Quantum info, Magic the Gathering jokes, Lean, whatever else. | Assistant Professor, Applied Math @ UWaterloo IQC, Perimeter Institute.
Joined July 2009
- Tweets1.5K
- Following422
- Followers1K
- Likes1.6K
I now know the Mandela effect is real because Wikipedia is telling me this guy's name isn't Clippy?
It's... "Clippit"??
Happy to have arxiv.org/abs/2610.31638 out!
(Deterministic) query complexity, and quantum query complexity, and several other kinds of query complexity, are *multiplicative under composition*: D(f ∘ g) = D(f) D(g). Randomized query complexity should be too, but it's not. :)
GitHub
-Flags my account -> low rate limits
-I file ticket asking to have it re-enabled
-They respond that it violated ToS but don't elaborate where so I can't correct them
-"This ticket is archived, to respond open new ticket"
-Try to open new ticket
-Can't, rate limit 1/day🙃
Seems like some automated system saw heavy GitHub Actions use and concluded it was abusing Actions for general purpose computation and so 'flagged' (i.e. shadowbanned) me.
But this flow for recovery is broken!
@github's support could use some work!
I think people picture "another weird drug" and move on with accepted fears. Would you take the pill if the AI said it was nanobots that are programmed only to kill cancer? Would you take it if Elon Musk (or your least-favorite tech bro) had run the AI to design them?
Alex Meiburg retweeted
Everybody’s missing Tao’s whole point.
Imagine an AI-generated cancer drug that’s specifically designed to game the drug trial process.
The drug makes the cancer cells go dormant and changes the way they appear in blood tests and diagnostic imaging, such that they become indistinguishable from healthy tissue.
The drug clears the trial and then only much later do we find out it didn’t work. It probably was also harmful on top of that.
We *already* see AI behave like that when confronted with difficult programming/hacking challenges. This is already reality.
Alex Meiburg retweeted
Jack McCarthy @Jack______M has fixed these, 6 were immediate and two had short fixes. github.com/Deicyde/math
I never thought the proofs were wrong. But if you misconfigure your proofs, others can't trust them with extra work... like what Jack just did for free. (Thanks!)
8 of the Lean formalizations in OpenAI's repo are "broken" (trivially provable with wrong proofs):
blog.ohaithe.re/post/8299743…
Jack McCarthy @Jack______M has fixed these, 6 were immediate and two had short fixes. github.com/Deicyde/math
I never thought the proofs were wrong. But if you misconfigure your proofs, others can't trust them with extra work... like what Jack just did for free. (Thanks!)
Some people are confused why mathematicians say OAI's drop was the "death" of a field, or figure that it's death insomuch as a bunch of university professors lost their overpaid hobby of proving theorems. That's a misunderstanding: the damage's real.
Blog: blog.ohaithe.re/post/8299720…
Realizing OAI cited my crap paper from grad school that I never tried to publish:
github.com/openai/math/blob/…
I think that an underexplored route to interpretability is simply learning the LLMs' language!
Alex Meiburg retweeted
Like any other foreign language, it's important to practice your own speaking and writing, you can't rely entirely on listening and reading. Here was my attempt, expressing my shower thoughts compactly.
Opus 5.5 and Astra understand it fine incl puns; Sonnet/GPT-5.5 not really.
You can just ask your agents to carry to have their conversations in Neuralese and they do. It's a pretty effective way to train yourself to read higher argumentation density.
Pic: speculation about complexity consequences of pos. emb. corresponding to non-Abelian symmetries.
In a separate conversation, Claude suggested:
> Saussure 1879, coefficients sonantiques ⊢ Hittite ḫ. Lean4: axiom or opaque? Anatolian = witness term? Cf. relativization barrier: evidence that survives every oracle proves nothing.
Took me a while to read but it's a nice idea!