Mu Li, one of the most recognizable figures in applied machine learning engineering, has taken on a second founding role: co-founder and chief technology officer of Aimotion, a startup building AI sales and marketing systems for the automotive industry. He will keep his position as co-founder and CTO of Boson AI, the model company he started in 2023 with his doctoral adviser Alex Smola. The two companies plan to jointly develop what they call industrial-grade world models for commercial AI agents, according to reports from Jiqizhixin (Machine Heart), PingWest's Silicon Star and Sina Tech.
Li's resume explains why the move drew attention across the Chinese-speaking AI community. At Carnegie Mellon he studied under Alex Smola and built his early reputation in distributed machine learning systems. He co-created Apache MXNet, which Amazon chose as an official deep learning engine on AWS, and rose to Senior Principal Scientist at Amazon leading large-scale ML infrastructure and AutoML work. His co-authored textbook "Dive into Deep Learning" and his paper-reading video series made him a household name among Chinese developers, where he is widely known as "Mu Shen."
Aimotion itself sits in a niche that general-purpose AI vendors have struggled to crack: long, high-value purchase cycles. The company builds AI sales and marketing solutions for car brands and dealerships, operating across the Americas, Europe and Asia-Pacific. A car purchase stretches over weeks or months and involves configuration comparisons, financing math, test-drive follow-ups and trade-in negotiation — a chain where generic chatbots tend to collapse into polite answer machines that never move a deal forward.
The technical pitch is a deliberate reframing of the world model concept. Where physical world models simulate objects, motion and space, Aimotion's version models commercial rules, customer decision logic and business causality: why a prospect went silent after a test drive, how a competitor's price change shifts a negotiation, what follow-up cadence actually converts. The claim is that agents built on such a model can move from answering questions to autonomously executing sales workflows, turning AI from a marketing assistant into a system that carries revenue responsibility. The company says technical and product details will be published later, so none of the capability claims can be independently verified yet.
The financing also carries reporting caveats. Multiple Chinese outlets describe a Series A in the tens of millions of US dollars led by BAI Capital, but the exact amount has not been officially disclosed, and at least one Chinese report described the round differently as a roughly $9 million seed with a Singapore headquarters — a discrepancy between HQ location and round size that suggests the company's corporate structure is still settling. What is consistent across sources is BAI Capital's lead and the two-company partnership.
The pattern Li represents may matter more than any single company. As pretraining returns flatten, senior systems people who once competed on benchmarks are moving toward verticals where proprietary transaction data is the actual moat. Cross-appointments — one scientist, two tables, model lab and application company — are becoming the mechanism for connecting the two without an acquisition.
For the auto industry, the test will be whether an agent trained on real dealership funnel data can beat the CRM-plus-LLM stacks incumbents already run. For the AI industry, the interesting number is whether "commercial world models" can show measurable conversion lift — the kind of evidence that would make verticalized agent companies fundable at growth-stage multiples rather than as demos.
Comments (0)
Log in to join the discussion
Log InNo comments yet