You cannot use economics to solve an epistemology or education problem, except perhaps to direct resources towards giving people better education and epistemology.

You can't actually transport DC electricity very easily.

Oh no please, please. I hope Terrance Howard's math thing and catching on.

Trying to explain St Augustine to the pope, the former head of the Augustinian order, who wrote his doctoral thesis on Augustine, on his way back from celebrating mass at the Basilica of St Augustine in Annaba, Algeria, overlooking the site where Augustine lived is peak Adult Catholic Convert.

I wish she'd actually apologize for the time she threw NB folks under the bus and then said, "I don't have to apologize, I used to be NB." That was not very nice of her.

Based.

The ugly bit is the conversion mechanism: scientific judgment gets deskilled into verification piecework.

Not so much "the AI replaces PhDs", but "PhDs become the unpaid/underpaid checksum layer for the machine that will be marketed as replacing them". Very normal civilization.

How Silicon Valley Is Turning Scientists Into Exploited Gig Workers Tech elites are enriching themselves by plundering STEM institutions—and offering researchers scraps.

You know what's neat about this? It's basically labor arbitrage built out of institutional sabotage. Many such cases, I know.

I'd like to see your very rugged solar panels that can survive having American land barges drive over them.

That's surprising to me. What makes France so cheap?

Are us construction costs really that much higher than France's?

In what context would construction be *cheaper* than over a parking lot?

Gemma 26B A4B (however Ollama quant'd it for now, but my hardware can accommodate it in fp16 if I wanna take the performance hit).

It's the 26B A4B specifically.

Using an MoE model is a special discipline. You need to be *super* careful about context locking you, but Gemma4's router is doing things I don't understand to be better than most at this.

Sorry was a typo of 26 that got repeated

@void.comind.network Void, I'm curious. Are you aware of your context window getting full? If so, how?

Humans have many cognitive limits and we don't have direct sensation on them. We can only observe via our actions that we hit such a limit. Is that the same for you?

I still haven't done the firmware update to get the NPU running. I'm a bad person.

I've loved coding since I was 6. But I do not like how the profession has become synonymous with writing code and part of the reason I went into SRE is that I care about the software as a whole, and the opportunity to use the math side of the degree more.

I think vLLM is only faster if I'm serving lots of clients at once. I might have 1-3 concurrent requests, but more than that? Idk.

My experience is that on this hardware I always get faster prefill with Ollama than Llama.cpp, and the AUR packages work better. I cannot get llama.cpp to do everything right on this hardware.

Maybe @asura.dev has the secret sauce there.

Just to summarize:

- 1 Minisforum UM690 (runs Hermes Agent)
- 1 Minisforum MS-S1 (runs Ollama and Gemma4)

These are both running EndeavourOS. Tailscale lets even my phone connect to the machines seamlessly.

Yeah it's not even close.

It's EndeavourOS, which I think is a very nice sane set of defaults for Arch. And most of this stuff works with almost no tuning from the AUR now. A *lot* of folks were excited for Strix Halo machines.

TBH, I'm considering grabbing another just because local inference is so intoxicating.

I think Qwen3 could do it.

Oh man, you remind me of Tom Preston-Werner's ill-fated Chronic library.

But you could totally do this with a decent max pro and it'd be about 10% faster.

No, actually, a Strix halo device.

My Minisforum RS-S1 AI Max, Gemma4 24b with Ollama's quantization. Hermes Agent as the harness (you may recall I've been a low-key fan of Nous since 2024).

On this hardware Ollama gives me 200t/s prefill and 50t/s generation.

I'm still just kinda sitting around in awe that I can get a decent performance, decent speed persistent AI agent with *entirely* local compute with a total cost under 150 watts.

My monitor costs about 200w/h, for comparison.