I can’t believe the sentiment in the comments here.

…Y’all think spammers are going to label themselves as “AI generated” if that’s detrimental? They will not, and it’s not something that can be objectively detected.

Getting customers to moderate this is self defeating, as the whole point of spam is to get in front of buyers’ eyeballs.

Valve needs to hire staff to cull this junk, they can afford to, but they don’t want to. That’s the crux of the issue.

Not really, as inference costs are plummeting because of the architectures open models use.

Wait… what do they even need live internet access for in the training stage? Shouldn’t tool calls and such be cached, if not preprocessed and filtered, for performance?

I get there’s probably an “answer competition” stage, but it still seems like an inefficient idea even if one sets security aside.

I mean, if every “AI” company went belly-up tomorrow, if development of any new model was banned globally, no exceptions… it would still not go away.

Deepseek 4.1-tier models would be useful for decades, and cheap to run. Heck, I think the scaffolding around them still needs years of development to not be so half-baked.

You aren’t wrong. That sentiment is everywhere in the comments below, and highly voted.

I mean, this isn’t true from my perspective. All the Chinese LLM labs are still developing full steam ahead. Model/architectures are getting updated so quickly it’s a meme amongst inference framework devs, because they can’t even support one model before the next one is already out.

…They’ve always been thrifty, though. The difference is they don’t need a trillion dollar bubble to fund their modest, more sustainable dev/inference costs.

Like others said, it looks better next quarter than a bunch of new projects. Short term gain vs long.

But also, their expenses from all that hosting and usage are high. They’ve masked it creatively, but that only works so long.

Maybe they’re trying to make their balance sheet look less catastrophic in the short term (again, at the expense of the long term).

Whatever it is, that will be the underlying theme. Things must look good immediately, at the expense of sustainability.

They needed this before “AI.” As there’s some evidence social media was already ruining attention spans.

TBH I have felt some degradation myself. And I’m more social media skeptic than most because, well… gestures at being on Lemmy.

Most companies won’t do this, though. That would require worrying about the company’s health years out, rather than next quarter.

You said:

Imagine 10,000 emails an hour through your enterprise and 1,000 times It fucks up.

So I interpreted that as an implication of “90% accuracy.”

I’m sorry I misinterpreted you, but I’m not sure why you have to be so abrasive calling that out? “Shillsville?” ???

Yep. This is why Nvidia is preaching open weights literally every chance they get, too.

Still. It’s a better stance, even if for selfish reasons.

It doesn’t literally mean “90% failure rate vs 100%”

In the HLE benchmark, for example, we may be talking about 45% test score for MiMo Pro or GLM or something, vs 55% for Sonnet and 61% for maxed out Claude.

But further complicating that, those particular open weight models do better than Claude in one particular hallucination benchmark, meaning Claude/ChatGPT might be worse at RAG/extraction type workloads. It’s like this across the board, with different models trading blows in different areas.

By “90%,” he really means “it’s so close the better model depends on the particular application, workload, harness, sampling configuration” and such, and that none of these models are close to “100% accuracy” across the board.

One other thing: Anthropic, OpenAI, and Grok have all been accused of “dumbing down” models under high demand to speed up inference. If your application depends on them, and they are the only provider, that could throw a wrench in the workflow.

But open weights? Pick your provider. You get an API as consistent as you are willing to pay for.

In other words, that “100%” isn’t really representing the potential inconsistency one gets being locked to a single provider.

Yeah. I see this attitude on Lemmy a lot. In Linux circles in particular, but elsewhere too.

No one is perfectly insulated from the rest of the world, as much as they’d like to think they are.

I mean, no. It will not. Some neat features or hacks may be LLM assisted, but it isn’t a miracle.

Besides, we can’t even convince people to switch to Linux, even with the RAMpocalypse in full swing. The vast majority still run Windows (or OSX/iOS/Android). How can we get even a fraction of them to hack a PS5 or something like that?

The other side is… what’s even out there to hack?

Old PS5s? Old PCs? Old Macs? Old servers? All these are skyrocketing in cost, and the instructions/features they physically support are aging. The only thing that’s plentiful are old smartphones, and you can already turn those into Linux boxes.

Hell, the RTX 3050 is still expensive.

My 3090 costs more than it was new… that is messed up.

Also, on open models: I can run Deepseek V4 Flash at 20 tokens/s on a single 3090 desktop, a Ryzen 7800, and 128GB RAM, with measurably low quantization loss. And 4.1 once the engine updates to support it.

That was eye-opening to me. Thats faster than some cheap APIs, and higher quality too. Once companies figure that out, what are they gonna do with all those idle datacenters?

Dev time is spent on mostly-idle GPU instances too. Code, test, doesn’t work, edit, test launch again. As one sometimes needs the full environment to even test whatever they’re ultimately trying to deploy.

And yeah… like I said above, I’m more afraid of actual utilization under load. Some deployments are so inefficient it makes me want to pass out.

I’d be more interested in efficiency under load.

There is evidence that some training runs at bigger companies are hilariously inefficient, or even that they run busywork to meet contract quotas. And that’s not to speak of the questionable efficiency of US “frontier models.”

And I’ve seen some staggering inefficiency even just deploying stuff, like “running Ollama on an H100 box” levels of ignorance. It boggles my mind. The heck are they even paying their employees for?

It’s a mixed bag TBH. The issue is that Qwen 3.8 27B is so good.

It depends on the quant you are running though, and the way you are running it. If you have 16GB or less VRAM and a generous amount of CPU RAM, and you are running a regular GGUF instead of something exotic, Qwen next is probably better, with the right hybrid inference engine.

LLM weights aren’t code; the engines that run them are completely separate, so they don’t have traditional “backdoors.”

It’s possible some elaborate prompt trigger was trained into the weights, but there’s never been any evidence of that, for years. I’d say worrying about that is paranoid unless you work at the DoD or something, and some full-weights finetuning would scramble it anyway.

In case anyone forgot, Meta was the pioneer of open weight models, with Llama 1 released/leaked February 2023.

They literally set the standard. Yes there was GPT-J before it, but basically everything in the open weights LLM space is derived from the Llama series, its footprint is still in newest Chinese models. The world was Meta’s oyster.

…And after one bad experiment, an experiment Zuckerberg allegedly meddled with (Llama 4), Zuck fired the whole Llama team, paid millions to a bunch of tech bros in their place, who proceeded to nothing for a good long while.

Muse Glimmer is obsolete, IMO.

Muse Spark is… not terrible? Its architecture (GQA + sliding window attention) is about a year or two out of date, but it… could be worse.

But this is just a reminder to anyone who thinks Zuck won’t find some way to feel insecure and fuck Muse up.

He will.

It is what he does. Zuck caused incalculable damage to his company, single-handedly, and he will do it again with zero repercussions.

Same with Musk. And Altman. Not to speak of Microsoft vacuuming up promising startups into a black hole. The U.S. AI industry was basically doomed from the beginning.