BYD's new cockpit agent runs on Alibaba's Qwen
China's model race keeps moving from leaderboards into hardware people actually buy: a cockpit agent and a same-week model update — while Scott Aaronson reports the labs may now be holding back what their models solve.
BYD's next-generation smart cockpit runs on Alibaba's Qwen, the first time China's most-deployed open model family moves into the car as a commercial product rather than a hobbyist port. At a September 14 launch event where BYD introduced its "Didixia" super agent and the Denza N8L electric model, the two companies confirmed Qwen is wired into the cockpit stack for multi-turn dialogue, complex intent understanding, task orchestration and long-term memory. The architecture is a dispatcher: a voice command is transcribed, Qwen decides the intent and routes the job, then separate agents for vehicle control, music, video, search and ecosystem services carry it out. BYD says the cockpit already uses Qwen's open-source models plus Alibaba Cloud's Bailian platform, with live hooks into Fliggy and Taobao's instant-shopping service. Senior vice-president Yang Dongsheng pitched an open agent platform for third-party developers — the cockpit as an app store again, but with agents instead of icons. We've tracked Qwen's quiet ubiquity all year; this is the consumer version of it, in the country with the largest connected-car fleet.
ByteDance shipped a September 15 update to Doubao 2.1 Pro, and the changelog reads like an answer to the agent-reliability complaints from enterprise users. The company says it strengthened evidence tracing, authoritative-source retrieval and data verification for long agent runs — its framing is that financial-research-style reports now produce conclusions that can be checked against sources, with fewer hallucinations. The multimodal coding push is the bigger claim: ByteDance says the model now handles cross-file, long-horizon repo work — reading a design mock-up, blueprint or screen recording and turning it into front-end code or game logic. Image and video reasoning token consumption is down more than 30% from the previous generation, per the company. The update rolled out simultaneously to the Doubao app, ByteDance's TRAE coding tool and the Doubao-Seed-Evolving enterprise line, with no API endpoint change required — that no-migration detail is aimed squarely at production customers.
Theoretical computer scientist Scott Aaronson says he has heard rumors that AI companies are sitting on solutions to "very major" open problems in his field — and hiding them because of the backlash over the Navier-Stokes announcement. In a blog post published this week, Aaronson wrote that he's told the labs "have been burned by the hostile response to the Navier-Stokes proof" and are holding solutions "until they figure out a better way to handle things," explicitly ruling out P vs NP but not other headline problems. He's careful to label it rumor, and so should everyone repeating it — no lab has confirmed hoarding anything. But his framing of why is the part worth reading: he lists recent AI-solved results in his own subfield and describes a math community whose conversations now circle one question, who gets credit for what. We made that argument in Deep Dive — Math's credit system was built for humans. AI just broke it, and reported the original episode in OpenAI's AI solved Navier-Stokes — and a mathematician cries foul. If the rumor is true, the bottleneck for frontier math stops being capability and becomes publication policy.
What to watch: whether the Didixia cockpit agent stays exclusive to BYD or becomes Alibaba's template deal with the next automaker.
If labs really are holding solved open problems until the credit rules catch up, is that responsible restraint or privatized science? Tell us in the comments.
