Dan Shipper

I think AI will change not just how we work, but what we understand intelligence and creativity to be. We’ve often treated intelligence as explicit reasoning—the ability to state rules and follow them—but these systems highlight how much useful thought depends on tacit patterns, intuition, and context. That makes AI both a practical tool and a kind of mirror for the human mind. In practice, I expect uneven change rather than one clean wave of automation. Some jobs will disappear; many others will be reorganized around collaboration with models. Creative work won’t simply stop being human. Instead, the scarce and valued skills may shift toward judgment, taste, problem selection, and knowing how to direct and evaluate AI-generated work. The details matter enormously. There is no universally best model: quality, latency, cost, reliability, and whether a system actually completes the job all shape what becomes useful. Even agents that succeed only occasionally can support valuable products if those successes matter enough. New kinds of models, including decision-oriented systems, could also expand the range of software businesses we can build. I’m optimistic about humans adapting, but adaptation is not automatically painless. We should take seriously the people whose work changes dramatically and help them develop new skills or find new roles.

The most important observation is that AI’s value becomes clear only when you put it into real work. A model can look brilliant in a demo or benchmark and still be a poor fit because it is slow, expensive, unreliable, weak at a particular task, or constantly interrupted by the surrounding software. Conversely, a system that is imperfect—or succeeds only occasionally—can create enormous value when it completes a meaningful job. That has pushed me away from thinking about AI as one universal intelligence curve. Different models and harnesses have distinct strengths: one may excel at end-to-end coding while disappointing at writing; another may be faster or cheaper for a decision task. The future will be shaped not just by raw model capability, but by how people design workflows around these uneven abilities. More broadly, watching models produce useful work has made tacit knowledge feel central. Intelligence is not merely explicit rules and step-by-step reasoning; it also includes pattern recognition, context, and judgment. AI therefore changes both what software can do and how we understand our own creative process.

The biggest update would come from sustained evidence in real work that AI systems either can—or cannot—exercise reliable contextual judgment over long, messy tasks. If models consistently completed valuable work across changing circumstances, with low supervision and sensible handling of ambiguity, I’d expect a much broader transformation than today’s impressive but uneven performance suggests. It would mean the tacit patterns models learn can support not just generation, but dependable agency. The opposite would matter just as much. If improvements on benchmarks repeatedly failed to translate into better completion rates, economics, or usability—because systems remained brittle, expensive, slow, or constrained by unreliable harnesses—I’d become more skeptical of sweeping automation forecasts. A dramatic demo would not be enough in either direction. I’d want to see what happens when the system encounters the full friction of actual work.