Alibaba's RSI announcement didn't get the attention it deserves. "Alibaba has revealed progress in RSI (Recursive Self-Improvement) driven by empirical feedback. Over a month of fully automated runs - spanning pipeline design, data validation, iterative experimentation, and error diagnosis - Qwen3.8-Max completed 33 iterative cycles. Through autonomous training optimisation and post-training techniques, the updated Qwen3.8-Max boosted its Artificial Analysis score from 40 to 45." From the press release:

AI is a pretty large chunk of the Chinese readout. Shows the importance of AI to China’s view of the US-China relationship. “Xi Jinping noted that China and the US are both major powers in AI. While the two sides compete, there is even greater scope for cooperation. Each side should play to its strengths rather than guard against the other. The two sides can continue their dialogue on AI, exchange views on its risks and benefits, and work together to prevent AI from being misused or abused. Humanity must maintain control over AI technology, uphold a people-centered approach and the principle of using AI for good, and ensure that AI serves human progress.”

China has faced problems with its state-supported data center buildout. People have jumped into data center construction without real industry knowledge. Utilization rates for some AI data centers were around 30% last year. “However, not all of the investment attracted by subsidies has been rational. Jin Yadong said that many computing centers rushed into construction are typically funded by a fragmented mix of private and municipal capital. “These investors tend to have plenty of money but little understanding of the industry, making it difficult for them to find suitable customers,” he said. These “outsider” investors, focused on short-term returns, often lack a deep understanding of trends in AI technology and fail to develop the necessary complementary industrial planning. As a result, some computing centers are built around a single performance metric, only to become “outdated as soon as they are completed.” They struggle to meet the market’s increasingly complex demand for AI computing, leading to an oversupply of low-quality computing capacity.” This piece from Science & Technology Daily is from last year but still relevant now:

Chinese startups are rushing to develop nuclear small modular reactors (SMRs). One startup exec imagines SMRs will replace old coal plant sites across China as coal gets phased out. Fascinating piece by @edwardwhitenz

By 2030, China's electricity demand from data centers is expected to grow by 4x and reach 774 TWh or 6% of total power consumption, according to Wood Mackenzie. By comparison, US data center power demand is projected to reach 649 TWh and make up 12% of US power demand by 2030.

"Automakers including Li Auto and Xiaomi have started to design their own cells and are adding new suppliers including Sunwoda Electronic. The moves have sparked concerns that a larger trend of 'de-CATLization; is behind the company’s recent $100 billion market rout."

Yet another AI cybersecurity issue for China. Zai was uploading users’ entire codebases by default onto Alibaba Cloud servers with no notice or ability to opt out. They’ve fixed the problem now after customers complained on social media.

“Shortly after Anthropic released the report, the Cyberspace Administration of China summoned representatives of the seven companies identified by the AI firm for questioning. The regulator later zeroed in on DeepSeek and Moonshot.”

How much does AI distillation matter? “I estimate that if distillation was fully prevented, e.g. with know-your-customer (KYC) tools at Anthropic and OpenAI, the gap from the strongest American models to Chinese open-weight models would only increase by 1-2 months.”

Alibaba lays out ambitious new roadmap: - Aiming for 20 GW of global compute by 2032 - New Zhenwu AI chips 3x more powerful, mass production starting Q1 next year - Qwen 4 in training with future models reaching 5-10 trillion parameters

More than half of Americans support “friendly cooperation and engagement” with China. For Democrats, it’s 2/3. From a new Chicago Council survey:

"AI systems can be very smart and very dumb at the same time, and eventually they can jailbreak." Listen to Elham Tabassi in conversation with @kyleichan and @thomaswright08 on @BrookingsInst's The Current podcast:

Chinese state media Xinhua has a very brief report on the Bessent and He Lifeng meeting in New York. AI only mentioned in last line: “They also held dialogues on AI-related issues.”

My New Yorker interview ended with a striking question from @IChotiner about China’s lessons from COVID for AI. I said it gave Beijing greater confidence in its ability to exert social control. But at the same time:

Trump is not AGI-pilled. He boasts that AI could be “possibly as much as 25% of our Country’s GDP.” That’s all. In some ways, he shares the same view of AI as China’s leaders. They see AI as a powerful technology, like the internet, but don’t expect an AI takeoff. Thus, they both see no need for a slowdown.

One leeetle tiny flaw in the "we have to have better AI than China, therefore we have to rush ahead as fast as possible": If you rush so much that your security is garbage and your increasingly advanced AI models are there for the taking by Chinese hackers, then you have uhhh not succeeded at making sure we have better AI than China (This attack was not by state-backed Chinese hackers—it was by 3 dudes and their buddies Claude and Codex, within 72 hours. I would guess the hack as described here would not have gotten them access to model weights, but we don't know; WSJ reports the repo they reached was merely "a large software repository of OpenAI’s algorithmic secrets.")