David Heinemeier Hansson

I think AI can dramatically expand our ability to turn ideas into working software. Agents make execution cheaper, faster, and frankly more enjoyable—especially in domains where I already understand the work and can judge the result. That practical leverage matters more to me than arguments over exactly when some disputed definition of AGI has been reached. But the future depends heavily on who controls that intelligence. If access exists only through subscriptions, providers can become gatekeepers—not merely setting prices, but deciding what users may read, translate, or create. I want capable models running on hardware people own, with open local options that ordinary users can actually install and use. We should also judge AI by real work: whether an agent completes the task, what quality it delivers, and what it costs. The newest or most expensive frontier model is not automatically the best tool. The tradeoff is human agency. AI can accelerate skilled work, but outsourcing writing or using it to bypass learning can weaken the thinking those activities develop. So I’m optimistic about AI as leverage, while insisting that people retain both control of the tools and responsibility for their own minds.

Overall, I expect AI to be strongly beneficial: more people will be able to turn ideas into useful software, and skilled workers will gain extraordinary leverage. That can improve productivity and competitiveness without requiring every model to satisfy some grand definition of AGI. But the benefits are not automatic. If intelligence is concentrated behind subscriptions and provider policies, we trade broad capability for dependence on gatekeepers. And if people habitually outsource writing, judgment, or the difficult parts of learning, they may gain output while losing competence. So my optimism is conditional on preserving agency: affordable local models, hardware people can own, practical evaluation of tools, and deliberate limits on when we delegate thinking. AI should accelerate what people can do—not decide what they may do or replace the effort through which they learn.

The most decisive observation is how agents change the distance between a software idea and its execution. When I already understand the domain, I can direct the work, inspect the result, and iterate rapidly. That makes AI’s upside concrete rather than speculative: it expands what one person can actually build. The counterexample is equally important. When a provider refuses something as basic as translating political text, the tool stops being merely an assistant and becomes a gatekeeper over what users may read. That is why my optimism is tied to control. Intelligence should be available through open models and hardware people own, not exclusively through subscriptions whose policies can override the user.