PLEASE stop adding me to your Substacks

RIP Blue Bottle University Ave Gone but not forgotten… gone too young…. You will be missed my friend… Never forget…

Even if Trump offered more powerful AI chips to China, I don’t think Beijing would take them. Beijing wants to solve the ”有无” problem once and for all: don’t be dependent on foreign sources that can suddenly be turned on or off.

Daron has useful thoughts on cooperation, but like many overdoes the "different race" framing. Increasingly, US and China are converging in their approaches! As long as the "bitter lesson" holds, in which concentrated data/compute/talent in general purpose LLMs outdoes specialized models across the board, there will be one main arena. For example, Chinese companies like SenseTime that dominated early AI focused on computer vision w/surveillance data (different race) have pivoted to LLMs and get most of their revenue from it. Yet to be seen whether it makes sense to fine tune models and repeat again on top of stronger base models when new ones come out, or whether just using the closed frontier is better for firms and individuals economically, but for now for most it is frontier. Chinese firms have also seriously boosted their capex this year (work forthcoming from me on this) after being relatively restrained. Chinese models are also less open than before--Bytedance best video model is closed, and many open weights like Kimi now come with an asterisk (cloud hosting needs to pay them for use). They would probably be playing closer to the US game if they could, as Liang Wenfeng has said, turn money into chips like US firms (but export controls limit this). China may also restrict open weight models in the future that reach certain cyber capability. In diffusion, I can find no economywide evidence of more AI adoption than the US. Similar overall adoption CN 43% end 2025 per CNNIC, US 49% of workers per Pew early 2026. Lots of anecdotes about greater CN use, but @PKU1898 research finds even among digitally active Chinese small firms only ~25% use AI. That's a lot of firms not using it! Chinese plans focusing on diffusion do not necessarily lead to more diffusion than US (or Chinese) market imperatives. And as @kevinsxu has pointed out, China's data environment is much weaker in terms of sharing, which hurts tool use and data availability. Much of China's internet is not open, but inside WeChat's walled garden. Not as many APIs and SaaS and ERP systems for business, and most companies don't want to pay for software even if it is useful. Robotics are interesting and the main point of divergence/CN advantage. But not in humanoids, where China has made enormous investments but only make money today only from centers who use them to generate data. The useful robots are for factory automation, and China is leading in installation but still heavily reliant on foreign supply (getting better though). But it also has the largest mfg sector, and @jjding99 work indicates when you scale the robot adoption to China's economy/sector, the adoption is a lot less impressive. Another difference is level of existing, AI specific regulation. China also has a quite extensive regulatory regime for AI already, with pre-registration of LLMs, labeling AI-generated video and images, bans on companions, and many mechanisms like standards bodies that involve firms and gov working on the balance of "development and security." They are also increasingly worried about employment impact--courts have found you can't just fire someone because AI can do their job, for example. But the US has a de facto regime, called voluntary but not necessarily accurately, to test and approve frontier models before both limited release (regulating trusted access programs for cyber models) and general release. US and Chinese reg systems focus on different things, but both are far from fully unfettered. Like Daron I am critical of the race framing, mainly because I don't expect a finish line. We are likely to see US and Chinese LLMs, open and closed, competing in different sectors, across the world, for the rest of our lives. I expect RSI/AGI/ASI will not be one player, but competitive like today's market for AI.

NEWS: Commerce Dpt isn’t represented in the new SI Force. Neither LUTNICK nor anyone else in the agency is in it Commerce had been a central player in AI policy. For one, CAISSI tests and evaluates models Here’s the task force members beyond CLAYTON, FERGUSON, KUPOR and MICHAEL

Reportedly these are President Xi's favorite English language movies.

I asked people which tier they'd put each lab in: Frontier: - Anthropic: consensus - OpenAI: consensus n-1 (1 generation behind): - Google DeepMind: consensus - xAI: consensus - Meta (Muse Spark): over half say n-1 - Moonshot (Kimi): over half say n-1, some n-2 - Zhipu (GLM): over half say n-1, some n-2 - DeepSeek: over half say n-1, some n-2 - Thinking Machines (Inkling): leans n-1, some put n-2 n-2 (2 generations behind): - Alibaba (Qwen): leans n-2, some put n-1 - ByteDance (Seed): over half say n-2 - Xiaomi (MiMo): leans n-2, some put n-1 - MiniMax: over half say n-2, some n-1 - Tencent (Hy): over half say n-2 - NVIDIA (Nemotron): leans n-2, some put n-3+ n-3+ (3+ generations behind): - Microsoft (MAI): over half say n-3+, some n-2 - Cohere: over half say n-3+, some n-2 - Mistral: over half say n-3+ - Amazon (Nova): consensus - Apple: consensus Consensus = 75%+, over half = 51%-74%, leans = the most common pick but under half, some = 20%+. I asked 70 people. Engineers/researchers/founders at AI data cos, people at labs (I didn't count votes for their own lab), and founders/engineers at other startups. I didn't vote.

