September 30, 2025 · JustSayAI

How Good Is AI at Coding, Really?

The ultimate question of AI coding: who takes the fall? (Disclosure: I’m not a programmer.)

Let me rip the band-aid off up front: programmers exist to be scapegoats. Shipping a new feature is easy; untangling a hairy bug is hell. When things blow up, whoever can fix it stays—otherwise, pack your desk. That’s the cold logic of business.

Yet the breathless AI-coding hype races down two parallel tracks that both dodge this ugly truth. Track one: “interactive” IDEs like Cursor and Windsurf—copilots glued to a human’s daily flow. Track two: “asynchronous” agents like Jules and Codex—overnight contractors you task before bed and wake up to merged PRs. Both promise salvation, both land in the same awkward middle: too clumsy for the 10× wizard, too clueless for the total newbie staring at a real-world hairball.

Track 1: IDE “human-AI co-drive”—UX is everything

Here the fight is over interaction philosophy. Cursor, beloved by pros, versus Windsurf, tuned for enterprise.

Cursor’s stardom isn’t some algorithmic leap—it just lucked into riding Claude Sonnet’s shoulders. Its moat: building a standalone editor from scratch instead of a plugin. Coding is immersive; a single stutter kills flow. Owning the stack let Cursor sweat the tiny stuff—like the git-merge view that sold me. It thinks like a dev, diffing and blending AI code instead of carpet-bombing your repo.

Windsurf took the pragmatic road: B2B first, legacy-code friendly. It flipped early to “AI drives, human rides shotgun” agent mode—perfect for companies that dread support tickets and blame ping-pong. That’s why OpenAI reportedly coughed up $3 billion: not just a sales channel, but oceans of process data—the messy loops of tweak, debug, revert—that static GitHub repos never show. That telemetry is premium fuel for next-gen models.

Track 2: Async agents and the “full-service” fantasy

Google’s Jules, OpenAI’s Codex, Anthropic’s Claude Code play a three-way game of asynchronous Thrones. Pitch: leave a ticket at dusk, find working code at dawn.

Jules is the poster child—Google-scale cloud, full-service wraparound. The dream of every cost-cutting founder: “Code is deterministic, devs are expensive. I’ll force everyone to use it—refusers can quit.”

But today it’s still concept-car territory. The hardest part of software isn’t writing code; it’s defining the problem. How many bosses stay as focused as Xuanzang marching west? Most detour into the KTV halfway. When requirements themselves are foggy, an overnight agent just hallucinates in the dark.

The shared wall: three human mountains no one climbs

Whether they live inside your IDE or run while you sleep, all tools smash into the same wall: the user. Lower apparent门槛 just hides a higher real one—good output demands better input.

To replace programmers, AI must clear three peaks:

Describable: Can you state the world (your spec) without ambiguity? Wittgenstein would laugh; by the time you can, you’ve practically written the code.
Decomposable: Hit proprietary knowledge or messy meat-space workflows and the agent stalls—no Stack Overflow to parrot.
Evaluable: Probabilistic models can’t guarantee v2 beats v1. Who signs off?

All three are people problems. You must be coach, referee, and grizzled sensei—know what you want, where the cliffs are, what “right” looks like. For those who “don’t know what they don’t know,” that’s Everest.

AI can hand you the Nine Suns Manual, but you still have to squat

Current reality: you’re given the sacred scroll, think you’re invincible, and the master orders three months of horse-stance squats. The flashy demo hides the grunt work—architecture, env setup, maintainability tests—that turns code into production steel.

The automation wave is unstoppable, but it won’t drown great devs tomorrow. It will flood the low-lying plains first: average front-end tasks that are repetitive and blame-light. AI mercilessly raises the floor—shipping “okay” products gets trivial, making “great” ones harder than ever.

Because in the end, AI can write code; it just can’t take the blame.

Viewpoints adapted from my podcast: People’s Park Talks AI

👉 Subscribe to “JustSayAI Daily Brief” · twice a day · one-click listen: https://justsayai.org/newsletter (VPN required)