AI总结
-
核心结论：特朗普政府对人工智能采取的监管手段从“名义上的自由放任”转变为“武断、不透明的强制预审”，这种做法不仅损害了美国科技公司的创新能力，还可能导致美国在AI全球竞争中丧失领先优势。
-
结构逻辑：
- 现状概述：指出AI模型发布受政府“事实上的许可制度”管控，行政干预日益激进。
- 政策背离：对比特朗普行政命令的“去监管”叙事与实际执行的“管控”做法。
- 实操困境：描述OpenAI和Anthropic等公司在模型发布审批中面临的混乱与不确定性。
- 多方博弈：探讨监管对研发成本、模型变现窗口及中国竞争对手（如DeepSeek）崛起的影响。
-
关键信息：
- 主要公司：OpenAI（GPT-5.6 Sol）、Anthropic（Mythos 5）、Google、Microsoft、DeepSeek。
- 关键人物：唐纳德·特朗普（总统）、J.D. 万斯（副总统）、霍华德·卢特尼克（商务部长）、萨姆·奥特曼（OpenAI CEO）。
- 行业趋势：从开源模型向受控部署转型；模型“蒸馏”技术应用；监管导致企业研发投入决策动摇。
-
问题与对策：
- 问题：AI治理如何平衡安全与国家竞争力？政府是否有能力实施有效的技术管控？
- 回答：文章指出当前“类牌照”制度缺乏明确标准与技术支撑，阻碍了行业发展，引发了企业对转投中国开源模型的担忧。
正文
The administration now controls who gets access to the best models
Illustration: Fortunate Joaquin
AMERICAN LIFE is full of things that are billed as voluntary but turn out not to be. Tipping. Joining the local parent-teacher association. And now getting permission to release your cutting-edge artificial-intelligence model to the world.
On June 26th Open AI said that its newest model, GPT -5.6 Sol, would be restricted to a handful of “trusted partners”. The same day, America’s government eased export controls on Mythos 5, another advanced model from Anthropic, a competing lab. But that, too, would be available only to a pre-approved list of American entities. All of this marks a haphazard revolution in America’s governance of AI. “In a matter of weeks”, observes Dean Ball, a former AI adviser in the Trump administration, “ US federal AI policy has gone from implausibly libertarian to increasingly draconian and opaque.”
The Trump administration assumed office railing against regulation, which it feared would suffocate innovation and allow China to catch up. “The AI future is not going to be won by hand-wringing about safety,” declared J.D. Vance, the vice-president, at a summit in Paris last year. As recently as June 2nd Donald Trump issued an executive order on AI governance that asked frontier labs for early access to new models and co-operation on cyber risks—both of which were happening anyway. The White House insisted that nothing in the order was to be construed as “a mandatory governmental licensing, preclearance or permitting requirement”.
In practice, that is exactly what now exists. Open AI had privately been sharing details of Sol, which exceeds the capabilities of Mythos on some benchmarks, with the government. But Sam Altman, the lab’s boss, was called by Howard Lutnick, America’s commerce secretary, last week, warning him against releasing the model without prior approval, according to the Information, a tech publication. Open AI is reported to have submitted a list of companies to be granted access; the government excluded some located outside America, according to the Washington Post. Anthropic has been permitted to share Mythos with 100 or so American firms and institutions, as well as their foreign-national employees.
The result of these rules has been a confusing cocktail of permissions. On June 26th Mr Altman suggested that the Sol release would be confined to America. “Working hard for worldwide,” he wrote in response to a question on X, a social-media site. But Britain’s AI Security Institute, a respected assessor of model capabilities and risks, has access to Sol, according to the country’s AI minister. It does not have access to Mythos, which remains formally under export control, though it can use Mythos Preview, an earlier version which is slightly less efficient.
“We don’t believe this kind of government access process should become the long-term default,” complained Open AI. The lab says it is working with the administration to develop a more predictable and “repeatable process” for the future. There has been some progress. Insiders say that the administration is now better placed to adjudicate claims that a model has been “jailbroken”, or tricked into performing a harmful task, the trigger for the Mythos export-control earlier this month. But the larger undertaking is working out which models should be restricted in the first place. Mr Trump’s executive order said that by August his administration would create a “classified benchmarking process” to adjudicate the cyber prowess of new models. That will not necessarily bring clarity, though. Only a handful of people in the frontier labs hold security clearances that would allow them to handle classified material and therefore understand the basis of the red lines set out by the government.
Then there is the question of who will serve as judge. The office of the national cyber director, part of the White House, has spearheaded policymaking, but senior staff have been leaving in recent weeks. Mr Lutnick’s commerce department has been leading the implementation of policy, but has limited in-house technical expertise. The bench of AI talent inside America’s government has never been large, but has shrunk further over the past 18 months.
Anthropic, Google and Open AI, the three main frontier labs, all agree that some sort of federal regulation is necessary, even if they differ on the details. Anthropic and Open AI want a government agency in charge. Google is keener on an industry-funded body like the ones that govern brokerage firms and the power grid in America. Anthropic—ironically, given how it has been treated—is the only one that favours a government veto on releases.
There is little dispute, though, that the current process is dysfunctional. One problem is that it holds back American labs at a time when Chinese competitors are catching up quickly. China’s leading models, most of which are “open weight”—meaning that they can be downloaded and run on one’s own computer—are roughly six to ten months behind their American equivalents, and far cheaper to run. In recent weeks Microsoft, a software giant, was reported to be considering the use of a model from DeepSeek, a Chinese lab, for its Copilot tool.
American labs are also responding to political pressure by tightening the guardrails on their publicly available models, meaning that these are more likely to refuse requests that would once have passed muster. Alex Stamos, who was previously in charge of cyber-security at Facebook, a social-media platform, notes that many companies have already prepared to switch to Chinese open-weight models in the event of further disruption.
Restrictions on the very best American models might slow China down somewhat, because many Chinese AI labs are said to train their models on the output of American ones, a process known as distillation. Yet that would be little consolation if American labs slow down too. Mr Ball, the former AI adviser, warns that companies may be reluctant to invest the vast sums planned for data centres in order to “serve frontier models to whatever 100 companies the US government will allow access”. He adds that labs tend to defray the enormous cost of training frontier models in the first few months of their release, while they have a temporary edge over competing models. “Every week of delay is eating into the narrow window that labs have to make their accounting work.”
Even some of Mr Trump’s closest allies are uncomfortable. “A year ago, President Trump declared that America was in a global AI race and that the way to win it was to be pro-innovation, pro-infrastructure, pro-energy and pro-export,” wrote David Sacks, a government adviser on AI who had initially defended the administration over its treatment of Mythos. “President Trump was exactly right; we deviate from that strategy at our peril.” ■
To track the trends shaping commerce, industry and technology, sign up to “ The Bottom Line ”, our weekly subscriber-only newsletter on global business.