← all notes
2026-09-01Naman Barkiya

AI Writes 95% of the Code Now. The Bottleneck Moved..

AI coding agents removed the typing bottleneck, not the decision bottleneck: 25% of YC's W25 batch had codebases that were 95%+ AI-generated, yet Andrew Ng says the constraint in AI startups moved to product management — deciding what to build, in what order, for whom. The fix isn't a faster coding tool, it's a one-page scope document before the agent starts.

A quarter of Y Combinator's W25 batch shipped codebases that are 95% AI-generated, and Andrew Ng says the bottleneck in AI startups isn't coding anymore — it's product management. What that shift actually changes about how to run an MVP build.

95% of the code in a quarter of Y Combinator's W25 batch was AI-generated, and Andrew Ng says the bottleneck in AI startups still isn't coding — it's product management. We agree with the second half more than the first: the code got fast. Deciding what to build didn't.

For two years, "we can code faster now" has been the headline. It's true, and it's also the wrong metric. Code was never the slow part of a good build — deciding what to build, and what to leave out, was. AI tools didn't remove that step. They just made it more visible, because now it's the only step left that a prompt can't do for you.

What did Andrew Ng actually say?

At a 2026 event covered widely and discussed on Hacker News, Ng said the bottleneck in AI-native startups has shifted from writing code to product management — deciding what to build, in what order, for whom (Hacker News, "Andrew Ng says bottleneck in AI startups isn't coding – it's product management," 120 points). The top reply on that thread put it bluntly: "code has docs. product fit does not." That's the whole argument in six words — code is a solved-enough problem that an AI can review it against a spec, but nothing reviews your spec against whether anyone wants the thing.

What does "95% AI-generated code" actually mean for the decision that's left?

Y Combinator's managing partner Jared Friedman said about a quarter of the W25 batch had codebases that were 95% or more AI-generated — and that these weren't non-technical founders coasting on a tool, but technical founders who, a year earlier, would have written that code by hand (TechCrunch, "A quarter of startups in YC's current cohort have codebases that are almost entirely AI-generated," March 2025). If 95% of the typing is gone and the founder is still the bottleneck on shipping, the 5% left — reviewing what the AI produced, deciding what it should produce next — is now effectively 100% of the job. Speeding up the 95% didn't touch it.

If code isn't the constraint, what is?

Three things, in the order we see them slow a build down:

How does this change how a founder should spend their time on an MVP?

Spend less time worrying about which AI coding tool is fastest — they're converging — and more time on the artifact that actually determines what gets built: a one-page scope with a named user, three to five features, and an explicit non-goals list. That's not a new idea; it's the same scope discipline that's always separated a fast MVP from a bloated one. What changed is the cost of skipping it. A vague brief used to cost you a few extra weeks of back-and-forth with a developer. Handed to an AI agent, a vague brief gets built literally and immediately — wrong, fast, and confidently. How to scope an MVP is the one-page version we use before any build starts, AI-assisted or not.

Does this mean the AI coding tools don't matter?

They matter for speed, not for whether the product is right. Bolt, Lovable, Cursor, and the rest compress the distance between a decision and a working screen — which is genuinely valuable, and genuinely useless if the decision behind it was wrong. A fast wrong answer is still wrong, just sooner and with more confidence behind it. The studios and founders who come out ahead in 2026 aren't the ones who found the fastest agent. They're the ones who spend the time they saved on typing back on the one decision an agent still can't make for them: what actually deserves to get built.

Code stopped being the bottleneck. The decision about what code to write next never stopped being one.

Written 2026-09-01 by Naman Barkiya.

FAQ

Questions this usually surfaces.

If AI writes most of the code now, what's actually left for a founder to do?
Deciding what gets built, and what explicitly doesn't. AI coding agents compress the distance between a decision and a working screen, but they build whatever they're told to — including the wrong thing — just as fast as the right one. The scope decision is the part that didn't get automated.
Does AI-generated code mean an MVP needs less product planning?
The opposite. A vague brief handed to a developer used to cost a few extra weeks of back-and-forth before the wrong thing got built. The same vague brief handed to an AI coding agent gets built literally and immediately — confidently, and often wrong — because there's no one in the loop to ask a clarifying question first.
Is 95% AI-generated code a sign of a weak technical founder?
No. YC's own data says otherwise: about a quarter of the W25 batch had codebases that were 95%+ AI-generated, and YC's Jared Friedman was explicit that these were technical founders who, a year earlier, would have written that code by hand. The skill that matters now is knowing when the output is right, not typing it yourself.