Initiative Seeks AI Lab to Build 'American Truly Open Models' (ATOM)

"Benchmarking firm Artificial Analysis found that only five of the top 15 AI models are open source," reports the Washington Post, "and all were developed by Chinese AI companies...."

"Now some American executives, investors and academics are endorsing a plan to make U.S. open-source AI more competitive." A new campaign called the ATOM Project, for American Truly Open Models, aims to create a U.S.-based AI lab dedicated to creating software that developers can freely access and modify. Its blueprint calls for access to serious computing power, with upward of 10,000 of the cutting-edge GPU chips used to power corporate AI development. The initiative, which launched Monday, has gathered signatures of support from more than a dozen industry figures. They include veteran tech investor Bill Gurley; Clement Delangue, CEO of Hugging Face, a repository for open-source AI models and datasets; Stanford professor and AI investor Chris Manning; chipmaker Nvidia's director of applied research, Oleksii Kuchaiev; Jason Kwon, chief strategy officer for OpenAI; and Dylan Patel, CEO and founder of research firm SemiAnalysis...

The lack of progress in open-source AI underscores the case for initiatives like ATOM: The U.S. has not produced a major new open-source AI release since Meta's launch of its Llama 4 model in April, which disappointed some AI experts... "A lot of it is a coordination problem," said ATOM's creator, Nathan Lambert, a senior research scientist at the nonprofit Allen Institute for AI who is launching the project in a personal capacity... Lambert said the idea was to develop much more powerful open-source AI models than existing U.S. efforts such as Bloom, an AI language model from Hugging Face, Pythia from EleutherAI, and others. Those groups were willing to take on more legal risk in the name of scientific progress but suffered from underfunding, said Lambert, who has worked at Google's DeepMind AI lab, Facebook AI Research and Hugging Face.

The other problem? The hefty cost of top-performing AI. Lambert estimates that getting access to 10,000 state-of-the-art GPUs will cost at least $100 million. But the funding must be found if American efforts are to stay competitive, he said.
The initiative's web page is seeking signatures, but also asks visitors to the site to "consider how your expertise or resources might contribute to building the infrastructure America needs."

I can do that

[locater16]

prinft("Bald Eagle Screech* USA, USA, USA, *Tariffs you*");
Return 0;

Thanks China

[allo]

Who would think that investors could be the driving force behind making things open? But with China showing that one gets to the top by allowing the community to improve on it, it seems that now even the people who usually are the force behind enshittification think about open source.
Truly American sounds quite a bit MAGA, but the great thing of them wanting open models means one can then tune the models not to be too "truly american".

Re: Thanks China

[iggymanz]

You're funny, DeepSeek isn't the top. It proved China can make a passable but lesser AI using Chatgpt

Re:

[martin-boundary]

You're cute. Denying the competition exists is a losing strategy. I'm making popcorn.

Re:

[allo]

Show me a better Open Model that is not from China. American companies keep their top model for themselves.

Yeah, Gemma is nice, but no match to Gemini. gpt-oss is ehm... just try it yourself :D.
And the latest llama was not the great hit Meta hoped it would be.

Re:

[iggymanz]

why do you have the artificial constraint of "open model" there?

Those who want the best AI don't care about that.

Re:

[allo]

The whole article is about open models.

And my post was about China driving America to also release open models. Even OpenAI released a open model a few days ago. Nobody would have expected that before.

And those who want the best AI care about that, because if the model is not open (as in open weight, not open source as the OSI definition for open source AI), they can never be sure how long they will be able to access it. Look at the other articles this week about how people are upset that GPT-4o was gone on

Re:

[allo]

Yeah, what is the problem with proprietary cloud services. I have no idea either!

Re:

[allo]

It's your choice what risks you take. But even APIs can disappear, deprecate models, change price or change ToS. Then you can accept the change or need to be fast to learn how to use local models when things go downhill at OpenAI, because you either still need to accept the new ToS to bridge the time or have a downtime until you managed to migrate your stuff.

Also you seem to have missed the GPT-5 debacle.

GPUs are outlandishly expensive

[gavron]

"10,000 state-of-the-art GPUs will cost at least $100 million"
You mean if I buy 10K GPUs they still cost $10K each.

Only the goverrnment -- with our tax dollars -- can afford this crap.

AI is a hoax. Let the "AI companies" spend their shareholders' money and leave taxpayers out of it.

Re:

[ndsurvivor]

Maybe they are including the cost of things like circuit boards, connectors, a building... etc?

Re:

[buck-yar]

$100m is probably just for the GPUs. Search says A100 costs $8000-$20000 each, H100 $25000-$40000. So they're probably talking about older A100s. The math only allows for $10k per gpu, but then there's the server for each one (motherboard is huge), cooling is a big deal as the heatsinks for each GPU looks gigantic. They need the fiber optic network system, storage, physical racks and buildings. Then each server of 8 GPUs takes 5kW to run, so can't forget about the enormous power bill that'll show up each mo

oops, not their

[buck-yar]

*they're

Re:

[ndsurvivor]

I'm not a grammar nazi. I double checked your figures, and damn!! you are right. No wonder NVIDIA is worth 4 Trillion on the stock market today! The numbers are mind boggling to me.

Re:

[buck-yar]

A lot of it is future expectations. With all this hype about AI, and Google, MSFT, Meta, Elon, etc all going all in on AI, Nvidia still only brought in $130b. Google brought in $350b, Apple $391b, Amazon $637b. In 2023, Nvidia brought in only $27b. So one has to ask, is this sustainable? Or was there an arms race between tech giants to not be caught without processing power and now that its built out, sales slow? All this money spent on AI, how does it exactly generate profit for companies besides Nvidia? O

The British are working on this too

[nickovs]

Over in the UK apparently they are working on English Truly Open Neurals, known as ETON. Not only will these neural networks be exceptionally well educated but they will also have impeccable manners, although critics worry that they may occasionally be rather elitist and have a superiority complex. None the less, it is expected that more than a third of British Prime Ministers will use these models, despite protests from the Labour Party.