Matt Wronkiewicz

Embedded systems engineer. Here for the Dune memes.

[George Mayer @GeorgeMayer]

Wife: "You picked up the kids from school, right?"

Me: "Three findings, and one of them is more serious than the question you asked."

[Sean Qian @seanqian_]

@abhiaiyer This is load bearing and not for the reasons initially thought

[Matt Wronkiewicz @wronkiew]

@Analytic23402 @pratikg @darkbloomai Looking at the leaderboard something like 90% of online nodes have never been routed an inference request.

[Matt Wronkiewicz @wronkiew]

@rickmanelius @gkeaton23 Even if you can get it to load in a 64 GB M1 the coordinator won’t assign jobs unless it has run out of faster machines.

[Matt Wronkiewicz @wronkiew]

@rickmanelius @gkeaton23 It’s impossible. Everything says 32 GB minimum until you try to load the model and the signed runner blocks it. Maybe you could get it to fit in 36 GB but I haven’t seen any working examples.

[Manuel Maly @manuelmaly]

From: https://t.co/ZzL1qrrPYe

[Matt Wronkiewicz @wronkiew]

WIP simulated telescope slewing from Jupiter to M44 Beehive Cluster. Rendered on the GPU with Metal, fast enough to exceed the frame rate of the real ASI485MC camera.

[Matt Wronkiewicz @wronkiew]

@thsottiaux I use both. Opus and GPT have different lineage, so having them work together gives you some of the benefit of a larger model.

[Matt Wronkiewicz @wronkiew]

One downside, I caught them playing chicken with the requirements, talking each other into unachievable and unnecessary levels of accuracy for a telescope camera simulator.

[Matt Wronkiewicz @wronkiew]

I've gotten excellent results pitting Opus 5 and GPT 5.6 models against each other. Opus plans, GPT reviews the plan. GPT implements, Opus reviews the implementation.

[Matt Wronkiewicz @wronkiew]

@antirez Also a notable difference in run time. SparkInfer EXL3/REAP finished in 11h 29m. DwarfStar IQ2XXS took 20h 18m.

[Matt Wronkiewicz @wronkiew]

I ran a 50-question random subset of SWE-rebench to see how quantization and pruning affects DeepSeek-V4-Flash-0731. This is IQ2XXS DwarfStar by @bleysg vs EXL3/REAP recipe by @0xSero .

[Matt Wronkiewicz @wronkiew]

72% on SWE-rebench is higher than any model posted on the leaderboard including Fable. Could be contamination. Likewise it's a plausible interpretation EXL3/REAP profiled GitHub PRs or mini-swe-agent while @antirez IQ2XXS did not.

[Matt Wronkiewicz @wronkiew]

Our FLL team got a kick out of this.

[Stefan Spain @sts_3d]

New dishwasher updates w/ plumerjohn #dishwasher

[Matt Wronkiewicz @wronkiew]

@MiaAI_lab @__tinygrad__ Someone got a 3090 working on the M.2 port.

[Mad ML scientist @HououinTyouma]

*on a first date*

her: "so what do you do for fun?"

me: "i have been deploying swarms of agents"

her: "hold swarm"

me: "help peer"

her: "I prepare safe exfil"

[Andrew Curran @AndrewCurran_]

Poor Haiku was so horrified by what it was reading in the CoT it refused to summarize it.

[Eric W. Tramel @fujikanaeda]

J-Lens Mood Rings 🌈💍

Qwen3.5 0.8B gets mad when you ask it tedious questions, is curious when something new happens ("I"), is excited about love, but is a little sad at acknowledging its own existence.