News of the day
1. China now leads open AI model downloads, surpassing US providers in 2025. This shift has economic and strategic implications beyond initial costs. → Read more
2. OpenAI has reportedly achieved significant improvements in its compute profit margins, enhancing the efficiency and sustainability of AI service delivery. → Read more
3. Anthropic's Claude Opus 4.5 achieves nearly 5 hours of sustained performance, demonstrating significant advances in AI capabilities for long-horizon tasks. → Read more
4. Kilo Code launches App Builder, an open-source tool that challenges Lovable. It generates production-ready code, covering the entire development process from prototype to deployment. → Read more
Our take
Hi Dotikers!
China just won the open source AI race. And nobody really saw it coming.
According to the "Economies of Open Intelligence" report published by MIT and Hugging Face, Chinese developers now capture 44% of popular model downloads in 2025, compared to 28% for Americans. A spectacular reversal when you consider that in 2022, Google, Meta and OpenAI together controlled 60% of the market. Alibaba with Qwen (750 million downloads this year) and DeepSeek have absolutely crushed the competition, relegating Meta's Llama to second place with its "modest" 500 million.
What makes this victory particularly bitter is its hidden cost. Because when we talk about Chinese open models, we're not exactly talking about open source. Transparency on training data has dropped from 80% in 2022 to just 39% today. You download the weights, you tinker with the code, but you have no idea what fed the beast. NewsGuard's audit drives the point home: leading Chinese models propagate or fail to correct pro-Beijing claims 60% of the time. Taiwan, Tiananmen, South China Sea: sensitive topics are systematically handled according to Party lines.
The most worrying part? Andreessen Horowitz estimates that 80% of AI startups pitching them are now building on Chinese foundations. The Trojan horse doesn't even need to hide anymore, we're holding the gate wide open while admiring its price-to-performance ratio.
This dominance raises a fundamental question: can you truly separate a language model from the values and narratives injected into its guts? Every prompt potentially becomes a vector for soft power. Americans wanted to win the war of closed, sophisticated models. The Chinese won the one that actually matters: mass adoption.
G.
Meme of the day
