mg

[mg @luonurus]

@Rainmaker1973 Reminds of Taxi 3 movie

[mg @luonurus]

adhd psychosis development

[mg @luonurus]

lil nvim plugin for searching sites of definitions. Also clipboard mode to pick blocks and throw into NN

[mg @luonurus]

@leftcurvedev_ Cute. What you used for rendering?

[mg @luonurus]

@DominativeOrder @neogoose_btw Why do you mean TCP is an overkill? What are the alternatives in your opinion?

[mg @luonurus]

@Matjaz19820 @SBerber84058 @rinegade Have 0 info what happened. Seems to me like it is deleted. Github is private, and slonana chart isn't loading anywhere. No idea how it is possible and what are the reasons

[mg @luonurus]

@Matjaz19820 @iamgub @rinegade idk. It reminds me of psycosis. Predicted that, wanted to stop somehow. Seen before, scary as hell

[mg @luonurus]

@iamgub @Matjaz19820 @rinegade He dmed me to try out his RPC, then I sent him my code in order to evaluate if it can run efficiently. Did not track but he was gone in like few hours after that. I thought he banned and scammed me. Some private setup. Just trying to survive from unemployment 💁

[mg @luonurus]

@iamgub @Matjaz19820 @rinegade he also dissapeared from github after I sent him some code...

[mg @luonurus]

@sudoingX @BrainOfMine Interested how 5090 (mobile) compares to ur old desktop 3090. Test for nemotron & qwen some time please just to see hardware comparison

[mg @luonurus]

@zeeg I think it requires clever agentic framework, but it is only possible with local models to spin whole night and not burn generational wealth

[mg @luonurus]

That's why Rust is great for fastfood projects, if you do not build something for ages.
- GPT etc. oneshots bindings for Python, perf increase is usually insane for some numpy mem traversal (3.14)
- After a program is finally compiling after prompt loop, it is fast (use Arc)

[Stefan @schteppe]

“The C++ standard library's default containers are a performance disaster. Not a minor inconvenience. Not a ‘it depends on the workload’ caveat. A 58x P99 latency penalty compared to Rust's standard library, on an identical order book workload” https://t.co/aDLYaGCPgn

[mg @luonurus]

@svpino did not check perf difference, but: https://t.co/iUbIkdCz00

[mg @luonurus]

@jonsflow @svpino @AnthropicAI I had some experiments with tool calling on extremely constrained evironment (4gb ram). It is very raw tho, I will polish and add coding skills as I will have time (love road bikes btw :) ): https://t.co/hAS95z6dNp

[mg @luonurus]

@4shpool Seems like that. Interested how beneficial ECC for video memory is, too broke to try out. Tried llama 70b on 2x3090 without adapter and qwen 27b (Or smth like that) on single. Perf diff was noticeable, but not critical. Early 2025, 4bit gguf, worked well for most tasks (data sci)

[mg @luonurus]

If I had resources to hire 500k devs, I would be uncomfortable to leak project details to foreign models and let them develop local agents to run overnight half of time

[sunny madra @sundeep]

“If your $500K engineer isn’t burning at least $250K in tokens, something is wrong.”

[mg @luonurus]

NixOS is pleasant in 2026, when LLMs got better in flakes.
Hate tf/torch not seeing GPU after building ffmpeg, hate download speed over pip.
Idk how it is so RAM efficient

[vitalik.eth @VitalikButerin]

My self-sovereign / local / private / secure LLM setup, April 2026

https://t.co/zmG8wrg7pQ