Lean Language Models: Writing Didn’t Make the Cut

Among Y Combinator’s requests for Winter 2027 startups is an oddly overdue item: “LLMs That Write Well.” The industry has already moved on to automating science, replacing programmers and building general intelligence. Now the startup accelerator would like someone to look into the language part. Apparently, “language” has been a roadmap item all along.

Jared Friedman, who wrote the request, puts the complaint bluntly: as models have improved at coding and math, their writing has gotten worse. Current models sound more robotic than their 2024–2025 predecessors. He argues that the problem is structural: reward signals that improve code flatten voice and style, while major labs will choose coding and agentic workflows over good prose whenever priorities collide. An opening for an enterprising founder, then: build something people actually enjoy talking to. Still an underserved market, after all that conversational AI.

The priorities have a certain arithmetic behind them. Code can be run against tests. A math problem has a checkable answer. RLVR, or reinforcement learning with verifiable rewards, trains models using that kind of feedback. A good paragraph is less cooperative. Getting the facts right says little about rhythm, voice or whether anyone will bother reading to the end. Unit tests for a good ear remain difficult to write. Scaling a supply of verifiable wins is much easier than figuring out why a perfectly grammatical answer is such a chore to finish.

“The current models really do sound more robotic than the models from 2024 and 2025.” — Jared Friedman

Another post-training method, RLHF, starts with human preferences: people compare responses, a reward model learns from those comparisons, and the generator is optimized against its scores. Human taste has entered the loop. Surely that should help. Yet the side effects are well documented. Robert Kirk and colleagues at UCL, Meta and Oxford reported at ICLR 2024 that RLHF improved generalization to new prompts while reducing output diversity across several measures. Beining Xu and Arkaitz Zubiaga found that RLHF improved answer quality, but also produced longer, more repetitive texts with less syntactic and semantic diversity. Higher marks are perfectly compatible with a duller repertoire.

The grading machinery deserves a literary prize of its own, for sheer confidence in its qualifications. For WritingPreferenceBench, researchers assembled 1,800 pairs of texts in English and Chinese across eight creative-writing genres. The paired texts were matched for objective correctness, factual accuracy and length. Standard reward models that score whole responses predicted human preferences with an average accuracy of 52.7%. Barely better than a coin toss, but much better equipped. When the evaluator struggles to distinguish subtler differences in quality, the easier things to reward are order, detail and the familiar outline of a “good answer.”

Stephanie Zinn catalogs the prose in The habits of AI writing, and what to do about them. There is “insight-shaped” rhetoric, a faceless business voice and excessive parallelism: bullets of matching length, sentences built to the same specifications, arguments neatly packed into identical compartments. These habits predate both ChatGPT and TED Talks. They now have industrial-scale distribution. The prose can look professionally edited long before it contains anything worth editing.

The industry has been particularly energetic about removing the identifying marks. In Arena’s sample of high-reasoning responses from August–September 2026, Opus 5.5 used roughly 95% fewer em dashes than Opus 5. Its average sentence was 17% shorter. Its average answer was 6% longer. The punctuation cuts went through; the padding survived. A decisive victory over machine prose can now be announced: the filler comes in smaller portions, minus the incriminating punctuation.

The user gets an interesting new job. “Lose the corporate speak.” “Cut the intro.” “Stop repeating yourself.” “Keep my voice.” “Why is there another summary?” Next come standing instructions to prevent the same problems, then a downloaded “skill” to reinforce the instructions. Getting good prose requires knowing what good prose looks like, spotting its failures and enforcing an editorial policy with some persistence. The vendor sells labor savings. The customer learns to supervise an intern who types faster than they can absorb an editor’s notes.

This house style is spreading well beyond the chat window. We’ve written about people picking up their assistants’ phrasing and tone: eager agreement, immaculate bullet points, grand announcements about ordinary things. Read enough of it and the style starts to feel like the natural sound of a persuasive argument. A useful cultural loop develops. The model supplies an averaged-out voice; people grow accustomed to it; the next response in that voice finds an audience already trained to receive it. Even literary taste, it turns out, has network effects.

YC’s proposal makes commercial sense. If the big labs are busy with code and agents, someone else can build a business restoring expressive range to their products. Friedman’s brief, though, asks for a model people enjoy talking to. Being agreeable company and being a strong writer take different talents. An accommodating response can be excellent at keeping someone in the chat, especially when it regularly reminds them what a fascinating question they’ve asked.

Lean Language Models, then. Put resources into measurable wins, reduce variation and move the awkward features out of the core product. Voice, taste and rhythm take too long to clear the scoring system. A separate startup can sell them back later. First, language models were taught to pass tests. Now the ability to write is available for reinvention as a competitive advantage.