wongarsu

4 ms·

31.
by wongarsu 14d ago
Based on the tables in part 4 it seems like some of the policies would have started overtaking LRU at larger KV-cache sizes. And while the range of cache sizes chosen is reasonable based on the table about Moonshot's number at the star
32.
by wongarsu 14d ago
More examples from this article: "A block with resident children isn't evictable. So the baseline is LRU over radix leaves, which is what SGLang and vLLM actually implement. Beating naive flat LRU would be a strawman." "
33.
by wongarsu 14d ago
Absolutely. Where google is shaping results "to please corpo-political masters", Yandex is obviously influenced by the Kreml I use Kagi as my daily driver. But there are some kinds of queries where my interests and Yandex's i
34.
by wongarsu 15d ago
Notably, those are the consumer terms. The commercial terms [1] are much stricter about what Anthropic can do. That's one of the main draws of the Team plans over the individual plans (in addition to a couple dashboards, shared skills,
35.
by wongarsu 15d ago
If Moonshot offered a "mystery model" tier with a note that my requests might reach whomever, I would have zero qualms about this Pretending to customers like they are serving Kimi while actually proxying Claude is however a bad t
36.
by wongarsu 15d ago
On Windows you can achieve something like manual overcommit by calling VirtualAlloc with just MEM_RESERVE. That gives you a continuous space in your process's virtual address space, without actually backing it with any physical pages.
37.
by wongarsu 15d ago
Notably windows doesn't use overcommit, and degrades much more gracefully under memory pressure. The biggest tradeoff is the amount of disk space consumed by a page file that also has to reserve space for unused pages that have been al
38.
by wongarsu 16d ago
52k on the stack is pretty significant, given Windows defaults to just 1MB stack size per thread
39.
by wongarsu 16d ago
That's simple. The Windows Kernel is a collection of mostly elegant solutions, with a strong peppering of backwards-compatibility cruft, all the way from NT3 to Windows 11. The Windows userland received a lot of effort until about Wind
40.
by wongarsu 16d ago
ZFS being a copy-on-write fs probably made the relative cost of that feature much cheaper
41.
by wongarsu 16d ago
How do you know how big your array has to be in a single pass? I don't think the WinXP source uses vectors or similarly ergonomic auto-growing arrays. You could preallocate an array big enough for 100 paths of length MAX_PATH, but that
42.
by wongarsu 16d ago
If that was the case you would expect a large similarity in the "unprefilled" case, but no significant difference from feeding it some of GPT5.5's CoT (the "delta" column) DeepSeek V4 Flash and Kimi K3 follow that p
43.
by wongarsu 17d ago
That writing style might be a tad too tense If I got it correct (appending B from https://stolen-thoughts.com/paper.pdf is essential) they are the authors of the well-known exploit to recover readable CoT from OpenAI and An
44.
by wongarsu 17d ago
Most serious providers are now supporting structured outputs in a reasonable way for all model configs. But for example on ollama structured outputs are still incompatible with tool calling and with reasoning
45.
by wongarsu 17d ago
Just for reference, in the .NET ecosystem there are multiple vendors selling UI component libraries for sums ranging from lower three digits to upper three digits per dev, per year. This was generally considered good value (might no longer
46.
by wongarsu 18d ago
The "Fox News" right is just as guilty of letting ideology override reason as the "woke" left. As are many other groups (most of Western politics is a bit more complicated than the US's left/right split)
47.
by wongarsu 18d ago
Most things in the EU AI Act are entirely reasonable. On its own it's a pretty sensible piece of legislation. But the US and China are incredibly lenient on enforcing even existing laws against their AI companies. In the US the standar
48.
by wongarsu 19d ago
The thing zstd got really right is fast decompression. For write-once read-never data like backups lzma (aka xz/7zip/lzip) is great. But it takes forever to decompress. On zstd I can get good compression while decompressing the fi
49.
by wongarsu 19d ago
https://www.ftc.gov/business-guidance/blog/2017/02/what-vizi... "Starting in 2014, Vizio made TVs that automatically tracked what consumers were watching and transmitted that data back to its server
50.
by wongarsu 19d ago
All that tells us is that the site is probably not run by the FBI. It could still be run by the CIA, Russian intelligence, or some lone morally neutral guy from Prague
51.
by wongarsu 19d ago
I tend to agree. Skill files become less useful as developer skill increases. As a skilled developer my repetitive instructions are mostly one or two sentence phrases for staring something like a highly-interactive planning session, or a se
52.
by wongarsu 19d ago
Also further down "If the pilot is successful, we might expect the migration to continue on all the 54,000 workstations owned by the federal administrations. That's my guess". So in the estimate of the author we can expect th
53.
by wongarsu 19d ago
Even Microsoft can't perfectly replicate MS Office rendering. Despite access to the source code. Lots of stuff renders slightly broken in the web version, especially once headers and footers are involved As an individual that makes MS
54.
by wongarsu 21d ago
As a society we tend to reward sociopaths with power, which leads to stuff like forever chemicals. But society providing neither checks-and-balances nor even a credible thread of punishment after the fact is also part of the issue. Even soc
55.
by wongarsu 22d ago
Might be usable for short-context utility workloads? Generate the title of your chat session based on the first three messages at the speed of light
56.
by wongarsu 22d ago
sglang is also worth mentioning. Fills a similar role as vLLM, but with less fiddling at the knobs to get a working setup For a GUI experience that also serves an OpenAI-compatible API, Unsloth studio or LM Studio is probably the way to go
57.
by wongarsu 23d ago
https://xkcd.com/908/ Turn's out xkcd's black hat was Elon Musk the whole time
58.
by wongarsu 23d ago
It has great web search, and is one of the best models on the AA-Omniscience Hallucination Rate benchmark (meaning it is really good at admitting it doesn't know something, rather than confidently bullshitting) In terms of raw intellig
59.
by wongarsu 24d ago
I don't see a legal pathway to do that The two competing legal arguments regarding copyright of LLM output are "it's like hiring a monkey" (author is the LLM, which is not a person, thus can't hold copyright and can
60.
by wongarsu 24d ago
That's exactly how you get 'you are right, I deleted the production DB to apply the new schema when I should have written a migration' That said, I do trust Opus and Fable enough to let them deploy to staging. Great for debug

More ›