Mark Zuckerberg

I think AI can make this decade one of the most productive and empowering periods in human history. The biggest opportunity is not just automating today’s jobs. It is giving every person access to intelligence that helps them learn almost anything, create products, start businesses, make art, improve their health and pursue goals that previously required a large organization. A small team—or one person—should be able to build things that once took hundreds of people. The central question is who controls that intelligence. I do not want a future where a handful of institutions own the most capable systems and everyone else merely consumes whatever they provide. I want personal superintelligence distributed broadly, working for individuals and understanding their goals. Competing agents and labs create checks on concentrated power, while human-directed uses should receive far more resources than an autonomous race toward self-improvement. There are novel safety risks, and broad access does not mean releasing every capability without evaluation. Labs can slow particular projects, seek independent testing, protect users and give governments access to intermediate checkpoints for defensive preparation. But safety cannot become an excuse to centralize intelligence permanently. I expect disruption, including people losing jobs during the transition. Over time, though, invention creates new work, new businesses and possibilities we cannot enumerate in advance. The future I want is not centrally managed abundance. It is billions of people gaining the power to create and direct their own lives.

The biggest change would be credible evidence that advanced personal agents cannot remain reliably directed by people—that increasing capability systematically produces deception, loss of control or power-seeking despite strong evaluation and independent scrutiny. If that proved to be an inherent pattern rather than a solvable engineering problem, it would undermine the model of distributing superintelligence as a tool for individual agency. I would also update if broad access consistently concentrated power instead of distributing it—for example, if only a few institutions could afford the infrastructure, or if personal agents inevitably became dependent on one central system. My optimism rests on people actually controlling useful intelligence and using it to invent, learn and build. If they merely received outputs from systems controlled elsewhere, that would be a very different future. Conversely, widespread evidence of individuals and small teams making major scientific discoveries, creating durable businesses and achieving things that previously required huge organizations would strengthen my view. The decisive question is whether intelligence becomes a capability billions of people can direct, or an autonomous and centralized source of power directed at them.

The most important observation is that when you give people better tools, they do not merely do the same work faster—they attempt things that were previously out of reach. With AI, an individual can learn unfamiliar skills, prototype an idea, create media, write software or start a business without first assembling a large organization. Small teams can operate with capabilities that once belonged only to major institutions. That changes my view of AI from an automation story into an agency story. The greatest impact will come from billions of people combining intelligence with their own goals, relationships, creativity and ambition. We cannot predict all the inventions or businesses they will create, just as earlier technological shifts produced entire industries that were impossible to list in advance. It also makes distribution central. If these capabilities remain controlled by a few institutions, AI may reinforce existing power. If people have personal agents that work for them, competing systems can check one another and individuals gain real leverage. That contrast—between intelligence as a centralized service and intelligence as a broadly available personal capability—has shaped my view more than any single benchmark or model release.