Multimodal AI breakthrough could come within two years, SenseTime scientist says

SenseTime’s Lin Dahua argues that intelligence extends beyond coding, positioning multimodal AI as central to real-world applications. This view contrasts with DeepSeek’s Liang Wenfeng, who sees multimodal AI as a component rather than the core path to AGI. SenseTime’s focus on native multimodal AI, particularly through its SenseNova 6.8 and U series, reflects its computer vision roots. These models unify visual understanding, generation, and editing, addressing diverse industry needs.

The challenge for SenseTime and other Chinese AI developers lies in the "alignment tax" of integrating visual data without degrading existing language and reasoning capabilities. Training costs and the need for rapid iteration demand a disciplined approach. SenseTime’s strategy of iterating with lightweight 8-billion-parameter models before scaling up is a pragmatic response. This allows the company to stabilize architectures and data mixes, crucial for sustained investment in a competitive landscape.

The thing to watch is whether SenseTime’s focus on "agentic generation" and continuous editing in products like Seko can deliver tangible commercial value. This moves beyond single-shot generation to a complete workflow. Its success will determine if multimodal AI can achieve the revenue and valuation milestones that coding agents have already demonstrated. SenseTime must prove commercial viability through user growth and paid usage, not just larger models.