Andrew Ng: Opportunities in AI - 2023

Andrew Ng frames AI as a general-purpose "new electricity," with supervised learning currently driving the vast majority of economic value—such as Google's massive ad revenue—while generative AI emerges as a powerful new frontier. He highlights that the key breakthrough has been large-scale supervised learning, where bigger neural networks trained on more data keep improving, and that generative AI builds on this by predicting the next token in text. Prompt-based development is dramatically accelerating app creation, cutting build times from months to about a week and enabling a simple sentiment classifier in minutes, making prompting a core developer skill. The lasting opportunity, however, lies not in easily replicated AI wrappers but in deeply defensible applications, much like the iPhone enabled Uber and Airbnb. To spread AI beyond consumer software, low-code/no-code and data-centric AI tools are lowering customization costs, allowing domain experts to build custom systems on their own proprietary data across industries.

• ▶ 1:07 Andrew frames AI as "the new electricity" and emphasizes that it is a general-purpose technology with many different applications, not just one use case.
• ▶ 2:14 Supervised learning is one of the most important AI tools, working by learning input-to-output mappings for tasks like spam detection, ad clicks, self-driving car perception, and factory visual inspection.
• ▶ 4:22 Over the last decade, the key scaling recipe has been large-scale supervised learning: very large neural networks keep improving as they are fed more data, while smaller models plateau.
• ▶ 5:27 Andrew Ng says this decade adds generative AI as a powerful new tool on top of supervised learning, giving examples like ChatGPT and varied text completions.
• ▶ 6:20 Text generation works by using supervised learning to repeatedly predict the next word/token, training very large models on hundreds of billions or more words to create LLMs like ChatGPT.
• ▶ 7:54 LLMs are underappreciated as developer tools, not just consumer tools, cutting application build times from months; the traditional supervised learning workflow often took 6–12 months for a commercial-grade system.
• ▶ 8:58 Prompt-based AI dramatically accelerates development: writing prompts takes minutes to hours, deploying takes hours to days, and some apps that once took six months to a year can now be built in about a week.
• ▶ 9:52 A live demo shows a full sentiment classifier requires only a small amount of code—built with Python and OpenAI tools—and can be created by developers worldwide in maybe 10 minutes.
• ▶ 11:05 Andrew Ng emphasizes teaching prompting as a developer skill, reinforcing that prompt-based AI is opening a new, much faster path to building practical AI applications.
• ▶ 11:41 The vast majority of AI's financial value today comes from supervised learning, not generative AI.
• ▶ 11:49 Supervised learning's scale is massive: a single company like Google can generate over $100 billion per year from it, with millions of developers building applications.
• ▶ 12:10 Generative AI is the exciting new addition to the landscape, representing a newer and emerging source of opportunity.
• ▶ 12:25 Supervised learning is already "truly massive" and is expected to roughly double in value over the next three years.
• ▶ 12:36 Generative AI is currently much smaller but is predicted to more than double, driven by developer interest, venture capital investment, and corporate exploration.
• ▶ 12:59 The light shaded region represents the key opportunity area where both new startups and large incumbents can create and capture value as these technologies scale.
• ▶ 13:05 Andrew Ng emphasizes that all the technologies discussed are general purpose technologies, applicable across many tasks and industries.
• ▶ 13:17 The core ongoing work is identifying and executing concrete use cases for supervised learning, which is “nowhere near finishing” and still holds huge untapped value.
• ▶ 13:26 Generative AI is another powerful general purpose technology that extends the set of possible applications, adding a new frontier on top of ongoing supervised learning work.
• ▶ 14:25 Lensa was a thin software layer built on top of someone else's APIs, making it a useful product but not a defensible business.
• ▶ 15:06 Like the flashlight app, easily replicated AI wrappers will get commoditized and absorbed into platforms, failing to create long-term value.
• ▶ 15:26 The real opportunity is building deeply defensible, hard AI applications that create lasting value, similar to how the iPhone enabled Uber and Airbnb.
• ▶ 15:39 Andrew Ng emphasizes "very long-term value," framing the discussion around durable impact rather than short-lived hype.
