David Sacks，直到最近都是特朗普的加密货币和人工智能负责人，昨天对此发表了如下看法，借当前形势机会来提倡他最喜欢的观点：美国的主导地位和对人工智能的最小化监管：
这是对商务部长霍华德·卢特尼克（Howard Lutnick）推文的回应。
CAISI 报告在这里：https://www.nist.gov/news-events/news/2026/07/uk-aisi-caisi-preliminary-assessment-kimi-k3s-cyber-capabilities。
虽然我并不怀疑 CAISI 报告本身，但我对 Sacks 的观点感到困扰，并在这里写了相关内容：
正如几天前所讨论的，我认为美国至少应该考虑转变方向，尝试与中国合作，实现人工智能的全球利益，而不是制造一场新的冷战，这场冷战很可能以对所有人都不舒服的方式结束，没有明确的赢家。
但这意味着 Scam、Dario 和 Elmo 的钱会少。这些人根本不关心全球利益或治愈癌症等问题。他们唯一关心的是下一笔十亿美元资金从哪里来。如果他们关心全球利益，他们就不会试图在美国每块空地上建数据中心。
写得很好，Gary。我想就像核不扩散条约一样，由于人工智能现在在国防中至关重要（根据乌克兰和无人机战争），关键在于细节，美国如何让另一个世界大国看到合作的优势，而不是直接竞争并试图重新获得一点执行优势。
从经济角度看，你可以说中国已经赢了，因为全国范围内将非常优秀（如果不是最好的仅有一点差距）的人工智能广泛融入他们制造的每件产品，这种规模化正在发生，而美国相对昂贵的模型（IPO 成功所需）则阻碍了这种普遍应用。
那么‘西方’将如何竞争呢。中国已经是世界的制造中心。他们将变得更强大，由于政府在人工智能模型软件中强制要求后门，美国和欧洲的公司必须尝试用成本高十倍的人工智能来匹配中国公司——这是一个将持续加剧的结构性劣势。
当然，中小企业都会忽略数据安全风险，并且在成本仅为 Claude 10% 的情况下无法抵抗 Kimi 3 的推理能力，但上市的美国公司就做不到这一点。
在开放权重模型（以及相关的工具、编排、内存）追上之前，这都是一个相当大的挑战。或者公司重新拥有自己的“硬件”，训练自己的专有模型，我想这甚至会进一步推高对 Nvidia 芯片的需求！
除非封闭模型的开销和不可证明的使用寿命（ROU）导致企业使用量下降，到那时我们才能以更有吸引力的价格从 Google / Microsoft 等租回所有数据中心的容量。
David Sacks, until recently Trump’s Crypto and AI Czar, had this to say yesterday, using the current situation as an occasion to champion his favorite points on US primacy and minimally regulating AI:
This posted in response to a tweet by Commerce Secretary Howard Lutnick
The CAISI report is here：https://www.nist.gov/news-events/news/2026/07/uk-aisi-caisi-preliminary-assessment-kimi-k3s-cyber-capabilities .
While I don’t doubt the CAISI report per se, I am troubled by Sacks’ take, and wrote about it here:
As argued here a couple days ago, I think the US should at least consider pivoting and trying to work with China towards AI for global good, rather than creating a new cold war that is likely to an end in uncomfortable way for all, with no clear winners.
But that means less $ for Scam,Dario and Elmo. These guys dont give a shit about the global good or curing cancer or any of that. All they care about is where their next billion dollars is coming from. If they cared about the global good, they wouldnt be trying to put data centers on every piece of vacant land in America
Well written Gary. I guess as with Nuclear anti proliferation treaties, because AI is so central now to defense (per Ukraine & drone warfare) the devil will be in the details on how the US could get the other big world super power to see an advantage in cooperation vs straight out competition and trying to regain a small execution advantage.
Economically you could argue China has already won, because country wide industrial impregnation of very good (if not the best by a small margin) AI into everything they manufacture, is going to (or is) happening on a pervasive scale that the relatively expensive models in the US (needed for IPO success) prevent.
Then how will the ‘West’ compete. China already is the world’s manufacturer. They’re about to get stronger and because of the govt mandated backdoors in the AI model software, corporations in the US and Europe have to try and match Chinese corp’s with AI that is 10x as expensive - a structural disadvantage that is going to compound.
Sure, the SME’s will all brush off the data security risk and won’t be able to resists Kimi 3 inference at 10% of the cost of Claude, but listed US Corp’s are not going to be able to do that.
It’s a pretty significant challenge until open weight models (and harnesses, orchestration, tooling, memory) catch up. Or corporations go back to owning their own ‘metal’, training their own propriety models, which I guess drives demand for Nvidia chips even higher!
Unless the closed model expenses and non demonstrable ROU cause a drop in businesses use, and then we can lease back all those data centre capacity from Google / Microsoft etc at more interesting rates.
扫码直达文章详情