Steeve Morin

Building @zml_ai (and we're hiring), ex @zenly, ex Exalead, ex @google. Skydiver and wingsuiter.

[Steeve Morin @steeve]

@mikiane Pour le coup non

[Steeve Morin @steeve]

@knowrohit07 @dzbarsky you *think* you like meson but bluer skies await!

[Steeve Morin @steeve]

.@dzbarsky (OpenAI): “At super scale, everything breaks, except Bazel”

[Steeve Morin @steeve]

The room is overflowing for @corentinanjuna and @dzbarsky’s talk on HermeticBuild at BazelCon 26

[Steeve Morin @steeve]

@antirez With the Euro

[Steeve Morin @steeve]

@DorianWolffMD j'accepte l'étron si tu veux, pas la crotte

[Steeve Morin @steeve]

MERCI Ô SAINT JENSEN DE BENIR NOS PÉPITES DE NOTRE PAUVRE CONTRÉE FRANCE

mais putain Doctolib c'est pas une pépite c'est un lingot un peu de fierté bordel

[BFM @BFMTV]

🔴 ALERTE INFO
La plus grosse entreprise du monde mise sur une pépite française : Nvidia entre au capital de Doctolib
https://t.co/OJ9h6SsVzV

[Steeve Morin @steeve]

you know, I'd argue that in TP with DPDK that would probably be viable
aggregate is probably a few hundred GB/s

[David Hendrickson @TeksEdge]

So Anthropic claimed it ran a 26B AI model across fifteen $55 Android TV boxes.

Setup
📺 15× off-brand Android TV boxes
💾 4GB RAM each
💰 ~$55 each
🧠 Gemma4-26B-A4B
⚡ ~0.39 tok/s

Total hardware cost = ~$825.

The model was sharded across the boxes using pipeline parallelism, so each tiny machine only had to hold and execute part of the network.

Yes, 0.39 tps is sloooooooow.

But, that 26B model was actually running across a pile of cheap consumer boxes that were never designed to be AI servers.

🎯And Anthropic also said that Qwen3.5-9B agents managed to set up inference successfully on 8 of 9 hardware types they tested, including Macs, Jetsons, and RTX systems.

Source: Anthropic Frontier Red Team

⚠️Note that the 15-box demo was distributed inference. The self-replication experiments in the paper were separate.

[Steeve Morin @steeve]

> a German demoscene group
IT'S FARBRAUSCH GOD DAMMIT

[Arika @Arikalabs]

In 2004, a German demoscene group released a game called .kkrieger.

It was a full 3D first-person shooter with dynamic lighting, textures, sound effects, music, and monster AI.

The entire executable was 96 kilobytes. Instead of storing images and audio files, they wrote C++ and Assembly that generated the 3D meshes, textures, and sound waves mathematically in RAM when the game booted.
Today, a single website's "Accept Cookies" banner is 10x larger than that entire 3D game.

[Steeve Morin @steeve]

@ccgencc @feurlooO @CamiyKrab tant mieux, la competition est une bonne chose

[Steeve Morin @steeve]

@minviable_org 😍😍😍

[Steeve Morin @steeve]

@qubitium @AIatAMD mind if we pull this in zml?

[Steeve Morin @steeve]

JFC

[Max Weinbach @mweinbach]

My favorite Gemini 4 Argon experiment was converting Alan Wake 2 from x86 and Windows to native Arm64 on macOS

Argon write rust bindings for every x86 byte and op to Arm as well as Windows to macOS + convert to Metal Shaders and did it in a way that can be AoT compiled without Rosetta

It works, gets over 140fps on M3 Ultra. What I'm doing now is having Argon build out generalized bindings for more games, to see if this could be applied to get fully native games on macOS without developers needing to support it!

[Steeve Morin @steeve]

@sarddou

[Steeve Morin @steeve]

@tkanarsky @duckdb i wished i didn't have to do this

[Steeve Morin @steeve]

you know what screw it, all llm serving benchmarks suck. they don't measure the same thing and it's a huge mess.

we're writing our own and it'll use @duckdb because that how it should be.

[Steeve Morin @steeve]

@chevreuil ah oui, dommage

[Steeve Morin @steeve]

CA ALORS

[Le Parisien @le_Parisien]

« C’est révélateur » : Kiev détruit des data centers russes, l’activité de bots pro-Trump et pro-AfD s’effondre sur les réseaux sociaux https://t.co/En7D4A9KMI

[Steeve Morin @steeve]

@CamiyKrab ben désolé mais c'est le seul argument en fait ?

[Steeve Morin @steeve]

@CamiyKrab 3,6 Md de valo, 420M d'ARR le lépreux bon..