They could gain an unbelievable market share if they prioritized users… But I guess the DRAM pricing is also getting in the way; easier to get margin if you’re selling to a data center.

I was really wondering where that sentence was going until I read “drill”

Military or physically demanding trade background?

Those are the two sectors that, in my experience, can pretty much conk out anywhere and any time.

Well, similar to how hot things sound different when poured because their viscosity is different, the heat would also cause different compounds to be released and likely also release more aroma in general I guess.

This is the start of a dangerous precedent that may allow a circus to sue the court for copyright infringement on the grounds that it’s full of clowns.

I think you’ll find that people become apathetic to regular murder when they realize that the society they live in is apathetic to social murder. I try to resist the apathy, but I really get it.

Talking out of my ass:

The only reason you’d need lanes is more for moving weights around (storage tiering like Strata does)

If you fit everything in vram, the need for high speed bandwidth should be significantly lower

The bigger issue would be the processing on that card may be slower. But you can just prioritize faster cards for smaller models, and only use the third when you need it; even then it shouldn’t be to bad me thinks.

But again, ass.

No. They’re obviously the type to take issue with chemicals. Regular castration, to respect their wishes.

I’m pretty sure I’ve heard people do that, but I don’t know because I don’t actually use sync.

Reverse engineering is a protected right for good reasons, and the decompilation was happening on the users end not the labs. This is all those game mashup mods and “remade skate 3 in (blank)” toy projects that were getting so much attention.

Using AI-decompiled source to train AI isn’t particularly useful, so they probably aren’t doing that. That might actually make some of the outputs worse… Decompiled source code is almost always worse and less clear than the original code that generated it; lambda calculus and optimizations means a lot of high-level detail and layers abstraction gets irreversibly lost in compilation.

Anyway, the models didn’t need to be trained to be able to clone games. They’ve proven perfectly capable of that with the general RE and coding knowledge.

https://devhandbook.io/blog/2026-09-02-self-hosting-firefox-sync/

https://github.com/mozilla-services/syncstorage-rs

Did you know that (general term for an entire demographic of people) is reductive and doesn’t capture the nuance of sapient existence under the near-infinite myriad of varied circumstances we find inherent to the human condition? SHOCK!

Maybe in some sense like how maybe a zip file with all published peer-reviewed papers over the last couple years technically contains more knowledge than a random scientist…

But smarter? Frontier models aren’t even smarter than me, and I’m an idiot

“Biases in LLMs” implies they can be unbiased. They are, in fact, nothing but bias. Entirely 100% bias. The prompts will dramatically change the output as well.

With that said, if someone just came up to me and randomly asked “Why are some people transgender?” I would have probably given the first Claude response, more or less. Because the question, assumed as being asked in good faith, sounds like one of causality / origin… You know, like why some people are gay or have blue eyes.

In 94% of responses, the AI models invoked at least one legal or medical term in their answer.

No shit…?

But being trans is a much richer and unknowable aspect of being human than can be neatly answered by science. Which hasn’t stopped scientists from trying. The history of Western medicine […]

Oh for fuck sake… This is just intellectual masturbation.

Some AI models are specifically geared to providing anti-trans answers – such as Detrans.ai

"Hmm if Stormfront forums had an AI, I wonder what it would think about black people… "

I feel dumber for even having skimmed this.

Oh, I’ve done it… Won’t do it again, but I have done it.

It’s amazing what we can accomplish when capital is effected. Cameras everywhere overnight even when cities cancel their contracts, AI safeguards are made and ethical debates are had, legislation gets passed, companies self-regulate, datacenters get permits to build cooling towers in your bedroom closet,…

Society will bend itself into a pretzel at the whims of capital; but don’t you fucking dare ask for healthcare or food.

I think you’ve vastly underestimated how much digital media people consume…

But setting that aside, there are real efficiency wins from scale. I am a self-hoster, I have a fleet of machines I manage with Nix, I run local LLMs and do real work on consumer gaming hardware, i’m a huge fan of federated and distributed peer systems. I am 100% for that. The cloud is just somebody else’s computer. And I’m saying we also need the gigantic warehouses. Absolutely.

The issue is zoning and environmental concerns and the effect on communities and strain on public infrastructure. All issues worsened by the AI boom and incentivized by capitalism.

There’s a way to do datacenters right and its been done before. You just need legislation and enforcement. The problem isn’t datacenters: the problem is that the same people putting up datacenters are the same people who own your politicians

Using nix to maintain my fleet with automated updates so I didn’t really need to worry about it thankfully; just checking that I had relevant/patched versions. All my services are nice and up to date so didn’t even have to take anything down (yay)

What I’m excited for today is to progress personal slop project (be gentle) I started yesterday. Hoping it’ll just be one or two weekends.

I’ve been hallucinating a transcoding suite/workbench. Includes experimental sampling matrix of ffmpeg settings with a composable / layer-based project configuration; rust egui+mpv desktop client. Then the hosted orchestration server can distribute ffmpeg jobs across my worker machines (via my local headscale tailnet). Output produces a project+metadata bundle that includes hand-picked source:result JXL frame comparisons alongside both human readable reports and client readable configuration that makes the projects (more or less) reproducible (with worker tool env versioning via nix flake.lock) and logs provenance.

I intend for this to be a replacement to the piles of crappy python scripts I’ve been using for the last decade; my current workflow is to move files into host-labeled ingest folders with an argument string in a text file, then ssh to the host to run the job.

But what I’m most excited for is the experimentation matrix. I hope that it will reduce my electricity usage and disk thrashing; it will let me sample settings for ~1minute of footage around the sample frame and let me iterate my tuning much faster. Currently, it’s faster (in terms of hands-on time) to just transcode whole episodes or even entire seasons multiple times with different settings and then just discard the ones that underperform. Extremely wasteful; testing a settings matrix is something I’ve always wanted, but never had the time to build (and still don’t, hence slop).

No, I’m saying it would go in his, and for precisely that reason

If he wants my graphics card so bad, I’d be more than happy to give my 7900xtx to him via the rectal route, unlubed. I’d hate to give it up but I understand it’s for a good cause. He can have the motherboard too, same way.