ET@Davos: Andrew Ng on Job Displacement, AGI Myth and India’s Crossroads

In this conversation, Andrew Ng delivers a pragmatic view of AI: he insists AI proficiency is now non-negotiable across roles, yet he dismisses near-term AGI hype, arguing the term has lost meaning and no known technology is a clear path to it. He explains that modern LLMs work through pre-training on massive data followed by alignment to make them helpful and safe, and he remains confident that scaling still has untapped potential. Ng sees "people change" as the biggest adoption hurdle for CEOs, and he urges India to upskill rapidly to leapfrog established players rather than risk being left behind. He also defends AI's capacity for original thinking as a practical, behavior-based judgment, supports open-weight models and global cooperation, and criticizes US immigration policy as a self-inflicted wound. Ultimately, his challenge is clear: learn the skills, ignore the hype, and seize the opportunity to build boldly.

• ▶ 0:00 AI skills are non-negotiable: Andrew Ng says he would not hire a software engineer who doesn't use AI tools sophisticatedly, and this expectation extends to marketers, HR, and other roles.
• ▶ 0:05 AGI hype is not credible: he dismisses near-term AGI claims, noting no known technology today is a path to AGI, and the term has been redefined so often it has lost meaning.
• ▶ 1:26 India's upskilling is critical: with AI disrupting IT services, job displacement could be significant unless professionals upskill rapidly—the risk is "dramatically less productive" individuals, but the opportunity is leapfrogging ahead.
• ▶ 4:43 Andrew Ng says the hype curve rose faster than the business reality, and substantial work remains on real applications, agents, and agentic workflows.
• ▶ 6:34 He argues LLMs alone are not a path to AGI, warns that redefining AGI at a lower bar misleads leaders and students, and says the term has almost lost meaning.
• ▶ 8:54 Ng identifies "people change" as the #1 adoption issue for CEOs, noting that decentralized "thousand flowers" strategies often fail because they produce only point solutions.
• ▶ 11:03 The interviewer defines LLMs as operating by predicting the next word or sequence of sentences based on training data.
• ▶ 11:12 They pose the central puzzle: if models only predict the next word, why do they sound so intelligent?
• ▶ 11:26 They question whether this next-word-prediction view is an oversimplification of how the models actually function.
• ▶ 11:33 LLMs are trained in two stages: first, pre-training teaches the model to predict an expert on the internet, but this alone makes it regurgitate random internet-style content.
• ▶ 11:49 The second step, alignment, uses only a small amount of data to make the model produce more sensible, factual, and helpful outputs.
• ▶ 12:12 Alignment is done via reinforcement learning or related techniques to make the model helpful, honest, and harmless — refining pre-trained knowledge into a controlled, beneficial tool.
• ▶ 12:55 The core recipe behind modern LLMs is building a sufficiently large deep learning model and feeding it enough data, allowing it to mimic patterns in data—such as helpful responses to questions—remarkably well.
• ▶ 13:32 This scaling approach dates back to the Google Brain project, where Andrew Ng was famously warned by colleagues that it was a "bad career move," though his confidence in the data proved justified.
• ▶ 14:01 The scaling-and-data recipe still holds untapped potential; in Ng's words, the field is "not yet done squeezing the juice out of this particular lemon."

• ▶ 14:10 The interviewer asks whether scaling AI is true progress, using a model solving a 300-year-old unsolved math problem as a benchmark for civilizational advancement.

• ▶ 14:53 Andrew Ng responds: "I think we'll get there. I think it's not easy," noting AI already assists rather than replaces human researchers in biology, materials science, and computer science.

• ▶ 15:18 The interviewer interrupts, questioning whether such successes are just "brute force" computing rather than original human-like thinking—a concern cut off mid-sentence at ▶ 15:26.

• ▶ 15:40 Andrew Ng affirms that AI will achieve original thinking and creativity, but immediately warns that the key terms need an important caveat.
• ▶ 15:43 Ng explains that "original thinking" and "creativity" are philosophical concepts, not scientifically defined terms — creativity is "in the eye of the beholder" and not measurable.
• ▶ 16:26 His pragmatic conclusion: if AI behaves in a way that looks creative or original to humans, he is happy to call it creative and capable of original thinking.
• ▶ 16:50 Andrew Ng says AI learns very differently from humans; while AI has absorbed far more data than any human could, it remains “dumber” in key ways yet surpasses humans in others.
• ▶ 17:51 Neuroscience has only provided “vague inspiration” for AI; Ng recalls reading stacks of neuroscience papers during Google Brain’s early days, but the field “barely understands how the human brain works,” so it was “mostly not useful.”
• ▶ 18:38 Using the analogy that “we didn't build airplanes by mimicking birds,” Ng explains that understanding intelligence—not directly copying the brain—could one day guide AI, but we still lack a true “theory of intelligence.”
• ▶ 19:23 Andrew views the AI arms race positively: recent frontier models from many labs—Claude Code, Gemini 3, OpenAI Codex, GPT-5.x—are all strong, giving the field "many strong horses in the race."
• ▶ 20:43 On sovereign AI, Andrew argues countries like India should invest in and cooperate on open-source/open-weight models rather than control all AI infrastructure; he also notes Chinese open models are a huge force and carry geopolitical influence.
• ▶ 23:08 Calling the US's less welcoming stance toward immigrants "awful" and "a huge unforced error," Andrew says much of America's strength comes from its people, including Indian and Chinese engineers and executives.
• ▶ 23:38 The US is making an “unforced error” by becoming less welcoming to skilled immigrants and international students, hurting a key historical advantage.
• ▶ 26:28 Because AI is so disruptive, “old rules no longer hold”—giving India a real chance to leapfrog entrenched US and China players.
• ▶ 27:52 Andrew’s challenge to India: learn the skills, ignore the hype, and seize the opportunity to build something bold and new.

• ▶ 0:00 AI Skills, Hype, and India’s IT Services (0:00 - 4:11) - Andrew discusses AI hiring skills, AGI skepticism, sovereign AI, immigration, and AI’s impact on India’s IT sector.
• ▶ 4:12 GenAI Business Reality and AI Adoption (4:12 - 10:53) - Andrew covers the ChatGPT moment, GenAI hype versus business reality, the LeCun debate, and how leaders should drive AI adoption.
• ▶ 10:53 LLM Mechanics and the Frontiers of Intelligence (10:53 - 19:23) - Andrew explains how LLMs are trained, why scaling makes them sound smart, and the open questions around AI, creativity, and human learning.
• ▶ 19:23 AI Arms Race, Sovereign AI, and Open Models (19:23 - 23:58) - Andrew discusses frontier-model competition, sovereign AI, China’s open-source influence, and Indian AI talent and immigration concerns.
• ▶ 23:58 India’s AI Leapfrog and Closing Remarks (23:58 - 28:21) - Andrew gives final advice on how India can leapfrog with AI, then closes with thanks and a sign-off.