suprjami

5 ms·

31.
by suprjami 2mo ago
You have understood correctly. One really would think these companies (including Google) who spend many millions of dollars on compute could write a few hundred lines of Jinja correctly, so their investment works optimally or at all. But th
32.
by suprjami 2mo ago
Our desire for better local models just happens to coincide with China's desire to destroy the western AI company business model by releasing local models. I doubt there's any philanthropy involved.
33.
by suprjami 2mo ago
Unsloth imatrix data puts their quants at lower KLD than almost all others. It's true they make architecture-specific changes like keeping certain layers at F16 but it's also more than that.
34.
by suprjami 2mo ago
Spoiler: Rogue just implements its own Linear Congruential Generator. #define RN (((seed = seed*11109+13849) & 0x7fff) >> 1) I've poked around in a couple of libraries from this time. At least QuickBasic and QuickC a
35.
by suprjami 2mo ago
I have the same opinion as your first paragraph, but I don't want to spend weeks or months vibe-coding basic features which come built into almost every other agent. Yeah maybe Claude/OpenCode/KiloCode/Hermes/whatev
36.
by suprjami 2mo ago
Tactics Ogre (PSP and Reborn) solves this but giving you a map of the storyline. You can travel back to "anchor points" where the story diverges, so you can experience the whole game without starting over. https://ogreb
37.
by suprjami 2mo ago
Ultima Underworld runes also had a word associated with them, and the words had meaning, which was both amusing when you used them for spells, and iirc could result in experimentation to find new or undocumented spells. https://w
38.
by suprjami 2mo ago
This reminds me that as a child I taught myself Tolkien's runes from The Hobbit, which is mostly another substitution of the modern English alphabet: https://en.wikipedia.org/wiki/Cirth#Runes_from_The_Hobbit I got
39.
by suprjami 2mo ago
When buying a new phone, doesn't everyone sit down and set everything up so it all works when you want it next?
40.
by suprjami 2mo ago
It's stated in many places that inference is profitable (margin 70% to 90%), only research is expensive. Hyperscalers are burning through their cashflow training new models while inference-only providers are printing money. If all the
41.
by suprjami 2mo ago
I have seen an employer approve a degree under training budget. Just needs the right employer.
42.
by suprjami 2mo ago
Depends on your setup. If you drop $10k on an RTX Pro 6000 then yeah Qwen 35B MoE will absolutely fly. If you have a pair of 3090s and run Qwen 27B, or an old Threadripper with heaps of system RAM and Deepseek or MiniMax or Kimi, no it won&
43.
by suprjami 2mo ago
This is a correct (though shallow) take on two problem solving approaches which have been formally described for over a decade: https://materiasiis.uson.mx/docs/control_de_calidad/1.CALIDA... Notably, when the Luc
44.
by suprjami 2mo ago
You can get 3080 20G from China.
45.
by suprjami 2mo ago
You can also get 3080 with 20G from China.
46.
by suprjami 3mo ago
No. Your link is the original post by Nikhil. This page is Gruber's commentary on Nikhil's original post. We make sure to click the link and read the article before posting here. Please follow that rule in future.
47.
by suprjami 3mo ago
Just merged into main llama.cpp a few hours ago: https://github.com/ggml-org/llama.cpp/pull/25165
48.
by suprjami 3mo ago
You had me up until you equated LLM usage to unreliable human delegation. No human being is going to genuinely suggest to glue cheese onto a pizza to stop it sliding off. These things aren't people. Stop anthropomorphizing a computer p
49.
by suprjami 3mo ago
If you specify the random seed and set temperature to zero, you'll always get the same response. As you said, it's just math. That doesn't mean the response is correct though. You've just hard-locked the model to one pos
50.
by suprjami 3mo ago
The gap has been steadily closing over time. Opus 4.8 (May) to Kimi K3 (July) has apparently just dropped it to two months. China also does efficiency improvements. Qwen 3.6 27B is better than Sonnet 4.5 and you can run it on a couple of ga
51.
by suprjami 3mo ago
My previous comment from 24 hours ago is now irrelevant. Newly released Kimi K3 is benching better than Claude Opus 4.8. The only better models are Claude Fable and GPT 5.6 Sol Max Effort.
52.
by suprjami 3mo ago
You're not just tired (em dash) you're exhausted!
53.
by suprjami 3mo ago
Another one besides Wikipedia https://tropes.fyi/tropes-md
54.
by suprjami 3mo ago
DS3 isn't even looked at anymore. GLM-5.2 is the best in that class right now. It is competitive with current GPT/Claude/Gemini.
55.
by suprjami 3mo ago
> LLMs... impedes the ordinary process of theory-building As I have said on here before, I actually really enjoy LLMs for code understanding precisely because they are imperfect. They're good enough to point you in the right directi
56.
In defense of not understanding your codebase(seangoedecke.com)
5 points by suprjami 3mo ago|3 comments
57.
by suprjami 3mo ago
Yes, seriously everyone should watch this: https://www.youtube.com/@AngeTheGreat Other engine simulators work by approximating the engine. Ange's engine simulator works by approximating physics of air fluid dynamics th
58.
by suprjami 3mo ago
Finally a believable headline about CATL. Next week we'll be back to CATL increasing energy density by 1 million percent with a battery that charges in 3 nanoseconds.
59.
by suprjami 3mo ago
Absolutely not lossless: https://www.reddit.com/r/LocalLLaMA/comments/1twz9ur/cyankiw...
60.
by suprjami 3mo ago
Here is the post: https://xcancel.com/_vkaku/status/2071469740141224272 It's mainly just an image and link to https://github.com/guilt/3DGFX Check the commit history for the addition of

More ›