← Back to the main Geek Mode article
What was tested: an 80-section specification for a personal task/calendar/tracking application, given to Gemini 3.8 Flash (Flash Extended) and GPT-5.6 Sol (Extra High), asking first how much work a conventional implementation would require and then how much AI coding tools would reduce that. Outcome: both models dramatically overestimated the calendar time an actual AI-agent-heavy, owner-directed workflow would take.
The application handles tasks, recurrence, calendar items, reminders, push notifications, food and nutrition records, medication data, sleep, long-form logs, incoming email, AI capture, rules, audit history, attachments, reports, shipment information, expected future events and related automation, under an architectural rule that AI interprets while deterministic server-side PHP owns authorization, validation, calculations, deduplication and database mutations.
The estimates
Gemini’s AI-assisted estimate landed around 300 to 450 engineering hours, roughly ten to fourteen weeks at the schedule it proposed. ChatGPT was even more conservative in calendar terms, estimating that a daily-driver version might take months and the full specification could take roughly a year or more for one person directing AI heavily.
What actually happened
The application was already being built at the time of the comparison. Roughly ten or eleven days into the project, around 95 percent of the code was already written by rough estimate, 60 to 70 percent already deployed and functioning and useful, with most of what mattered expected in production by around the two-week point.
The biggest bottleneck wasn’t the ability to write PHP — it was running out of AI. Five-hour usage windows and weekly limits on inexpensive consumer subscriptions were the meaningful constraint; spending substantially more on the highest usage tiers likely would have moved the project even faster.
These are firsthand estimates and progress reports, not audited engineering metrics. “95 percent written” is not the same as “perfectly hardened forever” — the last few percent of a complex system can expose recurrence bugs, races, strange email formats, migration problems, notification failures and things no design document predicted. But the two models’ estimates were still dramatically disconnected from the rate at which usable software was actually appearing.
The likely reason: both models started from conventional software-development estimates and applied an AI productivity discount on top, rather than modeling a workflow with no product manager, no ticket-writing meeting, no QA department and no two-week sprint before seeing whether a button feels wrong on a phone. Writing code is getting cheap; keeping the entire system straight in one head is not. The valuable human work is increasingly deciding what should happen, understanding how one part affects another, anticipating what a retry or failure means, maintaining architectural boundaries, and giving the AI enough precise context that it doesn’t have to invent the product while implementing it.
← Previous: plugin audit and repair · Back to the main Geek Mode article · Next: Anker charger documentation →