Claude digests 1,009 pages of Gödel shorthand, surfacing hundreds of unknown theorems
- The Finnish logician Jan von Plato has led a Helsinki University project since 2017 to read 16 notebooks Gödel wrote in the early 1940s, 1,009 pages of dense maths and logic in Gabelsberger shorthand, which was already obsolete when Gödel used it.
- Progress ran at roughly a page a week until September, when von Plato gave the material to Claude and got back a 650-page explanation of the whole corpus within a couple of days.
- Von Plato says the workbooks contain hundreds of theorems nobody knew about, dozens of anticipations of results not established until the 1970s and 80s, and 14 themes in logic that have not been pursued.
- Von Plato taught himself Gabelsberger (he says he is one of about six people in the world who can read it) with a €2.5m European Research Council grant, and fed Claude his own transcriptions plus six books he had written on Gödel and other secondary literature; the output still had to be checked page for page.
- Gödel fled Austria in 1940, settled at Princeton's Institute for Advanced Study, and died in 1978 of starvation after refusing food his wife had not cooked.
Hacker News 의견들
Nice timing for the IPO. The AI was trained on existing pre-AI transcriptions of Gödel's shorthand, which he learned in school around 1920 and which was widespread, only technically obsolete.
So what's the theory here, that von Plato is running a campaign for Anthropic? He's the foremost expert on the subject. A model generalizing from a few examples of the shorthand seems like a point in its favor, not against it.
I want to hear about these 14 themes in logic Gödel pursued and nobody followed up on.
This is genuinely exciting. Gödel published almost nothing after the 1930s, and von Plato is the world's leading expert here. If he's impressed by the output, that means something.
There's a slide deck with some of the results at aitp-conference.org, and the Helsinki research group has a site too.
I suspect Gödel's own critiques of modern logic will turn out to be the maddening limits of LLMs.
This might be a real example of someone making himself obsolete with AI. He learned the shorthand to stay indispensable after mandatory retirement.
Or the opposite: he kept himself from becoming obsolete by picking up the tool. Let me know when the AI gets a €2.1m ERC grant as PI.
You don't get credit for work you never tell anyone about. Newton pulled the same thing with calculus after Leibniz published. I wonder if AI makes mathematicians hoard results harder, or pushes them to post scratch work early.
What would the new incentive even be? Don't publish and get paid anyway?
The Gabelsberger thing sounds like hyperbole to me. My father wrote in a custom version of it until 2020, and he wasn't one of six people in the world.
I think the six are people who can read Gabelsberger and know enough mathematical logic to parse what Gödel is saying.
How much expertise did this actually need? Could a general purpose model one-shot the translation, or did the subject matter expert do the heavy lifting?
A decades-old properly trained OCR algorithm could probably have unlocked these manuscripts.
Gödel was a perfectionist who only published work he considered bulletproof. It's unlikely his notes contain finished theorems he simply failed to notice.