LLMs That Write Well
LLMs have improved at coding and math but declined in writing quality, producing more robotic output than earlier models. This tradeoff stems from structural reward signals that prioritize technical tasks, creating a market opportunity for startups focused on conversational quality.
Summary
The speaker argues that recent LLM development represents a counterintuitive regression in writing capability despite overall model improvements. While LLMs have demonstrably gotten better at coding and mathematical tasks, they have simultaneously gotten worse at writing—a decline that is perceptible and measurable rather than imagined. The current generation of models produces noticeably more robotic and less engaging prose compared to models from 2024-2025. The speaker characterizes this as a structural issue rather than a temporary setback, indicating a fundamental misalignment in how these models are being optimized. The root cause is identified as the reward signals used during training: the same optimization objectives that make models excel at code and math actively work against developing natural, stylistically varied writing. Given that major AI labs face tradeoffs, they consistently prioritize coding performance over literary quality when forced to choose. This structural preference creates an underserved market opportunity for a startup or team that specifically prioritizes producing models optimized for engaging, natural-sounding conversation and writing quality.
Key Insights
- Current LLM models sound more robotic than 2024-2025 models despite overall improvements in coding and math capabilities
- The decline in writing quality is a structural issue rather than a temporary glitch
- The same reward signals that optimize models for coding actively flatten their writing style
- Major labs will prioritize coding performance over writing quality when forced to choose between them
- The gap between current models' writing quality and user expectations creates a business opportunity for startups focused on conversational quality
Topics
Transcript
[0:00] As LLM's have gotten better at coding and math, they've actually gotten worse at writing. You're not just imagining it. The current models really do sound more robotic than the models from 2024 and 2025. This appears to be not just a temporary glitch, but a structural issue. The same reward signals that make models good at code actively flatten their style. If forced to choose, the major labs will always focus on coding over great writing. This has left open an opportunity for a startup that cares about producing a model that people like [0:30] talking to. Maybe that's the opportunity for you.
Full transcript available for MurmurCast members
Sign Up to AccessMore from Y Combinator
We’ll Be Inferior to Our Grandchildren—and That’s Great
A discussion about scaling societies and the relative ratio of good versus bad behavior, with one participant expressing concern that future generations will be intellectually superior to current humans, similar to how humans are vastly superior to ants.
YC Debates: Should We Slow Down AI?
Jacob Coxin argues governments should substantially intervene to slow frontier AI development, citing rapid capability growth (superhuman AI researchers by 2027), alignment risks, and misaligned corporate incentives. Robin Hanson counters that typical technological adoption patterns suggest slower real-world impact, that regulation is dangerous without concrete problems, and that liability-based approaches are preferable to centralized global regulation.
How To Build The Future: Eric Glyman, Co-CEO of Ramp
Eric Glyman, co-CEO of Ramp, discusses how the company evolved from Parabus to build a $60 billion financial operations platform, and how Ramp is leveraging AI agents and internal tools like Inspect, Research, and Glass to transform organizational productivity and reshape the future of work.
Grok Bot: What Changes When You Can Build AI Teammates
Lauren Tan and Roman Ugarte discuss Grokbot, a platform for creating AI teammate agents with their own identities and tool access that can work autonomously. The conversation covers how AI agents are transforming software engineering and non-technical work, with Lauren showcasing her 4,000+ PRs/month workflow using specialized bots in Slack and describing the shift from execution-constrained to execution-abundant organizations.
Grok Bot: What Changes When You Can Build AI Teammates
Lauren Tan and Roman Ugarte discuss Grockbot, a product enabling users to create AI teammates with their own identities and capabilities. They share how AI agents are transforming software engineering and business operations, demonstrating workflows where engineers ship thousands of PRs monthly by leveraging specialized bots integrated into team communication tools.