• ▶ 15:42 He introduces the first major trend: AI as a general purpose technology.
• ▶ 15:44 He begins to elaborate on this trend, setting up an explanation of AI's broad and versatile applicability.
• ▶ 15:51 AI's economic value today is concentrated almost entirely in consumer software and internet applications, with very early adoption elsewhere.
• ▶ 16:24 The "head" of the value curve holds a few billion-dollar projects (advertising, search, e-commerce recommendations), but these require massive user bases that most industries don't have.
• ▶ 18:16 The key barrier to wider adoption is the high cost of customizing AI for a long tail of tens of thousands of ~$5 million projects (e.g., pizza inspection, wheat height measurement).
• ▶ 18:36 High customization costs are a central barrier to applying AI across many use cases.
• ▶ 18:52 Low-code and no-code tools are emerging to let end users customize AI systems themselves, reducing dependence on AI experts.
• ▶ 19:19 Self-service tools are essential for domain experts to build and maintain custom AI on proprietary data that isn't available on the public internet.
• ▶ 19:29 Custom AI systems are becoming more accessible because businesses can build and maintain AI on their own data, not code.
• ▶ 19:48 Data-centric AI is a key enabler: instead of writing code, businesses can just provide data, which is far more feasible.
• ▶ 20:04 This trend matters because AI value is still concentrated in tech and consumer software, but these tools help spread it across all industries and the wider economy.
• ▶ 20:24 Andrew Ng recaps the two key AI trends before asking how to act on them.
• ▶ 20:25 Trend 1: AI is a general-purpose technology with many concrete use cases still to be realized across industries.
• ▶ 20:31 Trend 2: Low-code/no-code tools are lowering barriers to entry, enabling AI deployment in more industries.
• ▶ 20:36 The central question is posed: “How do we go after these?” to frame the next discussion.
• ▶ 20:40 Andrew Ng identified a problem five years ago: many valuable AI projects were becoming possible, but he needed to figure out how to actually execute them.
• ▶ 21:14 The diversity of AI opportunities (across shipping, education, finance, healthcare, etc.) made it inefficient to manage them all within a single big-tech team, leading him to conclude that starting many separate companies was the most efficient approach.
• ▶ 21:27 This insight led to the creation of AI Fund, a venture studio that builds startups, while he also notes that established companies can leverage their distribution advantage to successfully integrate AI into their existing products.
• ▶ 22:03 Andrew Ng notes the hardware/semiconductor layer offers fantastic opportunities but is very capital intensive and concentrated, with few winners—so he personally avoids playing there.
• ▶ 22:21 The infrastructure layer is similar: great opportunities but capital intensive and concentrated, so Ng tends not to play there either.
• ▶ 22:31 The developer tool layer (e.g., OpenAI's API) is hyper-competitive and crowded, but likely to have some mega winners; Ng "sometimes plays" there only when he has a meaningful technology advantage.
• ▶ 22:56 Building in the application layer, rather than infrastructure or tooling, earns startups a better shot at becoming "mega winners" in AI.
• ▶ 23:08 Despite media hype concentrating on infrastructure/tooling, that layer can only succeed if the application layer is even more successful and generates enough demand and revenue.
• ▶ 23:26 The application layer must be the economic engine: it needs to thrive enough to pay for and support the infrastructure and tooling companies beneath it.
• ▶ 23:31 Andrew Ng introduces Armor Raw, a real example of an AI startup focused on romantic relationship coaching, noting his active involvement with the CEO.
• ▶ 23:48 With self-deprecating humor, Andrew admits he knows "nothing really about Romance" — joking his wife would confirm this — highlighting the unexpected application of AI in a domain he lacks expertise in.
• ▶ 24:04 The startup’s uniqueness comes from pairing Andrew’s team’s AI knowledge with relationship expertise from Renata, former CEO of Tinder, showing how complementary skills can create something distinctive.
• ▶ 24:31 The key opportunity is the scarcity of teams with deep expertise in both AI and a specific domain (e.g., human relationships), creating a major opening for founders.
• ▶ 24:38 The application layer offers many exciting opportunities characterized by very large markets and very light competition relative to the opportunity’s magnitude.
