LLMs That Write Well: what YC wants and how to validate it
Original request by Jared Friedman, Y Combinator. Source: YC Requests for Startups.
The request in plain English
As large language models got better at coding and math, their writing got worse: flatter, more robotic, and more uniform. YC thinks the big labs will keep optimizing for code and agents, which leaves room for a startup that builds a model people actually enjoy reading and talking to.
What YC's request says
- YC's view is that current models sound more robotic than the 2024-2025 generation.
- The argument: reward signals that improve coding also flatten voice and style.
- Frontier labs are expected to keep prioritizing coding and agentic workflows.
Why now
Training for verifiable tasks rewards correctness, structure, and caution. Those are good traits in code and bad traits in prose. Many people now recognize model-written text on sight, which is a sign the style has converged on something generic.
That creates a quality gap in a huge market. Writing is part of nearly every job and many consumer products, and a model with genuine voice would stand out immediately against defaults that all sound the same.
The hard part
Good writing is hard to measure, which is exactly why big labs under-invest in it. A startup here needs an evaluation method that captures taste: blind preference tests with real readers, expert editors grading outputs, or retention in a consumer product where people choose to keep reading.
The other hard part is data. Training a distinctive voice needs high-quality writing with clear rights, and a feedback loop from readers that keeps improving it.
Three startup ideas to start from
A writing-first model for one genre
Fiction, long-form journalism, or speeches: own one genre where quality is obvious to readers and the market pays for it.
Taste evaluation as a product
Sell blind human preference testing for writing quality to labs and companies that ship AI-written text.
Companion or tutor with a real voice
A consumer product whose main advantage is that talking to it is enjoyable, measured by how long people stay.
How to validate it before you build
- Run blind A/B tests: your outputs against a frontier model's, rated by readers who do not know which is which.
- Pick a buyer who loses money on flat writing today, such as content teams or publishers, and measure whether your output changes their results.
- Check that you can source training data with clear rights before committing to a training run.
- Track whether people return to your product voluntarily. Enjoyment shows up in retention.
The founder trap
Competing with frontier labs on general capability. The opening is narrow: be clearly better at writing that people choose, and do not try to match the labs everywhere else.
IdeaCheck's take
A real gap with a measurement problem. Founders who solve evaluation first will know whether they are winning, which most writing-AI startups cannot say.
Test your version
Check your idea against Hacker News
Building something in ai models? See what 28,000 Hacker News threads say before you commit.
Your first evaluation is free with Google sign-in.