mswphd

7 ms·

91.
by mswphd 3mo ago
dense models are (more) compute heavy, so are generally worse to run on mac. mac tends to be better for (larger) MoE models. 27B dense can fit on a consumer graphics card. Even without getting into various "intrusive" ways to shri
92.
by mswphd 3mo ago
CPU moe offloading. see e.g. https://www.reddit.com/r/LocalLLaMA/comments/1t9eo83/running...
93.
by mswphd 4mo ago
I've generally heard it as manufacturing woes due to competition with China. in particular germany used to sell a significant number of cars to china, and now is being outcompeted (both in china and many other locations) by chinese
94.
by mswphd 4mo ago
that's broadly true, but there are some areas of CS that are at least as close (say at a minimum PL theory). it's also less math heavy than e.g. complexity theory (though that's admittedly smaller). It can also be less math-h
95.
by mswphd 4mo ago
I'm confused. China has been in a fairly bad recession the last ~2 years (their housing bubble popping). Why are we assuming that recessions can only happen in the US? I'm also confused how China managed to convince so many US VC&
96.
by mswphd 4mo ago
this is not an accurate picture of Chinese industrial policy. In fact, you could argue they have the opposite problem. Their industrial policy encourages too many companies to enter a space, where the resulting competition kills off profit
97.
by mswphd 4mo ago
this is a justification for why one frontier lab would have extremely high spend rates. There is >1 frontier lab though. More explicitly: if the Chinese can get near-leading performance models at a fraction of the cost by stealing from A
98.
by mswphd 4mo ago
As an explicit example, major revenue streams that Tesla (used to) take advantage of are 1. the EV tax credit, and 2. carbon credits so the richest man in the world/the US had significant tailwinds for a central business venture of h
99.
by mswphd 4mo ago
CS can (but not frequently) have the revolutionary result you mention as well. A candidate Fully Homomorphic Encryption scheme was first detailed in Craig Gentry's thesis, for example. That being said, this is much less common than a 1
100.
by mswphd 4mo ago
interesting links > Others think someone from the Rust (programming language, not video game) development community was responsible due to how critical René has been of that project, but *those claims are entirely unsubstantiated* huh &g
101.
by mswphd 4mo ago
This kind of reasoning leads you to reasoning that if he was an ineffective fraudster, it would be less moral, as he would have bought less mosquito nets. So it’s not only moral to do fraud, but you most extremely competently do fraud.
102.
by mswphd 4mo ago
the way to interpret the gigantic lean proof is not by inlining each lemma, looking at all the lines, and thinking "yeah that's a lot". That's also not the way to read a paper. Instead, you proceed in layers of abstracti
103.
by mswphd 4mo ago
The first half of your answer presupposes some platonic utilitarian calculus that, if it were applied correctly, would yield moral outcomes. This is very hard to believe. If I look at notable/well-known examples of EA-affiliated people
104.
by mswphd 4mo ago
there's pros and cons to it for them. Clearly, they get good branding (at least in enthusiast circles). Perhaps more important is they get community work on optimization. There have been significant performance uplifts on the Qwen3.6 m
105.
by mswphd 4mo ago
love thought-terminating cliches. really helps keep from actually thinking ever.
106.
by mswphd 4mo ago
both no in principle, and when you're used to reading LaTeX, word is ugly. It's a milder form of how if these notes were handwritten it wouldn't matter, but it would also be less appealing than them being typeset well.
107.
by mswphd 4mo ago
Yes and no. Qwen 3.5 was released 3/2/2026. It includes models up to a 397B-A17B model https://huggingface.co/collections/Qwen/qwen35 A day afterwards, a high-up technical leader working on Qwen was let
108.
by mswphd 4mo ago
|x| on a vector is not well-typed. The notation for norms of vectors is ||x||, namely 2 bars on each side. There still is a naming collision between |X| (cardinality of a set) and |x| (absolute value of a scalar). Sometimes this happens. It
109.
by mswphd 4mo ago
Funny: you can cf "sorting network", and see they use them within their own design even.
110.
by mswphd 4mo ago
this is broadly true. I've heard some criticisms of industrial QC research along these lines, namely that it runs into the issue that its entire TAM ends up being some combination of 1. defense contracting, which can definitely be lucr
111.
by mswphd 4mo ago
I am pointing out a particular cryptographer's abysmal track record in understanding the security of PQ schemes to call into question their current criticisms of PQ schemes. They've always been (in my opinion obviously) fear-monge
112.
by mswphd 4mo ago
Nation states have been pouring billions into QC. It's hard to collect the varous announcements into clean figures, but rough estimates are that the US has allocated ~$5B to QC computation research, the EU (via the EU itself, and indiv
113.
by mswphd 4mo ago
the rational argument is that this time is not particularly worse than prior transitions, and arguably is one we are doing much more clear-eyed (think about all the ECC vulnerabilities during their first few years of deployment due to not k
114.
by mswphd 4mo ago
note that OTP is only "perfectly secure" for a rather limited notion of security, namely IND-CPA. This is (roughly) an "honest but curious" adversary who looks at data on the wire (or wherever), but never tampers with it
115.
by mswphd 4mo ago
these have always been an issue, and were the motivation for starting the NIST standardization in ~2016. My point is more that recent developments in quantum computing have caused many cryptographers to go from "we should do this so pe
116.
by mswphd 4mo ago
Again, not everyone responds to a perceived social ill by gambling. Timing a short is still gambling.
117.
by mswphd 4mo ago
it's worth mentioning opinions have started to shift away from this. Quantum computing has made quite concrete progress in the last ~2 years. No guarantee this continues, but among people I know it has changed their perspectives from (
118.
by mswphd 4mo ago
First, it doesn't, because we don't use public-key encryption. Instead, we use key-encapsulation mechanisms, which you have to hybridize in another way. Second, hybridization can add weaknesses in several ways 1. Hybridization may
119.
by mswphd 4mo ago
> as they almost certainly have weaknesses for some of those new PQ schemes already lying around why believe this about PQ schemes vs about pre-existing schemes? Or any other schemes? It's also worth mentioning that it appears that
120.
by mswphd 4mo ago
it's not just those two institutions. South Korea is running their own standardization currently, and fundamentally similar algorithms are expected to win (some more modern insights might be incorporated, due to starting >=5 years a

More ›