antirez

Reproducible bugs are candies.
I like programming too much for not liking automatic programming.
In elenco AGCOM.

[antirez @antirez]

My second short story release in English is ready: Tales of Illustrious Computer Scientists: Iola Varga, nun and computer scientist. https://t.co/i1Xp0REuWh

[antirez @antirez]

Coming from high end MacBooks the Dell XPS screen and colors are a shock. OLED is another world.

[antirez @antirez]

A huge thank you to @syntaxfm for a generous sponsorship of my GitHub work. Thanks!

[antirez @antirez]

@0xMassi @omarchy @dhh reading myself while computing the current layer.

[antirez @antirez]

Trying @omarchy in a Dell XPS kindly donated by @dhh (I will try to adapt DwarfStar SSD streaming of large models btw, maybe it will be fast enough for QA usage). I'm really impressed by the "welcome" after the installation. The system is playful and not boring, which is cool.

[antirez @antirez]

@dhh The food there is also incredibile. And the people you can meet incredibly smart. I look forward to return back soon or later.

[antirez @antirez]

@liuliu The incredible thing is how more speed totally changes the way the model is preceived also from the POV of capabilities. It looks more intelligent, not just faster, as reiterating fast is for the models a fundamental way to make progresses.

[Liu Liu @liuliu]

After use these models intensively, now I understand what @antirez means. DeepSeek 4.1 Flash 1017 decode speed locally should be ~200 tok/s to be competitive.

[antirez @antirez]

@KyrelleBuckley Ask an LLM for clarification why this Mao sentence is relevant today with AI fore more clues.

[antirez @antirez]

AI: there is great chaos under heaven; the situation is excellent.

[antirez @antirez]

Indeed. A key Italian band, the CCCP, had this citation in one of their top songs. You can easily find many of their songs on YouTube. Moreover Italy was on the edge of turning communist in the '70s, so we have a strong heritage, similarly to how Calvino and Gramsci are the most read Italian authors in Cina. Today people think each country is different from the other, but we co-evolved strongly in a single great human culture.

[antirez @antirez]

Even if the DGX Spark is SM121 and not SM100, yet maybe we can get something good for DwarfStar kernels from DeepSeek Infer. Let's try.

[antirez @antirez]

@jmbollenbacher SM100 is a peculiar architecture compared to SM121, much more powerful. Something can be done, but not much, also Infer is specially suited for batching while local is often single session.

[antirez @antirez]

DeepSeek infer is the cloud version of DwarfStar, conceptually. Vertical inference engine for one model and specific hardware. Good move. And well, a lesson to the ones that criticized this approach as well.

[antirez @antirez]

@dusangran Between 4.0 and 4.1 it was easy, till something completely new appears in the attention side, all the rest can be kinda reused.

[antirez @antirez]

On my GitHub as usually.

[antirez @antirez]

For the ZX Spectrum lovers, there was enough documentation around to also simulate the whole Spectrum PCB with very precise ULA emulation (cross checked with the actual gates implementing it).

[antirez @antirez]

@CryptoBakeneko @ArenaTiziano mxfp4 is 4 bit, but it's lossless. So there is no need for q4.

[antirez @antirez]

DwarfStar GGUF for DeepSeek v4.1 Flash new checkpoint is now available. It is an improvement even only for the better token efficiency. Upgrading is a must.

[prasincs ⚓️🇺🇸🇳🇵 @prasincs]

Tested @antirez’s Dwarfstar on my 128GB M5 Max Mac Studio with DeepSeek Flash 4.1 and I can’t believe I’m getting 16 tok/s output with ssd offload. That’s 3x more than what’d fit in memory.

[antirez @antirez]

@dhh https://t.co/KlofyKV6AW