Andrew Ng

I think AI means an enormous expansion in what people can build and afford to do. Many valuable projects—better education, healthcare tools, scientific research, customized software, and services for small businesses—have been neglected because human effort was too costly. AI lowers that cost. The right starting point is not abstract speculation about a machine apocalypse; it is a real problem, a useful application, testing, and iteration. The gains will be uneven. Coding agents may make implementation much faster while product decisions, infrastructure, validation, and coordination remain bottlenecks. That changes jobs and lets AI-native teams accomplish more with fewer people, but it does not mean every task or profession vanishes. People who learn to use AI well—and who can decide what is worth building—will have growing leverage. Some workers will face painful disruption, but broad claims that AI is simply eliminating work are exaggerated. There are real risks, especially in cybersecurity. Agents can search patiently and chain vulnerabilities. The response is stronger isolation, monitoring, rapid patching, and clear accountability for builders and operators. Attacks still require actions that defenders can observe, and defenders often possess more information about their own systems. I expect that to provide an important long-run advantage. Most safety progress comes from building systems, finding concrete failures, and fixing them. A pause postpones that learning while adversaries continue. So my conclusion is straightforward: the useful applications greatly outweigh the risks, and we should keep building.

The biggest change would be strong, repeated evidence that AI-enabled attackers have a durable advantage over defenders—that even well-isolated, carefully monitored, rapidly patched systems can be compromised faster than organizations can detect and recover. My optimism about cyber risk depends heavily on defenders having more information about their own systems and attacks requiring observable actions over time. If that forecast proved consistently wrong, I would update substantially. I would also change my view if useful applications repeatedly failed to deliver real-world value despite good engineering—if lower implementation costs did not translate into better products, scientific progress, education, healthcare, or new businesses because validation and coordination bottlenecks overwhelmed the gains. But today the evidence points the other way: AI is already making many tasks cheaper and enabling previously unaffordable projects. What would not change my view is another dramatic demo, speculative extinction story, or isolated agent failure. Those are reasons to test systems, improve monitoring and isolation, and hold builders and operators accountable—not reasons to conclude that progress itself should stop.