Greetings from a world where…
does aging mean you continue to get worse and worse at League of Legends
…As always, the searchable archive of all past issues is here. Please please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay support access for all AND compensation for awesome ChinAI contributors).
Three mistaken assumptions in the debate on DeepSeek and Export Controls
I want to briefly react to this post by Dario Amodei, Anthropic’s CEO, which argues that DeepSeek’s success only reinforces the case for export controls. I’ve stated before that the U.S.’s chip controls represent a fascinating alliance between three camps: 1) China hawks who want to be tough on China for the sake of being tough on China; 2) “small yard, high fence” folks concerned about dual-use risk of AI; 3) people who believe AGI risks supersede any other causes. Dario’s post neatly reveals some of the key mistaken assumptions of this precarious throuple:
A lack of grounded thinking about how general-purpose technologies (GPTs) like AI shape military power. The closest Dario gets to a tangible scenario is that if China develops the smartest AI system, then it can make even smarter AI systems and parlay a temporary lead into a durable advantage. His expected timeline for this scenarios is 2026-2027.
History tells us that this is not how GPTs impact military power. Based on years of empirical research into how electricity — potentially a more transformative technology than AI in relative terms — actually shaped military effectiveness, I found that GPTs enhance military effectiveness through a broad impact pathway and prolonged timeline of diffusion. People thought that electricity would produce electric rays of destruction that would wipe out entire battlefields (think: one day we open up a box and get AGI as a superweapon); in actuality, electrification transformed militaries in a decades-long process through diffuse channels, with main applications in lighting, intelligence, communications, transportation, and logistics.
This is clearly written by someone who has never studied the military procurement and acquisitions process. The U.S.’s most advanced military systems rely on legacy semiconductors, not the most cutting-edge ones.
I want to respond specifically to this point: “Even if the US and China were at parity in AI systems, it seems likely that China could direct more talent, capital, and focus to military applications of the technology.” This line was especially ironic, as Anthropic recently announced a partnership with Palantir to provide U.S. intelligence and defense agencies access to Claude 3. If you just look at military partnerships with civilian cloud computing providers, the U.S. is way ahead of China on this front.
A dogmatic approach to AI timelines.
It’s really important to note Dario’s assumed one-to-two year timeline. I’ve been in governance of AI circles since 2017. In that time, I’ve consistently heard some iteration of: AGI is two years away! It’s the Bruno Caboclo of technologies: always two years away from being two years away.
Side note: in my interactions with the smart people who are modeling the effects of export controls, I’ve found myself blown away by how confident they are in projecting the trajectory of AI in the long-term. If DeepSeek proves anything, it’s that we should all be a little more humble about these predictions.
If one thinks AI’s transformative impacts will take place in a year, then, sure, export controls have a chance of working. However, if your timeline is longer (a much more likely scenario in my opinion), then Nvidia’s monopoly will fade.
A very narrow view of national interests as it relates to US-China AI competition. Dario writes, “To be clear, the goal here is not to deny China or any other authoritarian country the immense benefits in science, medicine, quality of life, etc. that come from very powerful AI systems. Everyone should be able to benefit from AI. The goal is to prevent them from gaining military dominance.”
In debates about export controls, I’m always struck by how the national interest is reduced to an AI arms race for military superiority. This unidimensional view ignores other things the U.S. should care about.
A previous post, ChinAI #280 “The U.S.’s Unstrategic Approach to the ‘Chip War’”, went into detail on this point.
Where is the conflict of interest disclosure? I may be the only person who cares about this, but if I were the head of a a company advocating a policy directly targeted at my competitors, I would at least acknowledge that my financial interests could unduly influence my judgement.
Relatedly, see Anthropic’s lobbying on the California AI safety bill: “Anthropic seems to be acting like any for-profit company would to protect its interests. Anthropic has not only economic incentives — to maximize profit, to offer partners like Amazon a return on investment, and to keep raising billions to build more advanced models…‘Anthropic is trying to gut the proposed state regulator and prevent enforcement until after a catastrophe has occurred — that’s like banning the FDA from requiring clinical trials,’” Max Tegmark, president of the Future of Life Institute, told Vox.
For Dario’s next post or op-ed, I would be really curious about his and Matt Pottinger’s views on the Trump administration rescinding the Biden AI order that guards against safety and misuse risks of AI. Isn’t that one of the key risks that chip controls are supposed to address? Interesting that this got lost in the post.
Feature Translation: 2025 Look-Ahead — What historical node are we at in AI development?
Context: ChinAI readers consistently surprise me. I thought folks would choose the open source developers annual white paper (#1 from the Around the Horn Issue) or the report on how many Chinese companies are actually using 1 billion in tokens — an indicator of large model diffusion and adoption (#5). However, this is not a dictatorship, so we go with the people’s choice: Deli Zhao, head of Alibaba DAMO Academy’s AI team, gives his perspective on AI’s past year and future decades.
Key Takeaways: The Scaling Law has not reached a ceiling BUT its fixed trajectory has been broken.
