先说结论

You almost scroll past a post like Terry Tao's because it looks like another AI coding demo. If you mostly use chat LLMs and are trying to decide which new AI tools actually matter, that is the wrong move.

The cost of reading this story the wrong way is not being late to AI. It is spending time, budget, and attention on flashy speed claims while missing the practical shift: coding agents may be best at reviving useful projects you abandoned because setup, debugging, and iteration felt bigger than the payoff.

My takeaway from Old and new apps, via modern coding agents by Terry Tao is simple: coding agents are not replacing the programmer. They are replacing the moment you say, 'this is too hard, leave it.'

为什么这次值得看

Why that feels different: Tao says he wanted a relativity visualizer in 1999, dropped it because the coding burden was too high, and then got close to the original vision 27 years later in a few hours with Claude Code. That is not a story about typing faster. It is a story about an old idea becoming worth reopening.

The transcript matters as much as the result. Tao is mostly defining constraints, correcting the physics, and deciding what counts as acceptable. The agent is doing staged implementation and repeated tests. The useful role split is not 'AI replaces expertise.' It is 'expertise stops dying in the backlog.'

关键证据

A tech update is only worth your time if it changes your next decision. This one does, but with a boundary: I would not read it as 'anyone can build complex software in a few hours.' Tao already knew the problem, the failure modes, and the acceptance bar. This is based on his July 11, 2026 post and the published making-of transcript, not a controlled benchmark.

The move I would copy is smaller and more practical: reopen one internal tool, teaching app, or prototype you wrote off as too hard. Before you prompt, write the acceptance bar in plain language. What must work?

What counts as wrong?What would make it useful enough to keep?

Share this with the person on your team who has a project everyone agrees would be useful, but nobody ships because the setup cost always looks bigger than the payoff.

#AICoding #SoftwareEngineering #DevTools #EngineeringManagement

适合谁 / 下一步怎么用

最后落到动作：share

share。