1/
Launching a new Substack: "Strait Up," focused on AI policy, semiconductors, and geopolitics, with one eye always on the Taiwan Strait.
First post is on Anthropic's vision for China compute controls, and the piece of the strategy almost everyone is leaving out.
Latest : Anthropic says Alibaba conducted the largest ever campaign to glean its IP via 28.8 million exchanges with Claude.
Senators Hagerty and Kim are introducing an NDAA amendment directing Commerce to blacklist Chinese companies found to be doing this.
@kimmonismus Sad that some people are seemingly happy that Chinese companies are engaged in massive theft. "At least I'll get part of the spoils."
I wonder if attitudes will change once people are actually holding Anthropic stock.
NEW: On June 17, @BISgov within @CommerceGov issued new guidance regarding export controls on AI chips, which seemingly acknowledges—but fails to close—one of the significant export control loopholes that BIS recently created, which makes it possible for Chinese front companies in third countries to make export-controlled AI chips at TSMC/Samsung.
For background, the June 17 guidance builds on BIS's May 31 guidance, which was issued in response to the government's realization that BIS’s non-enforcement of certain export controls on AI chips had created two significant and unintended loopholes: (1) firms no longer required a license to export AI chips to foreign subsidiaries of China-headquartered firms, which had been prohibited since 2023; and (2) TSMC/Samsung no longer required a license to fabricate and ship potentially export-controlled AI chips for untrusted entities in third countries, which had been prohibited by the Foundry Due Diligence (FDD) rule issued in response to Huawei's use of front companies to make millions of AI chips at TSMC. BIS's May 31 guidance resolved the first loophole, but not the second.
The interesting thing about the June 17 guidance is not the content, but rather the file name in the hyperlink: "fdd-faqs-06-17-26.pdf". The document itself issues a straightforward clarification related to the first loophole: it states that the May 31 guidance applies to all variants of controlled AI chips, not just some. But it does nothing to fix the second loophole, which exists because BIS is not enforcing the license requirements that the FDD rule relies on to have its intended effect. It does not even contain any content related to the FDD rule.
HOWEVER, the file name starts with the word "fdd"! This seemingly indicates that the June 17 guidance may originally have been intended to address and close the FDD loophole, but for some reason was watered so that it doesn't even address it, so the loophole now remains.
What is going on here? Why is BIS not fixing this loophole? BIS is obviously aware of the problem, that much is now clear as day. I understand a lot is happening in the export-control world right now, but BIS simply has to be able to walk and chew gum at the same time. It created this loophole through its own actions and non-enforcement of existing controls, but is dragging its feet on closing it. If BIS doesn't enforce the rules that are in place to counter Chinese circumvention of U.S. export controls, there isn't any point in having export controls at all.
All the more reason to tighten export controls further on SME. But it will take stronger international Cooperation on enforcement with countries like Netherlands and Japan. Excellent analysis.
Excellent work by @Hamish_Low5 projecting that China won't indigenize 7nm-capable DUV immersion tools until the mid-2030s.
When I made forecasts ~6 years ago, my estimates were that China's *EUV* indigenization -- an even more complex technology --- would take until the mid-2030s, which at the time felt quite bearish on China.
It turns out that China is even further behind on lithography than we thought, showing just how hard the technology is to master.
7/
Integrating Taiwan's membership may be the first real test of whether TSG can work.
Read the full article and subscribe to Strait Up here: https://t.co/tZ9JCTsIAR
1/
Launching a new Substack: "Strait Up," focused on AI policy, semiconductors, and geopolitics, with one eye always on the Taiwan Strait.
First post is on Anthropic's vision for China compute controls, and the piece of the strategy almost everyone is leaving out.
6/
Taiwan is essential to TSG: both as chipmaker and known smuggling route. But it's also already the #1 target of CCP cyberattacks and espionage. So while all TSG member should get priority access to frontier cyber models, Taiwan should likely be first.
China is Mogging Western Auto, and that’s Bad for Semis, National Security & War
If you live anywhere outside the US, you've noticed it: the streets are filling up with cars you've never seen before. Chery? Jaecoo? Zeekr? Leapmotor? BYD? No, you didn't miss a decade of car launches. They're Chinese. And they're everywhere. (1/10)🧵
And Dean is completely correct. It’s obviously insane to sell AI chips to China, but then export control the models that those same chips produce and run.
@deanwball@nickm4rro Yes, the housing situation in Taipei is a nightmare. Average salary/housing price ratio is worse than many global cities like London, NY, etc.
But should be noted this is almost exclusively a Taipei problem. In most of the rest of Taiwan, new construction is at a historic high.
If these chips were in fact smuggled through Japan, it shows smuggling is not just a Southeast Asia problem. For export controls to be a serious US policy, a permanent consortium of allied supply chain partner, like Taiwan, Japan, South Korea, etc. will be needed for enforcement.
EXCLUSIVE: Taiwan suspects Nvidia's AI chips were smuggled to China through Japan in the island's first public crackdown on the advanced tech https://t.co/1Srfhhr5gv
@MatthewBerman Dismissing CBRN threats because "it's already online" is like dismissing Mythos' cyber capabilities by saying, "the vulnerabilities are already in the code." They were there, but the model made it easy.