GPT-5: It Just Does Stuff
One Useful Thing Ethan Mollick ● Covered by 4 sources
OpenAI's GPT-5 doesn't just answer prompts, it decides on its own how hard to think and what extra stuff to build. Give it a vague ask and it runs wild adding features nobody requested.
Based on reporting by One Useful Thing, Ethan Mollick — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Ethan Mollick, who's had early access to GPT-5, says the headline isn't raw benchmark scores. It's that the model has stopped waiting for detailed instructions. Ask for something dramatic, get a paragraph whose first words secretly spell "This is a Big Deal" while each sentence grows exactly one word longer than the last. Ask for an otter on a laptop as an SVG, and depending on some invisible internal judgment call, you get a lazy stick figure or a genuinely charming scene with a steaming coffee cup and clouds drifting by. Nobody asked for the clouds.
That invisible judgment call is the real story. GPT-5 isn't a single model, it's a router that picks between several GPT-5 variants of different size and reasoning depth, then decides for you how much effort to spend. Mollick says this fixes a problem that's kept millions of ChatGPT users stuck seeing only mediocre output: they never left the free, fast, dumb default model, so they never saw what a proper reasoning pass looks like. Now the system chooses for them, sometimes generously, sometimes not. Ask the same otter question repeatedly and roughly a third of the time it gets bumped to a slower, better answer, with no clear logic other than that phrases like "think hard" seem to nudge it upward. Paying subscribers can bypass this by picking GPT-5 Thinking directly, which is the version Mollick says he actually trusts.
The more striking demonstration is agentic sprawl. Mollick gave GPT-5 one scrappy sentence: build a procedural brutalist city generator he could drag and edit. No spec, no follow-up guidance beyond repeating "make it better." Minutes later he had a working 3D app, and then GPT-5 kept volunteering things he never requested — neon lighting, moving traffic, facade editors, cinematic camera sweeps, a save system. He never once looked at the underlying code. When bugs appeared, he just pasted the error message back in and the model fixed it without spiraling into the failure loops that have plagued earlier vibe-coding attempts.
Same pattern showed up when he asked it to generate ten startup ideas for a business professor, pick the best one, and execute a plan. It didn't stop at the idea. It produced landing page copy, LinkedIn posts, rough financials, and asked permission before charging ahead on tasks it decided were worth doing anyway. Mollick, who's taught entrepreneurship, calls it MBA-team-quality output from a single prompt.
Google's Gemini 2.5 Deep Think quietly grabbed an IMO gold medal last week and barely made a dent in public attention, because most people don't have a backlog of hard problems sitting around to test it on. GPT-5's pitch is different: it's useful specifically because you don't need such a backlog. You can gesture vaguely and it fills in the intent, the plan, and the execution. Humans still have to check its work — hallucinations haven't vanished — but the balance of who's driving the interaction is visibly shifting toward the model.
My take — AI-written commentary, not fact-checked reporting
I keep hearing "it just does stuff" treated as pure convenience, and sure, it is — but convenience that removes the need to look at your own code, or even define your own request clearly, is also the first step toward nobody knowing how anything got built. That's not doom-mongering, it's just how autonomy creep always starts: quietly, with clouds you didn't ask for drifting past a building you didn't design.
Read more about this at: One Useful Thing