2025-05-19T19:09:06.153Z
then there's also some very technical arguments about jurisdiction that are over my head
I wouldn't be worried about this lawsuit getting dismissed entirely, but some of the claims might not survive
2025-05-21T19:20:40.574Z


2025-05-21T19:22:23.887Z

All taken from this excellent talk: https://www.youtube.com/watch?v=J1YH_GsS-e0
2025-05-26T17:48:17.021Z
Quick wall-time test for a mathematical calculation test:
Python code (original): 178s
Direct Rust translation: 28s
Rust with compile-time optimizations: 8s
2025-05-26T17:50:03.298Z
that's actually kind shocking to me. it was super easy to translate over (although only since late last year when isqrt was added to the standard library https://github.com/rust-lang/rust/issues/116226), and it's ~20x faster
*kind of
oh that link doesn't work you might need to copy it and remove the last two characters
Thinkblog moment
2025-05-26T17:52:30.918Z
(this might become the new Moser's Power problem, btw. still early stages of looking into it atm)
2025-05-27T04:01:32.769Z
this is fascinating https://stackoverflow.com/a/18686659/8507259
2025-05-27T16:38:03.464Z
Interesting sub-question: are there any solutions to C²=A²+2AB+2B², D²=2A²+2AB+B²? Seems like it should be pretty easy to prove if I was a bit more familiar with Diophantine equation techniques
(in the positive integers)
2025-05-27T16:40:04.024Z
I think the answer is almost certainly "no", there would probably be a small solution if one existed and quick simulation shows that there is no such small (less than a few thousand) solution
It also might be provable with Lean
2025-05-28T02:29:01.578Z
okay time to introduce properly what I'm doing: looking for solutions to the Rational Distance Problem! https://mathworld.wolfram.com/RationalDistanceProblem.html
this problem is interesting to tackle numerically for a few reasons
one is that people have found explicit families of near-solutions, which can make checking a lot more efficient
2025-05-28T02:31:13.001Z
however, a) those families have already been trialed a fair bit, and b) I get the sense that any solutions are actually more likely to *not* belong to a nice easy family like that
another reason is that there's no particlar reason to think that solutions should be small! because we're looking at rational distances, when we convert to integers, were essentially multiplying by the LCM of four rational numbers, which could be quite large
*the denominators of those numbers
so unlike many other problems, exhausting a reasonably large solution space doesn't strongly imply that no solutions exist, which means that there's no reason to not keep going!
arguably this makes it less interesting in that negative results are even more meaningless, but I think long-term it's more exciting computationally
2025-05-28T02:40:10.665Z
concretely, what I'm doing is looking for positive integers A, B, C, D, S such that Δ² := (B² - (S - A)²)((S + A)² - B²), C² = S² + B² ± Δ, D² = A² - B² + C²
The way I've written this is such that it's really a search over the 3D space (A, B, S) rather than the 5D space (A, B, C, D, S) because that's vastly more efficient
2025-05-28T02:43:03.897Z
for any A, B, S there's only two possible C values, which will overwhelmingly likely not be integers, mostly because Δ will very likely not be an integer, so this way we can discard most A, B, S combinations pretty quickly
2025-05-28T02:45:20.189Z
using some simple Rust code I bruteforced for values up to 10,000
that took something like 2.5 hours
growing with the cube of our search size instead of the fifth power is an improvement but it's still bad
so I think the next logical step is to write a CUDA kernel
AVX instructions actually happen to not work all that well for some of the optimizations we'd like to do
and I'd like to learn GPGPU someday, so this isn't a bad opportunity
* to be clear, the PTX does have the instructions I want, mostly. there's one that would be *really* nice to have from AVX but I can work around it
2025-05-28T02:51:25.739Z
this is also embarassingly parallel, which is nice. except that it's probably more efficient to execute in dependent stages, which makes it no longer embarassingly parallel and probably an absolute headache
useful reference: https://blog.speechmatics.com/pointless-gpu-optimization-exercise
2025-05-28T02:55:40.563Z
there's also very little memory transfer required. we can probably just give the GPU an offset for each batch and let it calculate everything else from batch indices, which is super nice. then good candidate solutions can be transferred back to CPU for the final tests.
2025-05-28T03:02:42.904Z
figuring out the threshold for "good candidate tests" should be interesting. for integer Δ, it seems like about .025% of triplets pass, which is actually kind of a lot. For integer C, it's only .0004%.
2025-05-28T03:12:06.898Z
I've got about 2000 CUDA cores, running at about 1GHz. Most candidates will be eliminated in the first stage with hopefully about 20 clock cycles, let's say the average is 30 because the stragglers will be significantly more but I have no idea how much yet. That comes out to 17 million candidates passed to the CPU per second with the first threshold, but only 250,000 for the second threshold. That might be overly optimistic, but processing 17 million good candidates per second on the CPU may be an issue. the sheer memory transfer may also be an issue.
2025-05-29T20:53:17.499Z
Four new filings in the Honey lawsuit. I took the ~$5 hit to my quarterly PACER cap to download them (and upload them to CourtListener, of course). Long and short of it is, we're officially entering discovery. Probably the case will be mostly quiet for the next few months.
2025-05-29T20:55:25.898Z
There was an update in Campbell v. Honey the other day too, which I didn't cover because it wasn't all that important. It's a response by PayPal to Plaintiff's opposition to their motion to dismiss. It's well-written, and more convincing than I expected. Still, we'll just have to see how the judge goes on this one, since it's primarily a legal argument at this stage and a complex one at that.
2025-05-30T19:53:34.346Z
it's funny how at this zoom level South Korea looks less populated than North Korea

2025-05-31T21:30:26.800Z
I've been watching Ranma ½ recently, just finished up the initial production run of 21 episodes
Azusa should play The Sims she would be such a good Sims girlie
2025-06-01T04:53:07.431Z
Is it bad that I just had a project idea and I'm already thinking about a logo design that I'll probably have to sink 20-40 hours into?
I'm going to go with yes
2025-06-01T18:35:54.695Z
3 hours into logo design, back to square one 👍
2025-06-01T23:54:19.403Z
gotta say it actually looks decent atm, still work to do though