• ▶ 24:48 While competitors exist, competition at the application layer is much less severe than at the DevTools or infrastructure layers, making it a more forgiving environment.
• ▶ 24:58 Andrew Ng shares the startup recipe his team has developed after many years of iteration, and it represents how they build startups.
• ▶ 25:15 His team generates many internal ideas from partners, which serve as the raw material for startup creation.
• ▶ 25:21 He introduces Bearing AI as a concrete example, using AI to make ships more fuel efficient.
• ▶ 25:27 Andrew Ng describes Bearing AI's core concept as making ships more fuel efficient.
• ▶ 25:29 Mitsui, a large Japanese conglomerate, approached him to build an AI business for reducing fuel consumption in shipping.
• ▶ 25:47 The vision was a "Google Maps for ships" that advises on steering to reach destinations on time while using about 10% less fuel.
• ▶ 25:59 The AI Fund spends about a month validating each idea, checking technical feasibility and customer demand as a go/no-go checkpoint.
• ▶ 26:21 A key lesson: bring the CEO in immediately after validation, rather than working on the project alone for a long time first.
• ▶ 26:29 Bringing the CEO in early reduces knowledge transfer burden and prevents the CEO from revalidating everything, making the process much more efficient.
• ▶ 26:42 Finding a strong, experienced CEO (repeat entrepreneur Dylan Kyle) is the first critical step in executing an AI startup within the AI Fund.
• ▶ 26:49 The team uses a structured three-month process (six two-week sprints) to build a prototype while conducting deep customer validation.
• ▶ 27:02 Roughly two-thirds of projects survive the first go/no-go check-in, showing a rigorous filtering bar before a company gets funding to hire a full executive team and build the MVP.
• ▶ 27:16 Get the product working and secure real customers first to validate the startup's value proposition before seeking larger investments.
• ▶ 27:20 After achieving product and customer traction, successfully raise additional external rounds of funding.
• ▶ 27:24 Use the raised capital to keep growing and scaling the company, with pride in the team's execution.
• ▶ 27:27 Bearing AI's collaboration with Mitsui has led to hundreds of ships steering differently, achieving 10% fuel savings worth about $450,000 per ship per year, with major environmental benefits.
• ▶ 27:54 Andrew Ng credits Dylan for the startup's existence, while noting Mitsui brought the original idea to the team.
• ▶ 28:05 Ng highlights that he would never have conceived the idea himself due to lack of maritime expertise, showing how deep industry knowledge combined with AI capability drives innovation.
• ▶ 28:28 Andrew defines his "swim lane" as AI only, having learned that mastering many other domains at once is unrealistic.
• ▶ 28:47 He focuses on providing accurate technical validation and building strong technical teams to turn AI capabilities into real products.
• ▶ 29:00 Success comes from partnering with subject-matter experts who supply industry depth, combining their domain knowledge with his AI expertise.
• ▶ 29:14 Andrew Ng prefers engaging only with concrete ideas, finding broad design thinking exploration "very slow" and inefficient.
• ▶ 30:16 Even a bad but concrete idea is useful because it can be validated or falsified quickly and gives the team clear direction.
• ▶ 30:49 Many subject-matter experts already have well-thought-out ideas and just need a build partner; finding these ideas is far more efficient than starting from scratch.
• ▶ 31:18 Andrew Ng transitions to the final topic of risks and social impact, framing it as a short but important segment.
• ▶ 31:26 He acknowledges AI's power and states his guiding principle: only work on projects that move humanity forward.
• ▶ 31:31 Ng reveals his teams have rejected cool, financially sound projects on ethical grounds, prioritizing ethics over profitability.
• ▶ 31:42 Andrew Ng observes that people show surprising creativity in coming up with really bad AI project ideas.
• ▶ 31:48 The key warning sign: these ideas "seem profitable but really should not be built," so apparent profit isn't enough to justify a project.
• ▶ 31:55 His team deliberately "kills" a few projects that fail this deeper test, even if they look good financially.
• ▶ 31:57 Current AI systems still have significant problems with bias, fairness, and accuracy.
• ▶ 32:05 AI is improving quickly and has become less biased and more fair than just six months ago.
