GPT-5’s real advance may be less code quality than initiative

GPT-5’s most important change, based on this demonstration, is not that it writes code at all. It is that it appears to carry a vague request farther before asking for help. If that holds beyond one author’s workflow, the practical shift is for non-experts and small teams: less time translating intent into specs, more time reviewing what the model decided to build.

In the cited example, the author says a loose prompt produced a working 3D city builder within minutes, then kept expanding it with features that were never requested. That matters because the bottleneck in many AI tasks is no longer only generation quality; it is initiative. A model that can choose next steps, add plausible components and recover from errors changes the human role from prompt-writer to supervisor.

But the evidence here is narrow. This is a first-person product demo, not a benchmark, not a comparative test and not a code audit. The author reports occasional bugs, and also reports that pasted error text usually fixed them. That suggests a useful but fragile workflow: the system may accelerate prototyping while still depending on human review for correctness, architecture and hidden failures.

The strongest alternative explanation is that the result reflects good prompting, a forgiving toy project and selective reporting of a successful run. Those factors do not negate the demo; they do limit what can be concluded. What would change the assessment is repeated testing across users, harder projects and objective measures of failure rate, recovery time and unwanted feature drift.

For now, the responsible reading is modest: GPT-5 may make some coding tasks feel less like command execution and more like delegated drafting. That is valuable, but it also raises the cost of oversight, because a model that “just does stuff” can just as easily do the wrong stuff faster.

Source: https://www.oneusefulthing.org/p/gpt-5-it-just-does-stuff