美国需要开源AI战略：中国开源模型崛起引发担忧

A new divide has opened up in AI. Open versus closed. Scaling of open source models. I think it's going down a very dangerous path. Open source models that are excellent should be used. Closed AI is what most people know. OpenAI's ChatGPT, Anthropic's Claude, Google's Gemini their most powerful models. They control the price, the rules, you pay for access. Open weight AI is different. The model can be downloaded, inspected, customized, run on your own infrastructure, and it's often cheaper.

China's deep tech was the wake up call, then z.AI's latest GLM model. Now moonshots Kimmy K three - China is taking over. They have a dominant share, no question. And the biggest players in American AI, they're now calling for open source AI too. But Washington's AI strategy is still focused on hardware. Protect the chips, build the data centers, add the power, all while China gains ground in the models, the world is increasingly running on.

The vast majority of humans on the planet will use Chinese models. Why? Because they're free. I'm Deirdre Bosa. America has a chip strategy for AI. Now it needs an open source strategy. China's open source AI momentum has Washington's attention. What is less clear is what it plans to do about it. So far, the U.S. response has centered on the risks. Looking at us, and we're looking at them. But we're leading now China substantially in AI, and we're going to keep it that way.

But it's hard to contain something built to spread with fewer advanced chips. Chinese labs, they had to compete on efficiency, smaller models, cheaper training, lower costs to run. And then they release many of them as open weight models, meaning developers can download and customize them, run them on their own servers. OpenAI and Anthropic. They have taken the opposite approach with their best models, keep them closed behind services that they control.

But once an open model is out, it's out. So if you restrict it here, you may just sideline American developers while everyone else moves ahead. Benchmark's Peter Fenton, one of Silicon Valley's best known venture investors, says that could backfire. There's some discussion that that might be a coherent strategy to limit access to open weight models. I think that will have the exact opposite effect. It'll disadvantage the U.S.

Disadvantaged because open models are the way smaller companies compete. Most startups cannot afford to train a frontier model from scratch, and they can't always afford to run their products on the most expensive closed models. Open weights give them a cheaper base to build on, to customize and deploy. So a ban or a broad restriction that might not actually stop Chinese models from spreading globally. It could just make it harder for American startups to keep up.

America has made this kind of mistake before. It invented the semiconductor. Then, it let most leading edge chip manufacturing move to Taiwan. That looked efficient until America realized that the technology powering its economy and military depended on factories in Taiwan, an island claimed by Beijing and at the center of U.S.-China tensions. Every single AI chip that Nvidia designs is made in Taiwan, and we need those in our cars and our phones and our military equipment.

Now, Washington is spending billions of dollars through the CHIPS Act to bring more of that manufacturing home. The lesson here was simple. Something this important shouldn't depend on one outside source. But chips are only the hardware layer. Open models are the software layer and may require a different playbook. More computing power for universities and startups, government contracts for American open source models.

Support for the security and software needed to run them. It is in America's interest to actually support open source and open meets. Right now, the closest thing America has to that is Nvidia. It's all being built in this country, basically, except for the open models, which is a real thing coming from China. Except for Nvidia open models which are world class. It's Nemotron models are open. So are the data and the tools companies can use to customize them.

Nvidia has also organized a coalition of AI labs to build more, but Nvidia is doing that because it fits Nvidia's business, one company acting in its own interest. That's not the same as the country having a strategy. And the pressure, for one, is not just coming from Washington, it's now coming from the customers. A backlash is building across corporate America. Companies have spent the last few years rushing to put their data and workflows into AI.

Now they're asking what they may be giving up in return. Palantir CEO Alex Karp, he put it bluntly on CNBC. In this country, at every single enterprise I deal with, these people are livid. They're like, I am paying for tokens that create no value. These people are stealing the weights and alpha of my business. Karp's argument is that a company's data isn't the only thing that matters. Its real advantage is everything it has learned about how to run its business, the connections between its data, its workflows, and its people.

When all of that flows through someone else's closed model, the company risks giving away the very thing that makes it valuable. By sending the data to a company that has extremely smart models, you are really giving up your business's recipe for them to copy. And then it has to keep paying to get that intelligence back. The closed labs Anthropic and OpenAI, they say their enterprise products protect customer data and do not use it to train their general models without permission.

But the concern is bigger than whether a lab trains on the data. It's about who owns what the AI learns about your business. They want to know they own the means of production. It's not being transferred to someone else. Satya Nadella, the CEO of Microsoft and one of OpenAI's biggest partners. He's now warning about the same thing. Nadella says companies need to control their own learning loop, the knowledge their AI picks up every time employees and customers use it.

So if they switch models, they don't have to start over. With a closed model. The company is renting access. The provider can change the price, can change the rules, or cut off access entirely. And that isn't a theoretical risk. Earlier this year, the US government ordered Anthropic to suspend access to its new fable and mythos models, customers and employees around the world. They lost access overnight. Anthropic, disabling access to its newest AI model.

After the government rang alarm bells about national security concerns. The restrictions were later lifted, but the episode showed how quickly a model someone is building on can just disappear if you don't own it. Someone else can decide when you're allowed to use it. An open model cannot be taken away overnight. Running an open model. It does take more work, GPUs, engineers, safety checks, but at least a company can run the model itself and keep control of its data and costs.

And this isn't only about companies. Governments don't want their most important systems dependent on a handful of American labs either. They want AI that runs inside their own borders, under their own laws on infrastructure that they control. Do you really want a repeat of the internet future where, uh, like all our data just lies in the hands of 2 or 3 companies? Or do you want a lot more control and privacy and sovereignty over that?

For years, the selling point of American AI was trust. Now customers are asking whether trust without control is enough. And that's what makes China's open model push so powerful. Now the US tech industry is pushing Washington to pick a side. For the last few years, open source looked like a China story. But now it's an American industry story. In late July, Nvidia CEO Jensen Huang used his first post on X to share an open letter to Washington.

The message open weight AI is a strategic asset. The letter wa

原文超出正文长度上限，此处截断——上游还有内容，完整版见上方「原文 ↗」。