Paul Graham: Why LLMs Crush Math but Lag at Writing

Paul Graham’s one-liner is a training-data economics tweet dressed as a culture take.

On August 3, 2026, @paulg wrote that he was curious why LLMs had gotten so good at math while still lagging at writing — so he asked an expert. The answer: not because math is easier, but because math has clear right and wrong answers and is therefore easier to train on. Punchline: “But they are coming for me next.”

That lands next to Graham’s own essays on good writing and writes vs write-nots, his May 2026 refusal to finish AI-written founder emails, and explainx.ai’s earlier Fable → 2031 speculation post. The useful builder read is about rewards — and, per @goyashy’s reply below, about what gets into the corpus when the domain has no narrow grader.

TL;DR

What Graham actually said

Quoted in full (paraphrase risk is high on viral one-posts):

I was curious why LLMs have gotten so good at math and still aren't that good at writing, so I asked an expert. It's not because math is easier. It's because it has clear right and wrong answers and so is easier to train on. But they are coming for me next.

Key moves in that paragraph:

1. Rejects the folk theory that “math is easier than writing.”
2. Names the real asymmetry: outcome checkability.
3. Predicts convergence — writing is next on the capability curve, not forever safe.

He does not name the expert publicly in the post. Treat the explanation as Graham’s reported synthesis, not a cited paper — still consistent with how labs talk about verifiable rewards for math and code.

Why “right and wrong” accelerates models

Modern capability jumps on math and coding are not only “more pretraining tokens about algebra.” They are closed-loop improvement:

If you can say pass/fail without a committee, you can:

• Run best-of-N and keep winners.
• Train outcome RL against graders.
• Let models critique and revise against a checker.
• Scale synthetic curricula where wrong answers are unambiguous.

Writing lacks that. Two essays can both be “correct” English and one still be dead on the page. Preference models help, but they approximate average taste, not Graham’s sanded-bin sense of good writing — where sounding good and getting the ideas right are entangled because the writer is the first reader.

That is also why biology agents lag coding agents in explainx.ai’s VirBench / agent biology coverage: fewer simple, verifiable rewards comparable to tests pass.

Math is not “easy” — and writing is not “safe”

Math side: FrontierMath-class and research proofs still break models. A correct boxed answer can hide a rotten proof. Verifiability explains relative progress vs prose; it does not mean AIME is trivial or that labs have solved mathematical research.

Writing side: Models already draft usable emails, docs, and marketing. What they still fail at — for Graham’s audience — is voice, trust, and thinking-on-the-page. His May 2026 stance is revealing: he objects less to “AI can’t write” than to AI writing signed as a human feeling like a lie. Capability and social acceptability are different curves.

So “coming for me next” can mean two things at once:

1. Capability: RL and better critics will raise prose quality.
2. Personal stake: Graham’s comparative advantage as an essayist gets contested — even if he never accepts AI-signed email.

Replies that sharpen the axis

Useful quote-tweet patterns (not independent research):

• Design has no single right answer → same training bottleneck as writing.
• Pointers to older Thiel / Cowen conversations about AI and what is hard to automate when evaluation is political or aesthetic.
• “Brute-force where variables are quantifiable” — folk RL intuition that matches lab practice.

The unifying axis is reward clarity, not STEM vs humanities branding. Chess, leetcode, and AIME share a property with unit tests. Essays, logos, and product taste share the opposite property.

@goyashy: the slop feedback loop Graham’s frame misses

explainx.ai founder Yash Thakker (@goyashy) replied to Graham with a second failure mode — not “no right answer,” but contaminated corpora:

I would say they're relied on a lot of bad data, for math this might not be the issue since the target audience is narrower. Imagine models re-crawling ai slop and then newer versions re-produce all of that. All writing will likely become ai slop at some point. I start using anti-slop skill to write content = this content gets crawled by ai = becomes ai slop.

Three claims worth separating:

Graham explains why RL with clear labels favors math. Yash adds why pretraining and continual crawl punish writing: the web is filling with model-shaped English, so each generation’s “improvement” can be style collapse toward the mean of prior generations — including pages written with anti-slop design/content skills that were meant to fight that mean. See also explainx.ai’s what is AI slop and slopocalypse threads.

That tightens “they are coming for me next.” Models may get better at producing fluent essay-shaped text while the distribution of public writing gets worse — which is exactly the world where Graham’s ear for AI email still works, and where human essays become scarcer training gold.

