dave white

energy, hardware, and llm psychosis

[dave white @lmsorrydave]

I honestly don’t know who keeps paying for my Netflix subscription

[dave white @lmsorrydave]

@regsky_

[dave white @lmsorrydave]

@regsky_ Given the precedent of US labs appropriating copyrighted works for their own use, it's almost certain that they are. And what's a little IP theft between friends

[dave white @lmsorrydave]

@nisten So what the trick to make offload work quickly? Ddr5 and pcie5? Is it faster to use RAM or NVMe in this use case? Thx

[dave white @lmsorrydave]

@plotarmordev How slow would it be to offload to 8 channel DDR4?

[dave white @lmsorrydave]

@ParkerThayer great attribution work. It's actually funny how sloppy they were here

[dave white @lmsorrydave]

@Reederey It's a no lose situation. Glm is already so good, can't wait to see what deepseek brings

[dave white @lmsorrydave]

@emollick To me that doesn't matter that much. There's a million dumb tasks that I have to do, and a competent open weight model will do a more thorough job. I don't need Einstein to help me shovel shit

[dave white @lmsorrydave]

I can attest to this recipe from @MiaAI_lab. It's much faster than before.

Got it to help me fix and reconcile three years of QuickBooks in about 4 hours.

They can pry my sparks from my cold dead hands 😂

[Volatile Markets @volatilemarkts]

Stop what you’re doing and look at this: (2) Sparks for open-source, frontier-level AI at home!

No subscriptions. No API rate limits. No hourly caps.

Raw benchmarks are placing this recipe right above the entire landscape—trailing only absolute giant frontier models (fable and gpt 6) and I do not make that statement lightly.

By ditching brute-force hardware scaling and moving to an optimized stack (EXL3 4bpw on 2 nodes vs traditional heavy FP8 on 4 nodes), Mia Labs just proved that clever engineering is the ultimate margin expander:

Prefill: 1,554 tok/s (vs 961 tok/s)

Structured Decode: 67.9 tok/s (more than 2x faster)

Context Pool: 850k to 1.85M tokens

Quality Loss: Essentially zero (0.0245 KLD vs teacher)

This is what monetizing compute actually looks like. Can this even be real? This is the one.. @Zai_org can you start getting @MiaAI_lab early access to models for zero day support on GLM 5.5 and 6? I need to schedule days off..

#thisismysuperbowl #fantasyfootballisoverrated @NVIDIAAI

@MiaAI_lab https://t.co/JEipuzueuu

[dave white @lmsorrydave]

@zainhas That's pretty compelling, but if you use Codex do they pipe all your secret sauce back to openai?

[dave white @lmsorrydave]

@imost I switched back. Qwen is still faster but glm is just unreal for two Sparks

[dave white @lmsorrydave]

I used to have to ask Claude to help stand up new models on main cluster... Now one of my Qwen 27b workers just does it. What a time to be alive

[dave white @lmsorrydave]

@GPTWare if it's on sparks, qwen is better for responsiveness. glm is exceptionally smart, but it gets really sluggish as context grows. I traded a little quality for much higher productivity by using qwen/next.

[dave white @lmsorrydave]

@MiaAI_lab the two sparks recipe is really good. thx

[dave white @lmsorrydave]

@MiaAI_lab I'm running this. It's great

[dave white @lmsorrydave]

I was wondering where the gpu modding scene was...
I love what these guys at GPU Lab are doing, looks like this guy Greg has been at it for years. Very cool!

[Greg Sky @regsky_]

48GB 4090 water blocks available now from https://t.co/CCxcZigEKU

if you've been holding out because the 2-slot air coolers are loud, you can now use them on your own custom loop's. Video review and install guide here: https://t.co/zQix5slc6j

[dave white @lmsorrydave]

@jensenloke @Tech2Wild @QwenDevs Yep, I'm trying it out now. Thx

[dave white @lmsorrydave]

@stepbystepnomad @sudoingX @turboderp_ @superalesha notifs on. thanks :)

[dave white @lmsorrydave]

@stepbystepnomad @sudoingX @turboderp_ This is post is tremendously helpful. Thank you

[dave white @lmsorrydave]

@runsonai @MiaAI_lab do you like it better than glm5.3-flash?

[dave white @lmsorrydave]

@MiaAI_lab how would you compare this to glm5.3-flash using your EXL3 2x spark recipe? I don't want to lose the quality from glm, but I need something faster as my orchestrator. glm gets really bogged down >200k context