mswphd

13 ms·

181.
by mswphd 5mo ago
the rust they've written (so far) is highly unidiomatic (and with a ton of unsafe). I can't speak to the zig part, but it seems plausible to me it is line-by-line, horrendous rust. Whether or not they can clean it up is an interes
182.
by mswphd 5mo ago
the rust port (at least currently) heavily uses unsafe as well https://github.com/oven-sh/bun/compare/claude/phase-a-port#d... that isn't particularly surprising, but the point is I would expect get
183.
by mswphd 5mo ago
yes, deriving all of the math cryptography depends on independently would not be easy. Fortunately, that's not really how anybody learns. Along those lines, you do not need to understand the proof of Euler's totient theorem to und
184.
by mswphd 5mo ago
in general the math is not actually that hard. It will be things you don't know beforehand, but a general undergraduate cryptography class will not assume the undergraduates have that much of a better math background than you. Typicall
185.
by mswphd 5mo ago
yeah I generally would say that learning about the actual schemes (tends to be) doable by a casual enthusiast, but learning about how the SOTA attacks work (which motivate scheme design for sure) is much more difficult.
186.
by mswphd 6mo ago
didn't PS3 have a hardcoded nonce for their ECDSA impl that allowed full key recovery? I would agree that I doubt LLMs let people mount side-channel attacks easily on consumer electronics though.
187.
by mswphd 6mo ago
This does happen some on other websites. I've often seen people in math discuss things that Bill Thurston (a fields medalist) has posted on mathoverflow https://mathoverflow.net/users/9062/bill-thurston note
188.
by mswphd 6mo ago
In math, Grothendieck's correspondences are quite famous. Here's a book that is entirely 300 pages of letters between him and J.P. Serre https://webusers.imj-prg.fr/~leila.schneps/grothendieckcircl... There a
189.
by mswphd 6mo ago
have any of their risky bets paid off though? most of their main products have been acquisitions.
190.
by mswphd 6mo ago
inflation has been persistently > 2% (and arguably much more, as the current methodology on how to measure inflation is quite flawed). There's a definite risk of inflation expectations shifting, which central bankers really want t
191.
by mswphd 6mo ago
this is a dense model, so that's expected. On a mac you'd want to try out the Mixture of Experts Qwen3.6 release, namely Qwen3.6-35B-A3B. On an M4 Pro I get ~70 tok/s with it. If your numbers are slower than this, it might be
192.
by mswphd 6mo ago
that's not it. 35B-A3B is a "Mixture of Experts" model. Roughly, only ~3B parameters are active at a time. So, the actual computational requirements scale with this ~3B, rather than with the 35B (though you need high-bandwidt
193.
by mswphd 6mo ago
I'm not excited about it, but the only main way I've been able to discover LLM-isms that sneak in are 1. via seeing them glimpse by in the agents' window as its making edits (e.g. manual oversight), or 2. when running into an
194.
by mswphd 6mo ago
I'm expecting we'll likely end back up on agents making PRs, and having to review them. Either that or giving up on quality etc/dealing with very messy code. I've been trying various automated testing/linting/e
195.
by mswphd 6mo ago
An "obvious" point to make is that it is not particularly usable on a unified memory machine. Only getting 9 tok/s (for Q6 quants) using a Macbook M4 Pro 48GB memory (though with GGUFs, not mlx). The quality seems fine, but t
196.
by mswphd 6mo ago
context window for Qwen3.6 models' size increase isn't that bad/large (e.g. you can likely fix max context well within the 48GB), but macbook prompt processing is notoriously slow (At least up through M4. M5 got some speedup
197.
by mswphd 6mo ago
1. the 35B model is a "Mixture of Experts" model. So the earlier commenter's point that it is "larger" does not mean it is more capable. Those types of models only have certain parts of themselves active (for 35b-A3
198.
by mswphd 6mo ago
kind of a circular argument though? the reasonable definition of "unsecured" is "stuff you can't muck about with". That might change over time as attacks/exploits are developed though.
199.
by mswphd 6mo ago
They might be trying to talk about Rice's theorem? https://en.wikipedia.org/wiki/Rice%27s_theorem Formally, any non-trivial semantic property of a Turing machine is undecidable. Semantic here (roughly) means &quo
200.
by mswphd 6mo ago
LLM's train in some standardized ways to emit things like tool calls, right? if you make those tokens a fundamental part of your programming language, it's possible you'd be able to run into tokenizer bugs that make LLMs much
201.
by mswphd 6mo ago
you could do something like rent GPU time yourself, and use it to run a higher-quality local model (e.g. one of the Chinese "close to frontier" ones). Not guaranteed to preserve privacy of course, but it at least avoids directly
202.
by mswphd 6mo ago
there are no meaningful questions. The only way there are meaningful questions is if you think global cryptographers + governments are part of a cabal to build insecure schemes. The new schemes use 1. cryptography developed across the world
203.
by mswphd 6mo ago
you don't really need that tbh. you can get pretty good speedups using standard (vector) intrinsics. the new algorithms are (mostly) modular linear algebra (+ some concept of "noise").
204.
by mswphd 6mo ago
they are much more thoroughly vetted than other schemes. They're more thoroughly vetted than elliptic curves were before we deployed them. Much more vetted than RSA was ever. Practically though, there are some downsides. Elliptic curve
205.
by mswphd 6mo ago
they're almost assuredly talking about two things (maybe 3 if they really know what they're talking about, but the third is something that people making this argument like to pretend doesn't exist). 1. the main "eye ca