• ▶ 32:14 Bias and fairness remain real issues, but many practitioners are working hard to solve them.
• ▶ 32:26 Andrew Ng identifies job disruption as the biggest risk of AI, despite the tremendous value it creates.
• ▶ 32:33 Citing a Penn/OpenAI study, he notes that higher-wage jobs are now more exposed to AI automation, unlike previous automation waves that hit lower-wage jobs hardest.
• ▶ 33:08 He urges citizens, corporations, governments, and society to support people whose livelihoods are disrupted by AI.
• ▶ 33:25 Andrew Ng notes that every major AI breakthrough, such as deep learning and now generative AI, predictably triggers a new wave of AGI hype.
• ▶ 33:49 He estimates AGI—defined as AI that can do anything a human can do—is still 30 to 50 years away or more, and not coming anytime soon.
• ▶ 34:02 He argues that comparing biological and digital intelligence is awkward, since LLMs are smarter than humans in some dimensions but much dumber in others, making the "everything a human can do" benchmark a odd goal.
• ▶ 34:46 Andrew Ng directly rebuts AI extinction risk, saying "I just don't see how AI creates any meaningful extinction risk of humanity."
• ▶ 34:55 He argues humanity already knows how to steer powerful entities like corporations and nation-states, so AI can be overseen similarly.
• ▶ 35:13 Ng rejects "hard takeoff" scenarios, instead emphasizing that AI will develop gradually, giving society time to adapt and manage it.
• ▶ 35:33 AI can be kept safe through oversight and management, making it a controllable technology rather than an inherent threat.
• ▶ 35:36 The real extinction-level risks to humanity are pandemic, climate change, and asteroid impact.
• ▶ 35:57 AI, even with greater intelligence, is not considered among the real existential dangers compared to those threats.
• ▶ 36:02 Andrew Ng argues that increasing intelligence is central to solving humanity's long-term challenges, framing AI as key to helping humanity survive and thrive over the next thousand years.
• ▶ 36:11 He explicitly pushes back against proposals to slow AI progress, suggesting instead that AI should move as fast as possible.
• ▶ 36:15 Ng frames AI acceleration as aligned with humanity's long-term prosperity, not as a trade-off between AI and safety.
• ▶ 36:15 Andrew references letting AI “go as fast as possible,” then signals he is moving into his summary.
• ▶ 36:21 His core message: AI is a general-purpose technology that creates many new opportunities for everyone.
• [36:21–36:34] The key work ahead is building concrete use cases on top of AI—turning its broad potential into real, tangible applications.
• ▶ 36:34 Andrew Ng emphasizes the importance of building concrete, applied AI use cases over abstract potential.
• ▶ 36:37 He expresses hope to engage more with the audience on these opportunities in the future.
• ▶ 36:44 He delivers final thanks, with the talk closing to applause at 36:46.

• ▶ 0:01 Opening and Supervised Learning Foundations (0:01 - 5:27) - - Andrew is introduced, frames AI as a general-purpose technology, and explains supervised learning and the last decade’s scaling recipe.
• ▶ 5:27 Generative AI and LLM Developer Tools (5:27 - 8:56) - - Covers the move to generative AI, repeated next-word prediction, and large language models as developer tools.
• ▶ 8:56 Prompt-Based AI and Live Demo (8:56 - 11:27) - - Contrasts traditional AI build/deploy timelines with prompt-based AI, demonstrates a sentiment classifier, and mentions prompting classes.
• ▶ 11:27 AI Opportunity Landscape and Adoption Barriers (11:27 - 18:39) - - Discusses the value of supervised learning vs generative AI, long-term opportunities, and why AI adoption lags in the long tail.
• ▶ 18:39 Low-Code Tools, AI Fund, and Startup Process (18:39 - 27:30) - - Explores low-code/data-centric AI, the AI Fund approach, the AI stack, and the Bearing AI startup-building process.
• ▶ 27:30 Startup Lessons, Ethics, and AI Risks (27:30 - 36:52) - - Highlights Bearing AI’s impact, lessons on partnerships and concrete ideas, and closes with bias, fairness, and job-disruption risks.