Shared in Technology@lemmy.world 9 days ago
0/2000
k
kreskin@lemmy.world 7 days ago
China now has better cars and better AI. When they take Taiwan they will control most microprocessor manufacturing. The US has squandered its post WW2 advantages and fiddled away everything via bad leadership, low standards, and our tolerance for industry grift.
demonsword@lemmy.world 7 days ago
When they take Taiwan they will control most microprocessor manufacturing
The don’t even need to, as they are evolving their semiconductor technologies by leaps and bounds for quite some time. They even managed to reverse engineer the latest lithography machines that the taiwanese host.
o
ofcourse@lemmy.ml 8 days ago
I know everyone hates on Meta but they just released an open source 800B parameter model: Muse Spark Open.
Their smaller 30B parameter model Muse Glimmer has also been out for a few weeks and is highly competitive with Qwen 3.8 27B.
So I hope we will start seeing more open source models from US labs especially in the sub-20GB category which can be deployed on daily use machines.
b
brucethemoose@lemmy.world 7 days ago
In case anyone forgot, Meta was the pioneer of open weight models, with Llama 1 released/leaked February 2023.
They literally set the standard. Everything in the open weight space is derived from the Llama series, its footprint is in newest Chinese models. The world was their oyster.
…And after one bad experiment Zuckerberg meddled with (Llama 4), Zuck fired the whole Llama team, hired a bunch of tech bros in their place, who proceeded to nothing for a good long while.
Muse 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 Its architecture (GQA + sliding window attention) is about a year or two out of date, but it… could be worse.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.
Same with Musk. And Altman. The U.S. AI industry was basically doomed from the beginning.
R
RedGreenBlue@lemmy.zip 8 days ago
Neat! This stuff should not run in data centers anyway. And it should not be gated and controlled by billionares.
v_krishna@lemmy.ml 8 days ago
Training these models still requires massive data centers, and using any of them that at all compare to frontier models for inference requires large numbers of GPUs
theneverfox@pawb.social 8 days ago
It does for now, but the methods for modifying an LLM are getting increasingly sophisticated. From blending models to doing brain surgery on them to fine tune the weights, there’s a lot of new techniques being developed all the time
There’s also recently been a breakthrough that cut the ram requirement to 1/6, eventually someone is going to figure out how to train models orderes of magnitude more efficiently
Especially once the AI bubble pops and no one is willing to dump billions into constant iterations
C
CosmoNova@lemmy.world 8 days ago
Open-weight is not the point in the AI race so this comparison doesn‘t tell us anything about who‘s leading in whatever metrics they‘re testing right now.
I wish models had to disclose their training sets so IP holders can sue for damages. Those would get into the trillions no doubt. Hell I wish LLMs never got this much out of hand. But I don‘t understand what they‘re getting at with this or why make it a „China VS the West“ thing.
eicker@lemmy.world 8 days ago
Open weight AI absolutely matters if the question is who gets to shape the AI ecosystem. A slightly better closed model wins a benchmark: a good open model can be adapted, audited and run without paying a Silicon Valley landlord. And yes, training data transparency should apply to both Chinese and Western labs. That’s not a contradiction: it’s the point. Still: FOSAI » OWAI » ClosedAI
C
CosmoNova@lemmy.world 7 days ago
I‘m not sure what gets to shape the AI ecosystem in the long run and I doubt anyone does but open-weights probably isn‘t it right now considering everyone used either Claude orChatGPT. I don‘t think open-weights matters all that much in the AI race right now. Seems like an artificial race to me anyway.
m
makeshift0546@lemmy.today 8 days ago
Who owns the computer? The models are irrelevant without the hardware.
Quazatron@lemmy.world 8 days ago
Because then Western AI companies can go to the governments, point to China and say: “see how advanced they are? See how they can apply that to the economy and military? You’d better give us money and not try to regulate us.”
That is the whole story.
C
CosmoNova@lemmy.world 7 days ago
Just to clarify: This is about western open-weight models being weak not closed models that are objectively more capable than anything open-weight. It‘s like saying China dominated Formula 2 racing when the rest of the world is focused in Formula 1.
N
NoLemurs@lemmy.world 8 days ago
You’d better give us money and not try to regulate us.
Actually, right now, the western AI companies desperately want regulation.
An unregulated environment is one where they’re forced to competing on price against the Chinese models and startups running open-weight models. All signs point to them losing that battle.
OpenAI/Anthropic’s wish list right now is a bunch of regulation to prevent the use of foreign models, and to require expensive to implement “safety” requirements that make it hard for startups to compete.
mannycalavera@feddit.uk 9 days ago
Trump (and Biden’s) legacy.
The__Guy_Fawkes@lemmy.world 9 days ago
Well thanks! Now your comment made me picture Trump and Biden kissing!
eicker@lemmy.world 9 days ago
Silicon Valley spent billions building AI tollbooths while Chinese labs made models people can actually download and run. Then Washington restricted chips, and China got better at doing more with less. It’s almost poetic: America tried to corner the future of AI and accidentally made efficiency China’s competitive advantage.
Em Adespoton@lemmy.ca 9 days ago
It’s worse than that even… frontier models now have guardrails restrictive enough that I can’t use them for legitimate work. And the ones that DO work have token prices at exorbitant rates, and chew through them like there’s no tomorrow.
Meanwhile, I can’t use download Qwen3 and run it on my own hardware, and it’s about as good as the last generation US model without the annoying guardrails.
Hmm… which model am I going to choose to use?
adr1an@programming.dev 9 days ago
Did you mean qwen 3.8 max? Or qwen 3 coder next? They are very different.