Joa Ebert

• SVP Technology & Architecture @ltnglobal • Previously @maketv, @defrac_ and @audiotool • Probably not the @joa you mentioned in your tweet

[Joa Ebert @joa]

It's becoming increasingly annoying that https://t.co/QF2rHtMIGU hosts throttle their watts...but SGEMM moved to 88% of cuBLAS thanks to recent compiler changes and better/more auto tuning. Basically, any literal can be autotuned now, like the U in loop<unroll U> :)

[Joa Ebert @joa]

@PotluckAI The “every test, every card” section goes up to tg2048 on supported platforms. That’s with Qwen3.5-4B-Q4_K_M. See https://t.co/h4CtHEJWtB

Anything in particular you are interested in? The real world benchmark replays actual agent turns.

[Joa Ebert @joa]

https://t.co/GzRAACwsiY runs @OpenBMB MiniCPM MoE Q8_0 pp128 at 6x of llama.cpp 😅

Several compiler optimizations landed in the past two weeks and inference certainly benefits. Also faster on 3090, 4090 or 5090 cards now.

[Joa Ebert @joa]

@robnet @mrdoob It has always been like this. I write assembly, I am better than you. I write C, I am better than you. I write Java, I am better you. I write JavaScript I am 🤪. Joke aside, you give up control, you get leverage. That’s how it works and it’s no different with AI.

[Joa Ebert @joa]

Rocket jump 😍

[Tesla @Tesla]

Do a kickflip!

[Joa Ebert @joa]

And by the way 🎤🫳 : https://t.co/ItA2aITrDR

Phobos has now a home that's a bit more polished than a bare github repo :)

[Joa Ebert @joa]

👀
Integrating @OpenBMB 's latest drop into phobos-lang; I'm also pretty sure the elastic cache would be very efficient on small devices.

[Joa Ebert @joa]

@golang Hey folks, are you aware https://t.co/lvEfCGrr99 is having trouble? This will hang forever: wget https://t.co/ylPvTovyH7

[Joa Ebert @joa]

It also included a bit of uncanny existential crisis: "are we machines, or are we dreaming?". You can listen to machine_heart.js, a purely functional song, at https://t.co/xlwK11gfbm

[Joa Ebert @joa]

Æon learned to talk. My purely functional audio engine now has a formant speech synth and a vocoder. No samples. Just a (very basic) model of a throat as a function of t.

[Joa Ebert @joa]

Then I asked Opus 5.5 for a track. It wrote the lyrics as phonemes on a beat grid:

S IH S | T AH M | AA N | L AA IY N

That's "system online".

[Joa Ebert @joa]

Is it better than llama.cpp? Absolutely not. llama.cpp runs everywhere, and I optimized for one card hardly anyone cares about. But for that card I can tell you exactly where the time goes.

Blog Post: https://t.co/Loe7PMVCcV

v0.1.1, Windows and Linux: https://t.co/mxXidHTMDd

[Joa Ebert @joa]

My GPU is from 2019 and has 8 GB of VRAM. It now runs a 35B MoE that doesn't fit on it at 55 tok/s.

Every kernel is compiled at runtime from Phobos, a toy kernel language I wrote to learn what a kernel even is. No cuBLAS, no CUDA toolkit. Just the driver.

[Joa Ebert @joa]

No knobs. Phobos sizes its expert cache by free VRAM and the CPU computes the misses. When the desktop grabs 2 GiB mid-session, it keeps ~75% of its decode speed. llama.cpp at its fastest setting keeps 30% (-ncmoe 36 survives it, but runs slower all the time).

[Joa Ebert @joa]

This is @typesafeai #jev on the decks. I clearly need to tweak the context more. All parameters / decisions are driven by the model. It's similar to Jev playing Doom with a decision tree of what to do.

[Joa Ebert @joa]

Making launches where the compiler cache is cold more entertaining. #phobos #mlir #llvm

[Joa Ebert @joa]

@mitsuhiko The vast amount of generated output.

[Joa Ebert @joa]

I added @OmarchyLinux mode to Phobos. Because it's that's what you do in 2026 🤘(I had to use a tiny model, as otherwise the vid capture wouldn't work 😅)