gnubyte

Software engineer and Creator

503CM, LLC IT Manager

Previously: Sr. Software Engineer @ Western Digital, L4 Engineer @Verizon

[gnubyte @gnubyte_]

@landontgreen @NVIDIAAI I mean as of today Mia labs has a one spark flash recipe you can use supposedly

[gnubyte @gnubyte_]

@Zorbrix *promotes build 69420 to production anyways*

[gnubyte @gnubyte_]

@GeneralistAI Interesting. This isn’t a public model release I’d assume. Wondering what the play is whether it’s a product or an acquisition exit strategy after private demos.

[gnubyte @gnubyte_]

@Studyhelp4you @nalinrajput23 No beer, we’re about to go hard lol

Coffee works too though

I had my nodeJS server for my MMORPG rewritten in C lang recently and it’s about 2 days in with about 90% coverage so far.

I feel like depending on your LLM it’s possible. Not quite sure I’d put the sparks on that yet

[gnubyte @gnubyte_]

@MiaAI_lab You really feel that way huh?

I felt the opposite like deepseek would hallucinate often even when clustered across two sparks.

[gnubyte @gnubyte_]

@TheAhmadOsman @Anthropic is great! It’s just that @DarioAmodei should slow down his PR and product direction changes and be a good, unchanging CloudAI solution for ease of use for 12 months.

My experience should not be changing day to day with the LLM product or the interfaces to it.

[gnubyte @gnubyte_]

This is some of the greatest reasons to invest in a DGX spark if you just want to follow the money.

If your subscription to big cloud AI is subsidized at $7800 that free lunch can’t last forever: we are in the YouTube Pre-Ads, pre-premium era of AI.

[gnubyte @gnubyte_]

I got tired of there being no windirstat on mac because its such a great tool and put this out there. Its open source - if you want to save some tokens or make some edits just open a PR and I'll merge them in
https://t.co/2f7AbbukmR

Native windirstat on mac basically.

[gnubyte @gnubyte_]

If you own a DGX Spark and don't want to provide an API key for Google or Duckduckgo or one of these, I've built a skill you can put in any of your openclaw or hermes agents and have it setup and build its own search index/engine and host it locally as part of the spark
https://t.co/YEp5hOim0B

[gnubyte @gnubyte_]

Deepseek's open source harness with a Mac OSX GUI so you can have a desktop experience with the Deepseek harness.

https://t.co/waEYrPeQUz

[gnubyte @gnubyte_]

@NVIDIARobotics I'm really excited - I want to try folding laundry with it. I feel like it will be hard if I just toss anything at it but I think with time or simpler tasks I can figure something out.

Comes in tomorrow and I'll trying it with the spark.

[gnubyte @gnubyte_]

@NVIDIARTXSpark @UnslothAI @ggerganov That is awesome! Glad this is being promoted.

Now what does it take to get one of those sleek sweatshirts LOL

[gnubyte @gnubyte_]

@ivanfioravanti Qwen 3.8 seems to be able to actually be useful. Deepseek hallucinated far too often in agent harnesses for me. Qwen stays at it until the jobs done.

Both great tech though. Can't believe how fast flash runs.

[gnubyte @gnubyte_]

@DavidOndrej1 @AnthropicAI handling of everything post Opus 4.8 has been very dissapointing.

They can be a fantastic easy cloud partner in America. Just be that.

[gnubyte @gnubyte_]

@Alibaba_Qwen Incredible technology.

[gnubyte @gnubyte_]

@ClementDelangue @huggingface hey! I'd gladly become a subscriber of HF if you guys just offered faster downloads.

That is practical value. Personally not really in the realm yet of the other features (yet) but that seems like a good gateway offering.

[gnubyte @gnubyte_]

@MiaAI_lab What stack are you running the GLM 5.3 on?

[gnubyte @gnubyte_]

@nvidia @LGE_Global @JensenHuang I've never known @LGE_Global for open source but I'd really like to be able to see a robot from LG that is competitive that a developer could order for $5k-20K here in the US

IMO that is the bot to beat because its accessible and doesn't have US's engineering overhead costs.

[gnubyte @gnubyte_]

Wondering how Z AI Coding plan does rate limiting on its max plan. I noticed that @Kimi_Moonshot k3 you can smack that plan around all day every day but it does take a very long time to execute (feels almost like 10t/s). Where @claudeai has let me run 70t/s roughly it feels like.

There is a window of effectiveness where if I'm going to fail, I'd rather fail fast and fix it on a number of tasks. I think Anthropic with options provided has done well with that but the watermarking is still so crappy.

Looking forward to trying Z AI's newest model.