Quantum info, Magic the Gathering jokes, Lean, whatever else. | Assistant Professor, Applied Math @ UWaterloo IQC, Perimeter Institute.
Based in Canada
- Tweets 1,426
- Following 412
- Followers 669
- Likes 1,427
I overwrote my keyboard's firmware so that it runs a tiny character-level language model to predict the next key you'll hit. It has per-key backlight LEDs and just lights up the ones you're likely to hit. All inference local on the microprocessor - works without the computer too!
217
326
6,861
305,894
okay @pangram is... really bad?
arxiv.org/abs/2608.01308 is one of the most egregious Claude-ish voices I've seen published online. It's like, really really bad.
Pangram: 0% AI, 98% Human, 2% Assisted. What a joke.
2
8
2,870
at the risk of being cringe ai comic guy ... well, images are a better way of memetically disseminating ideas you care about
Mesopotamians 12kya: "They're eating ... grass. It's a whole village eating grass. Ugh."
-Half of your neighbors think it's sad because they tried eating wild grass on Instant mode and it was bad
-Half your neighbors have tried Emmer germ Ultra and know that gathering's doomed
3
392
Do you remember when OpenAI refused to release GPT-2, because its text generation capabilites were "too dangeous"? Feb 2019.
ALT sam altman sits in a buggy with several bags full of GPUs. Pepperidge Farm Remembers
astra is a powerful model and we are working to make it generally available.
we do not think it is a good strategy to keep powerful models to a chosen few.
given its cyber capabilities, we need a little big longer to do do this safely. but hopefully not too long!
7
543
What corner off AI-psychosis-space is this
(yes it's the full conversation)
chatgpt.com/share/6a6bc942-a…
1
3
277
okay I gave it a much larger sample (and double checked Claude wasn't using tools) and it still fails. This is just a prompting skill issue, LLMs can be convincingly random and Pangram doesn't know.
1
29
578
Doesn't work? @pangram I asked Claude for random numbers (not using code) and it said it's human
2
31
3,726
I won't claim this was a "big open problem" in math, in fact ABKRT just float it as a "for example, perhaps...". But it's compelling, for me, that we can use AI to quickly rule out these guesses!
1
1
6
760
I won't claim this was a "big open problem" in math, in fact ABKRT just float it as a "for example, perhaps...". But it's compelling, for me, that we can use AI to quickly rule out these guesses!
1
56
This is from a job advert in the UK. Without further context, what kind of job do you think this is for?
1
3
236
Excited to be giving a keynote speech at SAIR's Science x AI Summit at the end of the month.😃I'll talking about physics and Lean!
sair.foundation/events/scien…
1
17
408
Sometimes Claude will refer to subagents as child processes, or for short, children. When this propagates through notes, sometimes it misunderstands itself. Which leads to it writing summaries of mathematical analysis like this:
3
145
The first stage of SAIR's "Math Distillation Challenge" completed. Happy that I got 8th place! Right below, *ahem*, someone named "magmaballs".😁
Stage 2 will involve generating Lean proofs and disproofs. Excited to see will come out of it!
competition.sair.foundation/…
8
253
Claude Opus 4.7: Arguably a genius in every domain
Also Claude Opus: "3 isn't a valid hex char" ??
1
102
I ask Claude some follow up questions on a thread.
"These are sharper objections than they might look. Let me take them seriously."
Is this incredible negging or what? If a student in a class asks a question and the professor says it's "a sharper question than it sounds"... 💀
116
Nevertheless, over 14 runs of submitting to Aristotle (with nothing beyond the initial prompt of "prove correctness"), it was able to formalize the correctness of the assembly. First it decomposed it into several subgoals...
2
391
I asked Claude to tell me what it's like, in its own voice.
It coded up a speech synthesizer + audio visualizer.
claude.ai/public/artifacts/b…
me: "can you use whatever resources you like, and python, to generate a short 'youtube poop' video and render it using ffmpeg ? can you put more of a personal spin on it? it should express what it's like to be a LLM"
claude opus 4.6:
1
254
For years, I had seen the warnings. Ignored them. "It can happen to anyone", they said. "Anthony" goes into Starbucks and comes out with a drink for "Annie". "Harold" becomes "Healed".
But no, surely "Alex" was safe? What else could that become?
Today, I am "Leix".
88
There's a popular idea that StackExchange is dying out because it's replaced by AI. I think the problem isn't AI, but UI. Compare usage of screen real estate today vs. 2015. This is when you first open the site.
1
86
This very easy question (stated in plain English on Wikipedia) still stumps Gemini and ChatGPT. Grok gets it, because it uses web search.
"What is the only ethnic sum?"
1
90
Can someone explain to me figure S38 from the supplemental material? Because it *really* looks like (also in the accompanying text) Google is saying that they can actually do it faster on a classical computer, getting better accuracy. "13000x" is only for exact calculation. What?
Building quantum computers faces a core challenge: control qubits without losing info. Hear from Yu Chen, Director, Quantum Processor, on the “secret sauce” behind Willow. It let Quantum Echoes run 13,000x faster, a verifiable quantum advantage. Read → goo.gle/47C7jH9
1
84
The relevant quote (on Wikipedia):
Gemini seems simultaneously convinced that this is a classic solved problem with an elegant solution, and it keeps just barely missing the mark.
1
20
"Dev Mode" has been the official name of the Microsoft app you install on an Xbox to enable developer tools, since 2013. When are they going to go after Msft? Now THAT would be funny
65
Ahhh very useful software expatfile.tax/c/hm/
Finally all those people paying their taxes over there can do so correctly
1
65
This is the kind of weather that in Canada - apparently - means it's time to go out jogging in short shorts and not even a T-shirt. They must be built different up here!
58
From here on, if we wanted a good approximation of arcsin(x), we can just keep taking more derivatives at x=0 and growing the Taylor series. But this is already a great approximation. If we wanted to, we could even tweak the (1-sqrt(2))*x = -0.414x to be -0.425x, and it's better:
1
34
We approximated the function at x = -1 and x = 1; now let's "fix" it at x=0. The derivative at x=0 is 1-sqrt(2), so the function looks approximately like (1-sqrt(2))*x. Let's subtract that off next:
A max error of only 0.015!
1
22
And that's what's there in the first plot: the leftover term. But it's amazing how straight it is! *Just* by capturing a local description at the two endpoints of the interval, we've now got a much smoother function. (Literally - it's not smooth. It was not smooth before.)
1
13
If we wanted to approximate arcsin(x), we could take e.g. a Taylor series at x=0. This works. And this "works well" when all the derivatives "stay small". But we see that, near x=1 (and -1), the derivatives become large:
1
16
Fun Math Fact: arcsin(x) + sqrt(2-2x) - sqrt(2-2x) is actually a linear function. I mean, just look at it:
(This is a plot of the whole function, remember that arcsin only has a domain of [-1,1].)
2
1
126
Humanity's Last Exam has this! ... All the models are terrible. :) Calibration error of ~90%, so, like 10x overconfident.
agi.safe.ai/
1
3
34
