More

steeve · 2026-09-16T21:43:38

https://x.com/harshagundal/status/2100044305536889015?s=20

> They were building in stealth for 2 years, I was building in stealth for 2 hours…

> Happy to open source Qwen-2.5-1B-RLCD, 5x faster on-device inference for JSON workloads that need to be type-safe.

flockonus · 2026-09-16T22:31:52

No question OSS is amazing, but this video is a satire at best. It doesn't take much attention to see the results on right vs. left side are significantly different.

Jev is not interesting if it's not "smart", a 1B param model is most definitely not smart.

nullbio · 2026-09-17T02:01:02

Jev has a 32k context window. I doubt it's a large model.

toasty228 · 2026-09-17T08:54:35

It can be smart and not knowledgeable, as long as all the data is in the question you really don't need much in term of general knowledge, it might be even beneficial to stay on track. It mostly depend son what you do with it. For example in the doom demo I don't think you need astra level knowledge to know "enemy = bad" and "enemy is at my left = turn gun left"

regularfry · 2026-09-17T08:43:49

Taking their spiel at face value, it's intended as a mimic of System One in humans. System One is also not smart.

toasty228 · 2026-09-17T08:50:48

In their own demo the left and right panels almost never agree on anything. So it's X times faster but doesn't tell you if it's still right

steeve · 2026-08-02T10:14:19

XLA uses MLIR very much, the interface to it is StableHLO, which is a MLIR dialect.

ndesaulniers · 2026-08-02T16:07:27

Ah, I'm behind the times.

steeve · 2026-08-02T10:13:30

> Plus, no need to build XLA using Bazel either! You’re welcome.

I don't get it, Bazel is incredible.

melodyogonna · 2026-08-02T13:05:17

Many people find it complicated.

fragmede · 2026-08-02T15:31:26

It is. The question is if it's worth the complication. The cheeky answer would be to say it is, but that's not universally true.

steeve · 2026-06-03T07:16:20

sidenote: it is possible to use Vulkan to map GPU memory to CPU space and even map it back to CUDA: https://x.com/steeve/status/2055042304344231978?s=20

steeve · 2026-04-05T14:27:55

consider it done

steeve · 2026-03-31T14:03:09

sadly, sandboxing is something that can't be upstreamed. this way, sandboxing is kept in zml instead of patching mesa.

as for nvtop, great program, but we missed a few features (such as sandboxing)

pstuart · 2026-04-05T07:53:25

It looks cool and I was excited to get monitoring for the NPU on my Ryzen AI 395+, unfortunately it does not show. NPU support in linux really seems to be an afterthought.

steeve · 2026-04-05T07:55:48

Weird, because we tried it. It doesn’t show anything?

We use the amdsmi to get metrics. I’ll investigate.

steeve · 2026-03-11T23:21:20

hi, i'm the founder of https://github.com/zml/zml, very happy with Zig

steeve · 2026-03-03T22:59:54

Misleading title. It's about using their AI service, not the editor.

steeve · 2026-02-14T14:16:43

we (ZML) have been back to following Zig master since std.Io was introduced. It's not that bad tbh. Also most changes really feel like actual improvements to the language on a day to day basis.

steeve · 2026-02-14T10:40:44

pjmlp · 2026-02-14T10:44:15

What is even that? Not bothering to Google for it, which shows how irrelevant it is.

dmit · 2026-02-14T10:47:08

Good call, for bad reasons.

> At ZML, we are creating exciting AI produ[...]