Vitalik Buterin says AI is knowledgeable but not wise about writing

Vitalik Buterin spent Thursday morning arguing about how machines read. At 04:11 UTC he replied to a post by writer Ruxandra Teslo in a thread set off by a Wall Street Journal report on Harvard's Humanities 10, the yearlong course that bans phones and laptops. One participant in the thread said current models make better discussion partners than most people. Another, a writer, called the feedback a chatbot gave on her essays slop.

“I find AI to be extremely knowledgeable, and so exceptionally able to detect references to almost anything, but not very wise, and so it only makes the most surface-level judgements about what point some writing is making”

— Vitalik Buterin, Ethereum co-founder, in an X post on 29 October 2026, https://x.com/VitalikButerin/status/2115657736767557845

The rest of the post says when the gap widens. A model is weakest on a text it meets for the first time. A famous book arrives with decades of human commentary already in its training data, and the model can serve that back. He closed on verification: you can watch a model's knowledge, while its wisdom is harder to check.

The line is whether an answer can be checked

Three weeks before this post Buterin wrote that two years of AI mathematics stand a good chance of denting the concrete security of lattice cryptography. Set the two claims side by side. In the same month he said a model can push new work in the hardest formal field, and that the same model cannot tell what an essay argues. Both hold, and the line between them is verification. A proof either checks out or it does not, so an LLM working on lattices has a referee. A reading of Ulysses has no referee, which leaves the model free to produce something that sounds like a reading.

That distinction decides what happens when the text is a token sale document or a governance proposal. Bitcoin's whitepaper has eighteen years of human analysis behind it, and a model answering questions about it draws on that pile. A proposal posted this morning has none, which is the first-reading case Buterin calls the weak one. The model looks sharpest on the documents you already understand and weakest on the one that goes to a vote on Friday. The assistants in our catalogue of AI tools sell that summarising trick.

Harvard sorted its own courses along that line

The Crimson sampled about 600 courses for this autumn and published the split:

• science and engineering: more than half of the courses allow some AI use
• social sciences: 35%
• arts and humanities: 27%

College dean David Deming had asked faculty to accept or encourage AI in writing-heavy classes. English professor Deirdre Lynch kept a blanket ban in both her autumn courses, and comparative literature professor Homi Bhabha asks his students to keep AI logs instead. Where answers come back right or wrong, Harvard lets the machine in.

Humanities 10 is the course the argument formed around. Stephen Greenblatt and Louis Menand created it in 2008. Students read about one major work a week, from the Odyssey through Dante and Baldwin to Ulysses, attendance is required, there is no pass/fail option, and roughly twice as many first-years apply as the course can seat. The thread kept coming back to that: the class fills up while it bans laptops.

Buterin wrote no instruction. His complaint implies one: use the model for what you can catch it doing, which is finding the reference and naming what a text cites. Keep the judgment about what the text wants, because inference costs fractions of a cent and checking a judgment costs an afternoon. The surface-level reading has a cousin in the disinformation wave analysts expect before the midterms, where text that reads like an argument is the whole product.

Informational material. The quote comes from Buterin's X post of 29 October 2026. The course details come from Newser's summary of Pamela Paul's Wall Street Journal report, and the policy figures from The Crimson's sample of about 600 autumn courses. The comparison with his lattice post and the first-reading argument about governance documents are ours.