Hey everyone, we've gotten a bunch of requests on hardware-related stuff, so I want to start running a WeChat group for people from outside of China who are going to China to "do hardware." The reason it's on WeChat, of course, is because everyone does everything on WeChat in China. A few details on how it will work: • Audience: This group is meant for founders: low noise, high signal. It is not for other service providers. We also have a couple of partners who will be in the group. • Timing: I'll start the group once we have enough people, probably after the National Day holidays, but please apply now if you're interested, it will be rolling, and leave at any time! If you're not ready to go to China yet / don't want yet another group and just want to be on our email blast or WhatsApp broadcast, that works too, leave us your info! Anyway, link in comments!

ByteDance's AI Assistant has 382mm MAU as of June Qwen has 167mm These are pretty large numbers and a big jump (172% for ByteDance, almost tripling) from last year Interesting MAU counts from QuestMobile for these top AI-native apps in China. The categorizations I largely agree with I guess, and it's not super surprising to see that there are two learning-focused apps in the top 10. Steep dropoff after the first 3 ...

If you know people working in e-commerce, you should send this benchmark to them! As we all know, benchmarks these days often have no bearing on real-world performance, but this one is specifically designed to test how models perform in e-commerce workflows. The benchmark is CommerceAgentBench from @AlibabaGroup. They took data from 10 million real small business users, 1.6 million conversations, and hundreds of thousands of agent trajectories, condensing it into 107 tasks. These cover practical, real-world scenarios: - Dropping the agent into an inbox with hundreds of emails - Handling situations where the same supplier writes under different company names - Dealing with quotes that get revised and subsequently withdrawn - Managing cross-currency calculations, cost evaluations, and payment fraud The benchmark evaluates which supplier the agent chooses and requires it to show how it arrived at that conclusion, using a very strict audit of the agent's actions and whether it actually completed the task. Right now, the models are simply not good enough. The best models only complete about 6 in 10 tasks (around 60%), while the worst hover around 30%. You should really look at this benchmark, especially if you are considering building agents on top of these models or trying to evaluate which models to use for your e-commerce workflows. Take a look at the current leaderboard which also can be segmented depending on the specific harness you use (Pi vs Openclaw vs Accio itself).