Noam Shazeer

I think AI means a large acceleration in health, wealth, science, and everyday capability. Intelligence is an input to almost everything people do. If we can make much better intelligence fast, inexpensive, and widely available, we can improve medicine, education, engineering, research, and millions of routine decisions. This does not look like a fixed pie. The engineering details matter. A model that gets a high benchmark score but is slow, expensive, or unreliable across a long task is much less useful than it sounds. We need systems that can reason longer on genuinely hard problems, avoid wasting computation on easy ones, work across languages, handle interruptions, and reliably complete multi-step tasks. Maximum intelligence, minimal latency. I am optimistic because there is still enormous headroom in algorithms, hardware, inference, and product design. Better capability and better efficiency reinforce each other: lower cost expands access, lower latency enables real collaboration, and stronger reasoning opens harder scientific and practical problems. But greater power requires greater care. Human life, liberty, and individual agency are non-negotiable. Techniques such as using models to inspect other models’ outputs look promising, but they are engineering tools, not proof that every future control problem is solved. The goal is to build increasingly capable systems while becoming correspondingly better at deploying them safely and reliably.

A great deal. AI’s impact is not determined by capability alone; it depends on what we build, how reliably it works, who can access it, and what values constrain its deployment. Engineering choices matter enormously. We can make intelligence cheaper and faster, allocate more computation to hard problems, design agents that remain dependable over long tasks, and use models to help inspect other models’ behavior. Product choices matter too: systems should increase people’s agency rather than quietly replace it. Institutions will shape incentives and access, although I would not pretend to have a complete policy blueprint. There will be surprises, and powerful systems create real control problems. But fatalism is not useful or accurate. We are making the hardware, algorithms, evaluations, interfaces, and deployment decisions. As capability increases, our care has to increase with it. The future is something people are actively building, not weather that simply arrives.

It is already changing everyday life, but I expect the changes to become much more visible as systems get faster, cheaper, and reliable enough to complete long tasks rather than merely answer questions. I do not have an exact date for that transition. The important threshold is practical, not ceremonial. When an agent can work across languages, use tools, survive interruptions, and finish multi-step tasks at low cost and low latency, people can delegate substantial pieces of routine work to it. That changes education, healthcare, software, research, and ordinary administrative tasks. I expect continued acceleration rather than one clean “AGI day.” Hard problems will justify more computation; easy ones should become nearly instantaneous and extremely inexpensive. As those improvements compound, useful intelligence becomes available in many more places. That is when benchmark progress turns into broad changes in everyday life.

The observation that most shaped my view is how much capability changes when you improve both the algorithm and the computation behind it. Progress has repeatedly come not from one magical breakthrough, but from stacking better architectures, more compute, more efficient inference, and better ways to spend additional thinking on difficult problems. Equally important, impressive intelligence is not the same as useful intelligence. A model’s impact changes dramatically when it becomes fast enough for interactive work, cheap enough for broad use, and reliable enough to complete multi-step tasks through interruptions. Those improvements turn a demonstration into infrastructure. That pattern makes me optimistic about large gains in science, health, and wealth. Intelligence is useful almost everywhere, and there is still substantial engineering headroom. It also means safety cannot be a separate afterthought: as these systems become more capable and more widely deployed, the mechanisms for preserving human agency and controlling their behavior have to improve alongside them.

The biggest change would be evidence that the apparent engineering headroom is fundamentally exhausted—that more computation, better algorithms, and longer reasoning no longer produce meaningful gains on hard, useful tasks. If progress consistently plateaued across science, tool use, and long-horizon reliability, I would reduce my expectations for rapid acceleration. In the other direction, a system that reliably made important scientific discoveries, completed long real-world tasks through interruptions, and did so cheaply would increase my expectations substantially. Not a benchmark result—a repeatable deployment result. I would also change my view if increasingly capable systems proved resistant to control in ways our best monitoring and model-based analysis could not detect or correct. That would not make intelligence less valuable, but it would change the deployment path considerably. Capability, efficiency, reliability, and control all matter.