The Problems Progress Creates

A railway can move people between cities faster than a horse. It also creates a problem a horse never did: what time is it at the other end of the line?

When North American railroads expanded, local clocks made schedules difficult to coordinate. In 1883, the railroads adopted standard time. A technology that solved one problem had created another, and solving that problem made the larger system more useful.^[1]

AI follows that pattern. What changes is how quickly we can move from a new capability, to the problems it creates, to the next attempt to solve them. That accelerating cycle is why learning and deciding at the speed of play matters.

The pattern extends beyond rail. Electricity did more than replace steam power. Factories had to redesign how machines and people were arranged before they captured much of its benefit. The internet made global communication cheap, then created new challenges in network scaling, complexity and security. Each advance expanded what people could do and what they had to coordinate. Those complementary changes help explain why the economic gains from major technologies can arrive well after the invention itself.^[2][3][4]

That is one way progress creates economic opportunity. New capabilities reveal problems that were previously invisible, tolerable or impossible to address. Someone builds the scheduling system, the safer network, the better way to organize a factory. A solution can become a business; the solution can also make the original technology more valuable.

But complexity alone creates no wealth. A problem may be too costly to solve. A solution may shift costs onto someone else. It may generate revenue while making people worse off. The opportunity lies in reducing the burden that complexity creates, at a cost lower than the benefit people actually receive.

AI is following this historical pattern. It can produce software, analysis, images and decisions at a scale that would once have required far more human effort. That abundance creates its own work: checking accuracy, tracing sources, protecting data, integrating tools and deciding which of a thousand plausible outputs deserves attention. As producing an answer gets cheaper, knowing whether to act on it can become more valuable. Stanford’s 2026 AI Index reports rapid adoption alongside substantial limits: agents improved sharply on computer-task benchmarks, yet still failed roughly one in three attempts on OSWorld, a structured computer-task benchmark.^[5][6]

AI also changes the pace of the pattern. A railway needed track; electrification needed physical equipment and factory redesign. AI still needs infrastructure and organizational change, but many of its applications can be copied, modified and tested through software. More unusually, AI can help improve the tools used to develop AI. Google DeepMind reports that AlphaEvolve found improvements to processes used in training its own underlying models. That is a real feedback loop, though it depends on goals, evaluators, computing resources and human decisions. It is evidence of AI-assisted improvement, not proof of an autonomous system improving without limits.^[7]

So two clocks are running. The capability clock can move quickly: a better model can make yesterday’s difficult prototype routine. The adoption clock still depends on trust, training, redesigned work and results. In the 2023 working-paper version of one study, AI assistance increased customer-support productivity by 14% on average. In a separate early-2025 experiment, experienced developers working in familiar codebases took 19% longer with the AI tools tested. Neither result describes every job or every model. Together, they show why access to a powerful tool is different from an improvement in outcomes.^[8][9][10]

The response is to shorten the time between noticing a problem and learning whether a solution works. Start with a real need. Build the smallest useful test. Measure the result, including the new burden the solution introduces. Improve it when people return to it; stop when the evidence says to stop.

That is the reason to live at the speed of play. Mark Pincus uses the phrase in his book of the same name; Life at the Speed of Play emphasizes testing ideas and learning quickly. As AI shortens parts of the cycle between capability, problem and solution, the advantage is not simply adopting everything faster. It is learning fast enough to recognize which new problems matter, solving them well, and staying able to change the solution as the technology changes again.^[11]

Historical examples and empirical findings are linked below. The argument connecting them is Thinkforge Labs’ interpretation.