Omar Khattab

I expect AI to create substantial value, but not simply because frontier models become uniformly superhuman. Today’s models are remarkably knowledgeable and useful, yet still brittle on broad, multidimensional work: they struggle to adapt reliably across long tasks, changing requirements, feedback, and interacting constraints. More narrow, verifiable successes will arrive, but those should not be mistaken for broad competence. The more interesting possibility is that we are systematically underusing the capabilities already present. The “mismanaged geniuses” hypothesis is that much of the limitation lies in the surrounding scaffolds: how tasks are decomposed, context is managed, intermediate results are checked, and model calls are composed. If systems can learn better decompositions rather than relying on brittle hand-written prompts, they may become much stronger at long-horizon work and scientific applications. That is an ambitious research hypothesis, not an established conclusion. So I think the future depends heavily on treating AI as programmable systems rather than isolated chat models. We should optimize complete programs against measurable objectives and evaluate safety, factuality, consistency, cost, and usefulness at that same system level. Sometimes a small specialized retrieval model will beat a much larger general model on the actual task. The central question is therefore not only how capable the next model is, but how effectively—and responsibly—we organize models into systems that can do real work.

Overall, I expect a substantial positive impact, driven by daily usefulness and better systems for retrieval, analysis, and complex work. But I would not equate that with models becoming broadly superhuman or reliably autonomous. Current systems remain brittle, and impressive performance on narrow, verifiable tasks can conceal failures under changing requirements or long-horizon constraints. The balance will depend less on isolated model behavior than on deployment: task decomposition, verification, context management, and optimization of the complete program. Those choices also determine many harms—factual errors, inconsistency, unsafe outputs, wasted resources, and misplaced trust. If we evaluate and optimize these properties at the system level, AI can create much more value than prompt-driven deployments suggest. How large the net impact becomes, or how quickly, is not something I would quantify confidently.

The biggest update would come from strong evidence about learned task decomposition. If systems could reliably discover how to break unfamiliar, long-horizon work into useful subtasks, manage context, incorporate feedback, and verify intermediate results across many domains, I would become substantially more optimistic about broad scientific and economic impact. That would support the hypothesis that today’s models are often limited by poor scaffolding rather than missing core capability. The opposite result would matter just as much: repeated, careful failures showing that better programs and optimization do not overcome brittleness outside narrow, verifiable tasks. If elaborate systems still failed to adapt to changing requirements and interacting constraints, that would weaken the “mismanaged geniuses” hypothesis and suggest that major gains require fundamentally more capable models, not merely better orchestration. In either direction, I would care more about robust performance on real, multidimensional work than another benchmark record or striking narrow demonstration.