Open Source AI Better for US as China Will Steal Tech Anyway, Zuckerberg Argues

Meta CEO Mark Zuckerberg has advocated for open-source AI development, asserting it as a strategic advantage for the United States against China. In a blog post, Zuckerberg argued that closing off AI models would not effectively prevent Chinese access, given their espionage capabilities, and would instead disadvantage U.S. allies and smaller entities. He writes: Our adversaries are great at espionage, stealing models that fit on a thumb drive is relatively easy, and most tech companies are far from operating in a way that would make this more difficult. It seems most likely that a world of only closed models results in a small number of big companies plus our geopolitical adversaries having access to leading models, while startups, universities, and small businesses miss out on opportunities. Plus, constraining American innovation to closed development increases the chance that we don't lead at all. Instead, I think our best strategy is to build a robust open ecosystem and have our leading companies work closely with our government and allies to ensure they can best take advantage of the latest advances and achieve a sustainable first-mover advantage over the long term.

Re:Yes, steal

[LazarusQLong]

no, this isn't always about prejudice, it is about truth. I had a prof, MIT educated, who worked in this country for 45 years, then retired back to China, taking all his hard drives with him from about 30 of those years. Is that 'espionage?' well, kinda. But not what we in the US expect. See, apparently after xx number of years we start trusting people, this guy worked on missile re-entry problems, and after he had been a prof for some time, he was investigated, and got a high level clearance... but his loyalties always were with his home country, he never intended to stay here, and he brought all his valuable research with him back to China. All, classified research.

This is not a one - off, sleeper agents are all over the USA, our laws and culture is all about open-ness. Maybe you had heard of the Chinese employees at Los Alamos who were apprehended back around 1999 carrying bunches of hard drives containing classified data about to board aircraft for China? The time scales that the Chinese operate these types of operations on are nothing like the timescales that US persons think reasonable.

I myself have had grad students that were pretty stupid and up front about their desires to design nuclear warheads to destroy the USA with. Yes, they were Chinese. The Germans, French, Italian and other grad students I have had never were that.

Sure, there is racism I am sure. But, that sword is sharp on both sides... and, just as importantly, in this case at least, there are a lot of chinese sleeper agents here.

Re:

[Anonymous Coward]

People who bleat on about racism (like the GP) are only trying to disrupt a conversation, they're too unimaginative to think it's even possible for foreigners to have goals and interests that don't match those of a liberal, open society. They implicitly believe that those living in the open society must oppose the closed society from which the foreigner comes, and entertaining such thoughts must be evidence of racist thinking.
Either that or the poster is just trolling for idiot internet points.

Prejudice and Money

[Roger W Moore]

no, this isn't always about prejudice, it is about truth.

I suspect that in this case it is more about money and since Meta seems to be behind in the AI race having all their competitors Open Source everything is good for them. They are just leveraging prejudice against China to achieve it.

Re:

[Hadlock]

> since Meta seems to be behind in the AI race

That may have been the case 2 years ago but middle schoolers are reliably creating GPT 3.5 level LLMs reproducing steps outlined online and using freely available data sets. The only barrier to entry these days is time and about $400 to rent the A100 and/or H100 for ~72 hours from -pick your favorite GPU as a service provider-

Given that every middle and large corp is working on their own LLM I don't think facebook having GPT 3.5 level LLM is "

Re:

[LostMyBeaver]

If someone gets caught or makes statements about warheads, this is intentional.

The most effective weapon against the US is FUD. Build a crappy aircraft carrier for as little as possible and the US will waste trillions while blocking shipyards while China makes bigger gains in development.

Get busted with a hard drive? You chose to. It costs nothing to use data safe havens to send the data electronically.

Yeh, China plays the long game. No country with 4 year presidential terms will ever have the agility to co

Re:

[narcc]

China is winning in Taiwan [...] It takes 10-20 years to execute a plan like that

Obviously not.

Re:

[jythie]

Heh.. with regard to long games.. why do you think there has been such a big movement lately to 'dismantle the deep state'. The idea that the federal government should be flushed very few years and be personally loyal to the president aligns really well with the needs of China and Russia since it both gives a single authoritative person you can negotiate with AND kneecaps the contiguous institutions that DO think long term.

Re:

[jythie]

That isn't really exclusive to China though. I think the main difference is that when it happens and China is involved, peopel go 'China! Communism! Inferriors!', while when it happens with Americans or Europeans we just write it off to capitalism and a mobile workforce trying to get the best deal from their next, more lucrative employment contract.

Re:

[LazarusQLong]

I don't disagree with everything you have said here, but I do disagree about your comment that appears to claim that the US also plays the long game... Working in the federal government, we are given a 'strategic vision' every 5 years. it is a new and not derivative document. that is not the 'long game'. as far as I can see, the 'long game' played by China is in terms of generations.

