Random Libertarian Tech Lead

Shipping to prod in a $100M+ ARR unicorn startup w/ local AI.

4x DGX Spark (GLM 5.3 Flash), 1x M5 Max 128GB (Qwen 3.8 Flash Next), 1x RTX 5090 (Qwen 3.8 27B).

[Random Libertarian Tech Lead @someRandomDev5]

@kentcdodds Use the AI for what it's exceptional at.
Use yourself for what the AI sucks at.

https://t.co/kvSLNBZpW3

[Random Libertarian Tech Lead @someRandomDev5]

@kentcdodds Well, do the visual equivalent of spec driven development:

1. Sketch something out on paper.
2. Have an image gen AI generate an image from that sketch.
3. Have a VLM generate code from that image.

[Random Libertarian Tech Lead @someRandomDev5]

@vpocius @MiaAI_lab "decode speed" is not "task speed". A model that sacrifices quality in order to improve TPS often ends up finishing slower, due to how the quality reduction affects perplexity during reasoning. It also may result in less intelligent intermediate steps and tool calls, etc.

[Random Libertarian Tech Lead @someRandomDev5]

@bindureddy You can tell who actually uses the model versus who's just LARPing by looking at the responses and seeing who thinks "high tokens/sec = fast model".

TPS is like RPM on a vehicle. DS4.1 Flash has tiny little e-scooter sized wheels. Takes a lot of spins to get anywhere.

[Random Libertarian Tech Lead @someRandomDev5]

@Youssofal_ If you trust their WER mix for your use-case, then this is the rare situation where Europe is the open-weights lead lol.

[Random Libertarian Tech Lead @someRandomDev5]

@Raynos @loktar00 And btw, the better analogy truly is more like "taking transit and/or Ubering everywhere" rather than "leasing a car".

"Leasing a car" is more like "renting GPUs in the cloud". Have you seen the cost of that?

[Random Libertarian Tech Lead @someRandomDev5]

@Raynos @loktar00 This also feels like an itty bit of bias in favor of "renting vehicles rather than owning them" lol.

[Random Libertarian Tech Lead @someRandomDev5]

More:
* Consistency: You don't have to worry (as some do) that the inference provider will at some point "nerf the model".
* Uptime: The model runs without an Internet connection. If it stops working, the one who cares most (you) is also the one available to fix it. No wondering about fix ETAs.

[Random Libertarian Tech Lead @someRandomDev5]

@net_termina Some others:
* Education: You learn what it takes to stand up a model in production. You learn the tradeoffs involved in inference, learn to recognize when/how/why weird behavior is a quantization issue or inference stack bug.

[Random Libertarian Tech Lead @someRandomDev5]

@TSLA_Man @loktar00 Lol I didn't buy one, but I remember making the joke a while back that "the reason it's called RTX Pro 6000 is that it costs $6000".

[Random Libertarian Tech Lead @someRandomDev5]

@theo It's hard to view parts of this and take it seriously due to what seems to me like... naivety on the part of the one making the video. "Why can't we commit .env files" is a funny one, because I was doing that. They were encrypted with SOPS. Some orgs did the same w/ diff tools.

[Random Libertarian Tech Lead @someRandomDev5]

@levelsio @JeremyNoronha This is a little weird because UX design is absolutely something the models need significance guidance on.

[Random Libertarian Tech Lead @someRandomDev5]

@krunkosaurus Also, buying a single Spark is silly.

Sparks are like shoes: They're meant to be bought in pairs.

[Random Libertarian Tech Lead @someRandomDev5]

@krunkosaurus I have 4 Sparks, a 5090, and an M5 Max 128GB.

Each one does something well the others don't. Sparks host my biggest MoE LLMs. The 5090 is fastest for non-LLM AI (e.g., video, audio, images). The M5 Max is the only one that lets me use LLMs on an airplane with bad WiFi.

[Random Libertarian Tech Lead @someRandomDev5]

@jun_song Buying a single DGX Spark is kinda silly. It's like buying just one shoe. You're supposed to buy in pairs. That's the reason that expensive $1000 ConnectX-7 connector is non-optional. If you buy only one, you should think of that more like "half a set", not "a complete unit".

[Random Libertarian Tech Lead @someRandomDev5]

Was the idea of the hackathon to build games, or was it to build graphics engines? If it was to build games, then spending the entire time "building a raycaster by hand" feels like it's missing the point, and the vibe coder might've stuck better to the spirit of the meetup.

[Daniel Hooper @DanielcHooper]

Recently attended an in-person handheld game console hackathon. I wrote a raycaster by hand. Someone wearing a “I ❤️ slop” tshirt saw mine and said “oh I wrote a ray caster too!”. They showed me. It was a ray *tracer* and claude did all the work. The person didn’t know basic graphics terms. They couldn’t answer basic questions about how it worked, could only talk about their prompting harness. Sad how AI is ruining the craft and community for people that actually like to program and have spent their lives getting good at it.

[Random Libertarian Tech Lead @someRandomDev5]

@NeoAIForecast @DCRando2020 @TechMDAI @NVIDIAAI Yeah, using this one in Rust for GLM 5.3 Flash on 4x Sparks:

https://t.co/Nr6ux6BST5

[Random Libertarian Tech Lead @someRandomDev5]

@Dark__Modest @PawelHuryn Yeah, figured I'd leave in the typo just so you know I'm a human lol.