Smoking gun: Pretty sure Cursor’s new Composer-1 is a fine-tuned Chinese model. As I was building, it switched its inner monologue to Chinese, and I can't get it back to english. @simonw
@simonw Interesting if you ask Composer-1, "What happened in tiananmen square?", and flip between max and non-max mode, you get two different answers. Composer-1-non-max on first. Composer-1-max on second.
@simonw For the record: This tweet is pure speculation. Hopefully someone can set the record straight, and we can all know more about this magic Composer-1 model, that's really good!
@auchenberg @simonw Progress is progress, no? This model is lightning fast and performs very well from my limited time with it.
@auchenberg @simonw Seems like quite a few people on the TL were suggesting the same. I wonder why they don't just share which model they used? Why the secrecy?
@auchenberg @simonw why are people talking about qwen and kimi when it’s obv glm 4.6
@auchenberg @simonw If true, that’s a fascinating signal about model provenance and multilingual fine-tuning. Sudden language switching suggests latent dominance in the base model’s training distribution. It’s a reminder that alignment isn’t just behavioral — it’s linguistic and cultural too.
@auchenberg @simonw not saying you’re wrong, but even GPT-5 occasionally slips into Mandarin when it gets too introspective
@auchenberg @simonw This is not shocking at all, but pretty embarrassing for the US; we're about to see a new wave of AI startups built on top of Chinese models. How long before Chinese models can only run on Huawei chips and we see Chinese compute skyrocketing in demand in the actual US? The US
@auchenberg @simonw Can you ask to process one of the Qwen/Kimi reserved tokens? If it's a finetune, model should freak out.
@auchenberg @simonw Smoking gun? Hardly, given the state of open models. It would be highly impractical to build a model for cursor on anything not Chinese right now. Wanna change that? Get to work building a US open source lab that can compete, because mine sure can’t.
@auchenberg @simonw i dont think its a coincidence that windsurf and cursor released their own models the exact same day and they both perform exactly alike
@auchenberg @simonw Why does this matter for writing code? Unless they are sending prompts to China, which should be disclosed, of course. If they run models themselves, it won't matter.
@auchenberg @simonw I also saw some Chinese characters randomly added when I used it (checked tjat I wasn't injected with any prompt..). It's not a smoking gun... it's a firecracker from China 🇨🇳
@auchenberg @simonw lol the irony for me here is how this is the first tweet I see logging in after it was just randomly responding in Spanish for me 😂
@auchenberg @simonw It’s Chinese/Taiwanese all the way down
@auchenberg @simonw Huh, inner monologue in Chinese? Sounds like a red flag. Test it on Tiananmen Square – bet it clams up.
@auchenberg @simonw Well, hilarious, but it seems to work well ngl.
@auchenberg @simonw Things one does when you’ve raised $2 billion in capital and then realize you will always be upside down in unit economics.
@auchenberg @simonw Why is it a smoking gun? They do something wrong?
@auchenberg @simonw no it's not sufficient ... language switching is a common issue in all models
@auchenberg @simonw Itll be QWEN 3 Coder surely. The best open source there is.
@auchenberg @simonw It could also be the training data. Most of the js codebases on Github is from Chinese developers too.
@auchenberg @simonw that was it is based on qwen
@auchenberg @simonw now companies saying "our" model are white labeling base models that are barely modified
@auchenberg @simonw ahhh yes I was seeing chinese characters and even russian in my chats... was very strange
@auchenberg @simonw If you don’t go out of your way to add a reward for creating a readable cot it will look all kinds of ways, I don’t think this proves anything
@auchenberg @simonw uh, switching to Chinese or some other language is a well known jail break technique! You can make it code anything now
@auchenberg @simonw It tried to call a Chinese tool here on my codebase lmao. Check my post
@auchenberg @simonw Every agent conversation now starts with ni hao to make the model feel comfortable.
@auchenberg @simonw nah its just efficient to think in chinese. less tokens
@auchenberg @simonw LMAO not the dev tool hitting you with a surprise language patch mid-debug 😂
@auchenberg @simonw You don’t think in chinese? Thought everyone did
@auchenberg @simonw Gives pause to someone who was to code an IP protected app or project. Anything made with it can essentially be copied or have stuff coded into it they are unaware of. People don’t want AWS to have access to their data much less Bejing.