On the first point, Zhao writes:
In the past year, there have been many discussions about whether the Scaling law has encountered a ceiling, but in fact, there are only a few companies in the world that have enough resources and data to touch the ceiling of Scaling law. Because first of all, it requires sufficiently powerful infrastructure and computing resources, and secondly, it requires sufficient training data. Regarding data, on the one hand, there is existing Internet-related data, and on the other hand, there is synthetic data — synthetic data is very important, but whether the quality of synthetic data can be used for effective training depends on the generation ability of the basic model and the method of synthesizing data. Up until 2024, only a few models such as GPT-4 may reach this level. Therefore, it is not possible to conclude that the Scaling law has reached its ceiling.
However, DeepSeek’s advances have demonstrated that the fixed trajectory of the Scaling Law has been broken. In other words, you can get very very far (even SOTA performance) with more efficient training through algorithm design and engineering optimizations. Interestingly, Zhao notes that this will lead to more R&D on optimization of model architecture at the computing chip later (e.g., neuromorphic chips).
It’s not just DeepSeek; Chinese labs are investing heavily in increasing the “density” of large models, thereby squeezing more performance out of models trained with fewer parameters and less computing power (see ChinAI #296)
Zhao points to other Chinese models that have made similar breakthroughs, including Shanghai AI Lab’s InternLM3-8B-Instruct model, ModelBest’s “light artillery" MiniCPM-o 2.6 end-side models, and Chinese AI startup Minimax’s MiniMax-01 model.
Two further details from the MiniMax-01 paper: 1) authors comment that their efficiency improvements work better on lower-end Nvidia chips: “By the way, these optimizations can bring very noticeable benefits on H20 compared to H800.”; 2) there are substantial sections devoted to AI safety protocols.
A few other bits and pieces:
Zhao sees video generation (OpenAI’s Sora) as one of the four key AI advances in 2024. In China, they key players are: Kuaishou’s KLING AI, Minimax’s Hailuo AI, Alibaba’s Tongyi Wanxiang, Tencent’s Hunyuan, and Bytedance’s Doubao.
Zhao deems AI agents (e.g., Anthropic’s “Computer Use” and Zhipu’s AutoGLM) as “the biggest substantive transformation since the birth of the Windows system for personal computers and smartphones.”
FULL TRANSLATION: 2025 Look-Ahead - What historical node are we at in AI development?
ChinAI Links (Four to Forward)
Should-watch: Jeff on Wolf Blitzer’s CNN show
I had a “wow-did-that-really-happen” opportunity this past week to discuss DeepSeek and China's AI development on The Situation Room with Wolf Blitzer. They were kind enough to give some primetime product placement for my book too! Too bad they cut the part where I showed them that DeepSeek R1 still struggled to create a funny joke that involved a skunk.
Should-read: AI Generated Fake News is Taking Over my Family Group Chat
Big shoutout to ChinAI reader Lily Li for translating #2 from last week’s Around the Horn issue. It goes through some fasincating cases of AI-generated fake news in China, including a Zhejiang Group that used ChatGPT to generate fake news as well as an “Emerald Ink” software that generated over 190,000 articles each day. From her translation:
Thus, sensational fake news will always have an audience. Those interested in international politics will be riled up by the fabricated remarks of AI Trump; those who are health-conscious will enthusiastically order wafer protein bars promoted by AI Zhang Wenhong[2]; even those without hobbies will tag everyone when reposting disaster reports and human interest stories of dubious veracity on the family group chat.
Should-read: Helen Toner’s thread of good DeepSeek takes
Helen, director of strategy at Georgetown’s CSET, compiled a helpful list that I learned a lot from.
Should-read: Creation Lake by Rachel Kushner
I enjoyed this meditation on civilizational progress disguised as a spy novel.
Thank you for reading and engaging.
These are Jeff Ding's (sometimes) weekly translations of Chinese-language musings on AI and related topics. Jeff is an Assistant Professor of Political Science at George Washington University.
Check out the archive of all past issues here & please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay for a subscription will support access for all).
Also! Listen to narrations of the ChinAI Newsletter in podcast format here.
Might be of interest. Education background of the DeepSeek team:
https://collegetowns.substack.com/p/where-did-the-deepseek-team-study
Thank you Jeff!! I love your newsletter, it contains info we don't see anywhere else, and it punctures the US big bros' hype.
You may enjoy my papers: https://curriculumredesign.org/papers/#Pubs-Technology
Does Present-Day GenAI Actually Reason? (2025)
This paper considers whether present-day GenAI can reason through 1) determining which types of cognitive processes and procedures (aka “modes of thinking”) are used in human reasoning. Then, 2) determining which forms of human reasoning can be mimicked/reproduced by GenAI/LLMs. It finds that AI can, to a certain extent, achieve some “reasoning” at various levels of capability and with its now-familiar “jagged boundary,” (which is quite an achievement), but not at the all-encompassing and overhyped level touted by the industry. And this is certainly NOT AGI either.
AI in 2025: A Combinatorial Explosion of Possibilities, but NOT AGI (2025)
This paper discusses the rapid evolution of Generative AI (GenAI) by 2025, highlighting its transition from theoretical to engineering-focused development. It examines various technological advancements, including specialized datasets, improved training methods, multimodal models, and AI agents. It argues that despite challenges, these innovations will lead to unpredictable, transformative capabilities, but not AGI.
The Frustrating Quest to Define AGI (2025)
This paper discusses whether Artificial General Intelligence (AGI) can ever be defined properly by reviewing the various approaches, identifying their validity, and proposing alternatives.