Re:

[Growlley]

but they had to invent gun powder for you first ,

Forget AI, let China sabotage itself with it

[Anonymous Coward]

So-called 'AI' is crapware trash. I propose the U.S. forget about it for anything serious (it is, after all, amusingly bad, so okay for entertainment purposes), and let China and other idiot countries waste their time, money, and resources on it, sabotage themselves with it, America comes out on top.

"Hey, come back here with our... trash?"

[Pseudonymous Powers]

Yes, we wouldn't want our economic rivals to steal the massive benefits of [checks notes] burning ten terawatts of electricity to power a few hundred artificial fluent liars.

Re:

[gweihir]

Well, here is an idea: Give all that AI trash to China and stop using it ourselves! That might create a real advantage for us.

Re:

[jythie]

Even worse... the currentn crowd of ML people believe actually trying to understand things like the brain or speech or language are 'hacks'. Ignorance is seen as the correct path, since knowledge just corrupts your design.

Ole Zuck jealous!

[io333]

Umpteen billions spent by metabook results in nothing and some whippersnapper kicks their ass. Jealous bitch!

I believe him

[Rosco P. Coltrane]

If anybody know about stealing business ideas and customer data, it's Zuckerberg.

I don't

[martin-boundary]

Large AI models are not open in any meaningful sense. They are 99% data and 1% glue (aka architecture). The "open" part refers to the glue.

True open source requires *everything* to be made available so that it can be fully replicated by everyone. That's why in a true open source project, you get makefiles and support libraries and instructions for installing everything else that is needed.

What Zuck wants is to control the AI tools and infrastructure. By dumping pretrained models into the community he's

Re:

[Visarga]

Will you retrain the model on your GPUs? I didn't think so. The model is open weights. The source code almost doesn't matter at this point, it's 300 lines of PyTorch. But even a trained model is not so closed as you think. It can be finetuned on a much much smaller computer, with much less data to do anything you need in particular. It can be used to generate training sets for your small 7B models you can use at a lower price point.

Re:

[MachineShedFred]

Isn't it interesting that one of the largest tech companies in existence, who is demonstrably behind in all-things-AI, would advocate for open sourcing AI.

This is a not-so-clever use of jingoism to thinly veil his own selfish purpose.

China Will Steal Tech Anyway ...

[Savage-Rabbit]

Open Source AI Better for US as China Will Steal Tech Anyway, Zuckerberg Argues

That's a defeatist attitude, the US should get busy at getting better at counterespionage.

Re:

[Visarga]

It's not Zuck's business do do counterespionage. And China already has good enough models even without using LLaMA. It serves no purpose to close the model, will only defeat the American advantage of open creativity. Since it's too late to restrict this tech from China, now we better make it so widespread that it ceases to be a differential advantage.

It's about accountability

[Miles_O'Toole]

Zuckerberg: "...stealing models that fit on a thumb drive is relatively easy, and most tech companies are far from operating in a way that would make this more difficult."

Me: "Hey, Zuck, would a compulsory 10 year prison term for a tech company's board of directors help with that?"

Asking for a d-bag.

Re:

[Visarga]

Haha, he was dragged more than once in front of the Congress. They can't touch him, especially that he has his lawyers and deep pockets. And he didn't do anything illegal. If you want to change that, change the law first.

Just try stealing something from Meta

[Required Snark]

I'm sure that Zuckerface will go "no prob, dude, we're cool". He won't call the cops or sic $5000/hour lawyers on your ass. Just tell him it was bound to happen anyway and he'll let bygones be bygones.

Re:

[drinkypoo]

What are you going to steal from him? All the stuff you want is open sourced anyway. Facebook is a great OSS contributor and they've given away multiple fully-fledged and useful projects.

There are lots of reasons to be made at Zuck and Faceboot, but their stance on OSS contributions is not among them.

Re:

[Required Snark]

Wow, someone should take you aside and explain this thing called a "joke" to you. I don't think that I'm up to the task.

Obviously you didn't understand my sig either.

Re:

[drinkypoo]

It simply was not amusing.

You're just not that funny.

Re:

[Visarga]

Yes, remember React, PyTorch, and open designs for datacenter? All open sourced from FB. And now LLaMA. I used many of these products and found them well designed, a pleasure to work with, especially compared to Google's AngularJS, TensorFlow and lack of open sourced models until very recently, now they have a decent small open model. I'm wondering what kind of company culture makes such amazing software and tools?

He's not wrong about the models

[migos]

Yes it's true that China will be able to steal the newly trained close-source models. However, open sourcing it allows the entire SaaS and IoT industry in china to have access to some pretty powerful models. The stolen closed source models are probably harder to come by and productize. He's leveling the playing field for the entire world to prevent MSFT/GOOG/OpenAI from becoming too powerful. And yes he's definitely helping China.

Re:

[Visarga]

I prefer a level playing field, AI should not be an advantage. We need to make it common and easy to use so money doesn't accumulate in just a couple of companies.

yes, but

[groobly]

Yes, but even stealing requires diverting talent to do so. Talent that otherwise could be doing more productive things. Zuck is just a self-serving naif.