Less hype and more hardware: SenseTime banks on multimodal AI to regain its edge

Chinese artificial intelligence pioneer SenseTime is betting that its roots in computer vision will help it lead the next phase of AI, as the industry shifts towards multimodal systems and embodied intelligence in the physical world, according to co-founder and chief scientist Lin Dahua.\nIn an interview with the Post on Wednesday, Lin said the company’s long-standing expertise in vision-based AI put it in a strong position to become a leader in embodied intelligence, robotics and AI agents operating in real-world environments, at a time when there is growing debate about the limits of large language models (LLMs).\n“Our strategic approach is somewhat similar to Google’s in the United States, which primarily focuses on multimodal AI including the latest Nano Banana Pro. They also start with vision capabilities as the core, then add language abilities to create real multimodal systems,” said Lin, who is also an associate professor of information engineering at the Chinese University of Hong Kong.\nThe Hong Kong-listed company, long regarded as one of the world’s leading facial recognition providers, is trying to carve out a new role in the generative AI era that followed the launch of ChatGPT three years ago.\nExtending his comparison with Google – which has deep capabilities across the AI stack, including its own TPU chips for training models – Lin said SenseTime’s decision as early as 2018 to build out large-scale data centres had laid a solid foundation for its ambitions.\nAs of August, the company’s total computing power stood at about 25,000 petaflops, up 8.7 per cent since the start of the year, after surging 92 per cent over the whole of 2024.\nA petaflop is a standard measure of computing speed equal to one quadrillion calculations per second – a critical benchmark for companies building large-scale AI infrastructure.\nThat capacity was set to continue expanding at a “high double-digit to triple-digit” pace over the next two years, SenseTime co-founder and SenseCore president Yang Fan told the Post in April. SenseCore is the company’s AI infrastructure unit.\n\nSenseTime, which is on a US blacklist, is also among the earliest major Chinese tech firms to start shifting to domestically produced chips.\nThe share of its total computing power now provided by home-grown chips had reached “double digits”, Lin said, without disclosing a precise figure.\nFounded in 2014 by professors at the Chinese University of Hong Kong, SenseTime was part of China’s first wave of so-called “AI dragons”. Its profile has dimmed in recent years as a younger generation of “AI tigers” – including Moonshot AI, MiniMax and Zhipu AI, which focus mainly on LLMs – grabbed the spotlight.\nLin argued that the relative lack of attention around SenseTime did not mean it had fallen behind in the AI race.\n“Some companies release a model every few months or so, then disappear after three or four weeks,” he said. “Releasing open-source models will generate some buzz in the media, but it’s not a sustainable strategy.”\nAfter restructuring its operations late last year to refocus on core AI activities, SenseTime narrowed its adjusted first-half loss by 50 per cent from a year earlier, helped by surging revenue from its generative AI division.\nSales in that unit jumped 72.7 per cent to 1.82 billion yuan, lifting its contribution to overall revenue to about 77 per cent, up from 60.4 per cent from a year earlier. The company’s shares have gained roughly 32.5 per cent over the past 12 months.\nWhile Lin acknowledged that the language capabilities of SenseTime’s flagship SenseNova model were not “industry leading”, he said its strengths in multimodal AI were more critical for embodied intelligence and robotics – areas that had received more government support in China than in the US.\nHe added that SenseTime’s long-term relationships with major enterprise clients gave it a unique edge in gathering data to train the next generation of robots and AI agents with practical applications.\nThe company has also been backing robotics start-ups, including Ace Robotics, chaired by co-founder and executive director Wang Xiaogang, which is set to launch its first robot dog next week.\n“Our approach is similar to Google’s, which was first caught off guard when OpenAI came out with ChatGPT and then made some mistakes that dragged down its stock price,” Lin said. “But later on you see that their deep technological expertise, data accumulation, talent and infrastructure gradually wins out to secure them a lasting advantage.”\n