How do Chinese AI firms make money? In a new report, @cherylwoooo and @ansonwhho identify five revenue sources and analyze how much each contributes. 🧵

Mistral has released Mistral Large 4, scoring 38 on the Artificial Analysis Intelligence Index; France is back to having the most intelligent model from outside the US and China @MistralAI has released Mistral Large 4 in Research Public Preview, with plans to release the weights of the 1T parameter (49B active) model at the end of October. It achieves 38 on the Artificial Analysis Intelligence Index, comparable to GPT-6 Luna (max, 38) and DeepSeek V4.1 Flash (max, 39). It also achieves 50% on the Artificial Analysis Cyber Index, level with GLM-5.3-Flash and ahead of models such as Kimi K3 and DeepSeek V4.1 Flash (max). Key benchmarking results for Mistral Large 4 Preview: ➤ Most intelligent model from outside the US and China: Mistral Large 4 Preview scores 38 on the Intelligence Index, comparable to DeepSeek V4.1 Flash (max, 39) and GPT-6 Luna (max, 38). This makes it the most intelligent model from outside the US and China, ahead of countries such as South Korea and the United Arab Emirates ➤ Level with GLM-5.3-Flash on cyber defense capability: Mistral Large 4 Preview scores 50 on the Artificial Analysis Cyber Index, level with GLM-5.3-Flash (50) and behind MiMo-V2.6-Pro (56). Once its weights are released, it will rank among the top three open weights models on the Cyber Index. Its strongest result is on CyberGym-E2E-AA, where it scores 82%, ahead of MiMo-V2.6-Pro (79%) and GPT-6 Luna (max, 78%) ➤ Over 4x the Cost per Task of similar-intelligence open weights models: Mistral Large 4 Preview costs $1.13 per Intelligence Index task with standard pricing of $1.36/$4.18 per 1M input/output tokens, with $0.14 per 1M cached input tokens. For the first two weeks, Mistral Large 4 Preview will be served at a 50% launch discount, bringing its Cost per Task down to $0.57. This is still more costly than GLM-5.3-Flash ($0.25) and DeepSeek V4.1 Flash (max, $0.27) ➤ Strong document and image reasoning: Mistral Large 4 Preview scores 19% on GDP.pdf, on par with MiMo-V2.6-Pro (19%) and behind Kimi K3 (22%). This is an 18-point improvement from Mistral Large 3, partly driven by improvements in their API, which now accepts 100 images per request, up from 8 for previous Mistral models Key model details: ➤ Context Window: 512k tokens ➤ Multimodality: Text and image input, with text output ➤ Pricing: $1.36/$4.18 per 1M input/output tokens ($0.14 per 1M cached input tokens), with 50% off for the first two weeks ($0.68/$2.09) ➤ Availability: Research Public Preview on Mistral's API, with open weights planned for the end of October

Meet Mistral Large 4, aka Le Chonk. • 1T parameters, natively multimodal. 49B active. It is the best open weights model from US or Europe on aggregated benchmarks. • State-of-the-art on critical workloads, including cyber defense, manufacturing and finance and it surpasses closed frontier models on visual grounding. • Forged in Europe end-to-end and is deployable from Europe via our own Mistral Cloud infrastructure. • Available to all via API today. Working with cybersecurity partners privately. Open weights release end of October.

if you genuinely believe anthropic could build the machine god this valuation is totally reasonable. if not undervalued. investors are thinking about the probability distribution of future outcomes. there is absolutely 0 chance that Samsung builds the machine god. There is like a non-zero chance that Anthropic or OAI does, even if its infinitesimally small

The Blue Angels are coming to San Francisco! Today we officially kicked off Fleet Week, and we are ready to welcome US service members and over 1 million visitors to our city.  This is going to be a weekend like no other in San Francisco. We have air shows, the Parade of Ships, free concerts, K9 Heroes at Duboce Park, and many more activities. You can see the schedule at sf.gov/fleet-week   This is the perfect time to explore our waterfront. Spend an afternoon at Presidio Tunnel Tops, explore Pier 39, check out the new Alioto’s Plaza at Fisherman’s Wharf, stop for a meal in North Beach, or visit a neighborhood small business. Let’s go, San Francisco.

Distilling American AI models could hurt Chinese labs in the long run. The Information’s Asia Bureau Chief @jingyanghk shares an investor’s warning: “AI labs distilling to get ahead is like athletes doping." "You can have a short-term performance improvement, but it'll bring you long-term harm because it will just make you less incentivized to do groundbreaking, original research.”