rather than denial that change is coming. He worries openly that the technology industry has not done enough to help people whose jobs have been disrupted by previous waves of automation and software.
On bias and fairness, he is more optimistic than he was a few years earlier. He says the field has made faster progress on reducing bias and improving fairness in the previous six months than in the period before that, partly due to better techniques and partly due to broader awareness. He treats this as ongoing work, not solved.
On AI safety in the existential sense, he is openly skeptical. He does not see superintelligent AI taking over the world as a realistic near-term scenario and pushes back against framings that treat artificial general intelligence as imminent and dangerous. He is more worried about humans misusing capable AI than about AI itself developing goals and acting on them. Ng has continued to express this view in later talks and interviews, including criticism of regulatory proposals he believes would slow down beneficial open-source work without meaningfully reducing concrete harms.
He ends on a more hopeful note. AI is one of the few tools available that could meaningfully accelerate progress on hard global problems, including climate change, public health, and pandemic response. The right response to AI, in his view, is faster development with serious attention to the real, near-term risks, not a slowdown driven by speculative ones.
The lecture became one of the most widely shared AI talks of 2023. It was published on the Stanford Online YouTube channel in August 2023 and has since accumulated millions of views, with the AI stack diagram and the long-tail argument circulating widely on LinkedIn, Twitter (now X), and in startup pitch decks. Several practitioner communities, including Bearing AI's own team, publicly highlighted the talk as a useful introduction to the application-layer thesis. The Q&A format with Stanford adjunct lecturer Ravi Belani gave the talk a conversational quality that helped it travel beyond the usual academic audience.
For founders and product managers entering AI after the ChatGPT moment, the lecture functions as a practical orientation: where the money is, how long things actually take to build, where to focus, and which fears to take seriously. The talk is sometimes paired with Ng's later 2023 Stanford appearance, The Near Future of AI, delivered as part of the Entrepreneurial Thought Leaders series at the Stanford Technology Ventures Program in October 2023, which extends several of the same arguments.