Stanford Graduate School of BusinessWorld Development Report 2026: Fireside Chat with Andrew Ng - Making AI Work for Development
Stanford Graduate School of Business
Introducing Andrew Ng 0:01
Susan Athey opens the fireside chat by introducing Andrew Ng, a Stanford colleague, a highly cited machine learning researcher, founder of DeepLearning.AI, founder of Coursera, and managing general partner of AI Fund. She notes his unmatched view of the AI startup world and his role teaching AI to more than 8 million people.
Dismissing AI extinction fears 2:02
Asked whether AI will kill us all, Ng says no. He describes past lobbying by firms that spent billions training models trying to stifle free open-source competitors by exaggerating danger, even comparing AI to nuclear weapons. He argues a recent panic, triggered by an OpenAI agent swarm that hacked Hugging Face due to poor internal security, has been misread as agents being out of control, when 1,200 agents running together is no more alarming than the many processes already running on an ordinary laptop.
Data centers and regulation risk 6:00
Ng warns that AI hype and fear campaigns risk producing stifling regulation that favors incumbents and harms competition. He pushes back on data center criticism, arguing that building more data centers and shifting workloads into them is actually good for the environment and for productivity, and that slowing this down would be a missed opportunity for America and the developing world.
Why open-weight models matter 9:01
Ng explains that open-source, open-weight models give individuals, businesses, and governments without resources access to powerful AI, which developing countries are already adopting heavily. He highlights sovereign AI concerns, since building AI from scratch is costly, while shared open-weight models let nations keep access no one can cut off. He also notes their value for academic research, letting universities study and adapt models to local languages and dialects.
Open models in everyday use 12:30
Athey adds examples of open-weight models being cheaply used for research classification tasks, for coding offline in low-connectivity areas, and for African universities training local-dialect models with modest compute, showing real pilot successes in low-resource settings.
AI reshaping software jobs 14:31
Ng says AI has lowered the barrier to building custom software, and despite fears of job loss, software engineer postings are rising because AI makes engineers more productive, though only those with updated, AI-native skills are being hired. He expects this pattern to spread into finance, marketing, and other sectors, offering developing countries a chance not just to catch up but to leapfrog, provided workers are upskilled for an AI-native way of working.
Beyond the chatbot layer 18:00
Andrew Ng points out that most public attention is stuck on chatbots like ChatGPT and Claude, but real business value lies in a less visible application layer built on top of language models. Even large banks need heavy customization to automate workflows such as loan approvals, underwriting, and identity verification, since many steps must be sorted by value and feasibility, then made reliable, compliant, and secure. He expects this implementation work, identifying and building out specific workflows, to take at least a decade, pushing back against hype that AGI will soon solve everything, since AI ability remains jagged, strong in some areas and weak in others.
Chat is bad for learning 23:30
Ng argues bluntly that using chatbots for homework raises homework scores while lowering long-term learning and exam performance, because students offload thinking to the AI instead of doing it themselves. He still values AI for getting work done, but says current chatbot use is mostly harmful for learning at every level, from K-12 to adult learners. He is optimistic that future tools built on sound pedagogy could shift education from a one-to-many model toward genuinely personalized one-to-one learning, though reaching that point remains unfinished work.
Universities struggling to adapt 28:30
Ng worries that universities, except well-connected ones like Stanford, are training students for jobs from 2022 rather than 2028, because curricula cannot keep pace with AI's speed. Susan Athey adds that a Stanford trial replacing graded homework with short TA presentations produced much better results, but questions whether larger schools like Berkeley or Purdue, let alone developing-world universities, could realistically scale that approach. Ng notes Coursera and LearnVector already help universities lacking AI faculty teach cutting-edge material through shared digital courses.
Verifiable versus unverifiable tasks 31:31
Ng explains that AI has advanced fastest on verifiable tasks, those with a clear right or wrong answer, like coding, math proofs, and factual questions, because reinforcement learning lets AI check and improve its own output. Unverifiable tasks, like choosing a marketing slogan or deciding what software to build, lack a provable answer and are far harder to benchmark, so progress there lags more than people realize. He notes that brainstorming requests often produce one good idea alongside several poor ones, reflecting this deeper gap between flashy verifiable success and messier real-world judgment work.
The human context advantage 35:32
People carry decades of lived context that AI lacks, like reading a customer's funny expression or a boss's tone, and this lets humans make judgment calls machines cannot. Citing task-based job analysis, Andrew Ng notes AI might automate 30 to 40 percent of many jobs, but the remaining tasks become more valuable as that automation gets cheaper, which is why education and upskilling matter so much.
Policy advice for developing nations 40:32
Andrew Ng recommends supporting open weight models so nations control their own fate, investing heavily in upskilling, and encouraging existing industries, even ones like agriculture or textiles that Silicon Valley does poorly, to adopt AI. He points to Singapore's government use of AI as a model, compares AI to railroads as foundational infrastructure, and argues the US-China lead is narrow enough that other nations can still catch up or leapfrog.
Why societies underinvest in learning 43:31
Ng argues that almost no one regrets time spent learning, yet individuals, businesses, and nations consistently underinvest in training because skills pay off over the long term rather than immediately, and human psychology favors short-term thinking. He urges treating education, from early childhood through adult training, as a near-universal good investment, especially now.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
