ot

9 ms·

91.
by ot 2y ago
> You do need a fence in the unlock path though (at least a release fence). Well yes but on x86 that comes for free. The overhead of the full fence brought in by lock cmpxchg or lock xchg is in the order of ~10ns, which for an uncontende
92.
by ot 2y ago
> Reason: locks that have the ability to put the thread to sleep on a queue must do compare-and-swap (or at least an atomic RMW) on `unlock`. But spinlocks can get away with just doing a store-release (or just a store with a compiler fen
93.
by ot 2y ago
Yeah, that specific benchmark is actually likely to prefer undesirable behaviors, for example pathological unfairness: clearly the optimal scheduling of those threads runs first all the increments from the first thread, then all of the seco
94.
Why So Few Matt Levines?(gwern.net)
8 points by ot 2y ago|2 comments
95.
by ot 2y ago
Even if we were talking about real numbers (which we're not), it would still be true that (0, 1) and (1, +inf) have the same cardinality :) As touched on in the post, 1/x is an explicit bijection (but any set of real numbers that
96.
by ot 2y ago
VByte is a weird baseline to compare to: it is a byte-aligned encoding scheme, so it trades off some space efficiency for speed of decoding. The proposed method is bit-aligned, so it should be compared against bit-aligned encoding schemes.
97.
Uv: An fast Python package and project manager, written in Rust(github.com)
4 points by ot 2y ago|1 comments
98.
Twitter Takeover Is Now the Worst Buyout for Banks Since the Financial Crisis(wsj.com)
64 points by ot 2y ago|19 comments
99.
The Greenwich Meridian is in the wrong place [video](youtube.com)
1 points by ot 2y ago|0 comments
100.
by ot 2y ago
From the article > According to the Times, Shah wants CNET because it’s a “well-known industry brand” and still has an audience that’s large enough to be attractive to tech advertisers.
101.
Design doesn't have to end like this(jovo.design)
3 points by ot 2y ago|0 comments
102.
Meta 2Q'24 Update(mbi-deepdives.com)
1 points by ot 2y ago|0 comments
103.
Bing Generative Search(blogs.bing.com)
2 points by ot 2y ago|0 comments
104.
Phi-3 fine-tuning, new generative AI models(azure.microsoft.com)
2 points by ot 2y ago|0 comments
105.
by ot 2y ago
> there's a large company with an almost identical name Are you suggesting that this wasn't intentional? The name is clearly a play on "zero shot" + "xerox"
106.
by ot 2y ago
You could say the same thing about std::monostate, which is not a dummy type. If you need a unique sentinel type you have to make one for that purpose.
107.
by ot 2y ago
You can use void for that.
108.
by ot 2y ago
Folly has DeterministicSchedule, which also wraps atomics and it is used to test its core synchronization primitives, but I don't think it's as sophisticated as loom. https://github.com/facebook/folly/blo
109.
by ot 2y ago
> It's a shame because I know Google can put out easy to read code (see: the go standard library). My guess is that the difference is that go is managed by a small group of engineers that have strong opinions, really care about it,
110.
by ot 2y ago
I'm one of the authors of the paper, nice to see it on HN. I remember when we first experimented with this, the compression improvement compared to our previous heuristic was massive, but the algorithm took a day to run on a double-dig
111.
by ot 2y ago
I haven't worked in this field for a while, but if you look at the citations of the paper [1] there are quite a few applications. One that I find very interesting is optimizing function layout in binaries to improve their compressibili
112.
by ot 2y ago
> Google said this change is to allow the search company to serve the search results faster on more searches I'm surprised that this makes enough of an impact: in order to find the top 10 results the engine needs to retrieve and ran
113.
Forward? Error? Correction?(quic.video)
1 points by ot 2y ago|0 comments
114.
by ot 2y ago
LAN can be oversubscribed too, even if your switches have sufficient capacity you can have short packet bursts that fill up the queues. You cannot assume reliable delivery.
115.
by ot 2y ago
Everything at some layer has to run over datagrams, since that's what IP is. This is literally the point of the article: if you want to create a protocol over raw datagrams, you have to implement a lot of things that are very hard to g
116.
Formal methods: Just good engineering practice?(brooker.co.za)
179 points by ot 2y ago|149 comments
117.
by ot 2y ago
I would speculate that OpenAI is in a phase where speed of delivery is make-or-break, and any bloat would be a distraction. I bet they're extremely deliberate in their acquisitions.
118.
by ot 2y ago
I guess I stand corrected then :) (Hi!) EDIT: I forgot to say, with the recent hires and the Rockset team, OpenAI is building quite the infra dream team :)
119.
by ot 2y ago
Very unexpected acquisition. I don't think that Rockset is a suitable infrastructure for RAG, a purpose-built inverted index would be far more efficient (both in terms of compute and storage), so I'm not sure how much of the techn
120.
by ot 2y ago
You might be thinking of JBIG2 https://en.wikipedia.org/wiki/JBIG2#Character_substitution_e...

More ›