For publishers (including this site): treat “we used an anti-slop skill” as temporary differentiation, not a durable moat. Once that voice is public and crawled, it is training data. The durable moves are grounded claims, primary sources, and taste that still needs a human first reader — closer to Graham’s own good writing loop than to another prompt pack.

What this means if you ship products

1. Prefer verifiable agent loops

Where you can define pass/fail, do it. explainx.ai’s loop engineering and agent loops catalog exist because exit criteria beat vibes. Math tutors, code agents, data transforms, schema migrations — these are Graham’s “math” category even when no integral appears.

2. Invent writing proxies — don’t wait for a perfect grader

Practical proxies teams use:

None of these equals “clear right and wrong,” but they move writing toward partially verifiable — the same direction labs will push.

3. Separate “good enough prose” from “signed human prose”

Graham’s email rule is a product design lesson: disclosure and authorship norms matter when the reader is the evaluation function. Your users may accept AI drafts for internal notes and reject them for founder-to-investor first contact. Build modes accordingly.

4. Watch cognitive surrender

When writing gets “good enough,” people stop thinking through writing — Graham’s write-nots warning. Pair with explainx.ai’s cognitive surrender / “I don’t know” study: verifiable domains may increase reliance without the same epistemic humility costs as soft domains — or the opposite, if soft domains feel authoritative. Measure both.

Builder checklist (PG math vs writing)
□ Label each workflow: verifiable / preference / hybrid
□ Put automated checks on every verifiable path
□ For writing: rubric + critic + human gate for high trust
□ Assume public web prose is partly AI-contaminated (goyashy loop)
□ Don’t treat anti-slop skills as uncrawlable moats
□ Don’t ship AI-signed outreach without disclosure policy
□ Expect math/code agents to keep widening the gap until writing rewards *and* data quality improve
□ Re-read Good Writing + Writes and Write-Nots before declaring “AI essays solved”

How this fits Graham’s 2026 AI thread

Same person can cheer AI for founders who use it correctly and still refuse AI-signed intimacy. The Aug 3 tweet is the capability half; the email rule is the norm half.

Honest limitations

• Graham’s “expert” is unnamed — we have his paraphrase, not a paper.
• “Not that good at writing” is audience-relative (YC essays ≠ marketing copy).
• Math progress still has proof validity gaps.
• Preference learning does improve writing; the claim is relative difficulty, not zero progress.
• “Coming for me next” is forecast, not a ship date.
• Viral replies (design, Thiel, @goyashy slop loop) are pointers, not settled theory.
• Anti-slop → slop is a plausible feedback story; filtration and synthetic-data pipelines may blunt it for some labs.

Closing

Graham’s expert answer is the cleanest public compression of 2024–2026 training reality: models race where graders are cheap. Math and code got the cheap graders first. Writing still pays in human taste and trust — which is why prose feels behind, and why Graham can joke that they are coming for essayists next without admitting essays are “easy.”

@goyashy’s addendum is the data-side twin: writing also races through a crawl that increasingly is its own output, so “better writing models” can mean more confident slop unless humans keep injecting scarce, checkable judgment. Anti-slop skills help this draft; they do not permanently clean the internet.

Build like both are true: put checks on what you can check, treat the writing corpus as contaminated by default, and keep a human first reader wherever voice still matters.

Follow @explainx_ai when the next PG one-liner lands — and when writing RL papers start sounding like math papers.

Related on explainx.ai

• Thariq — Jevons paradox in math / rising demand
• Paul Graham — Fable / GPT-3 five-year speculation
• Loop engineering for coding agents
• Top AI agent loops with verifiable stops
• AI advice & cognitive surrender study
• Anthropic biology agents — weak verifiable rewards
• AI benchmarks complete guide
• Scalable oversight / RLHF
• Calacanis vs Musk — open vs frontier gap
• What is AI slop? SEO/GEO content quality
• Slopocalypse — AI slop and the internet
• Nutlope Hallmark — anti-AI-slop design skill

Sources

• Paul Graham on X — math vs writing / verifiable answers (Aug 3, 2026)
• Yash Thakker (@goyashy) — reply on AI slop feedback loops (Aug 3, 2026)
• Good Writing
• Writes and Write-Nots
• Simon Willison — Graham on AI emails (May 26, 2026)
• Business Insider — Don’t use AI when emailing Paul Graham

Interpretation of Paul Graham’s August 3, 2026 X post and related essays. The unnamed expert’s full argument is not public; training claims reflect widely reported lab practice around verifiable rewards, not a specific undisclosed paper.