Chinese AI agent Manus transcends chatbots, founder of start-up Butterfly Effect says
Founder and CEO Red Xiao Hong said the AI agent is ‘more like a human being’ than chatbots in terms of performance
Touting its edge over chatbots in terms of performance, Manus also claims it can quickly devise an itinerary for a trip to Japan, provide an in-depth analysis of Tesla’s stock and offer real-estate tips in New York based on a family’s requirements.
How AI model predicts heavy rainstorms 4 hours in advance
Hong Kong and mainland Chinese researchers have developed a pioneering AI model
Hong Kong and mainland Chinese scholars have developed a pioneering AI model that uses satellite data to forecast heavy rainstorms up to four hours before they strike, much earlier than current predictions allow.
Researchers from the university and mainland institutions, including national meteorological agencies, published a paper online last month on the “deep diffusion model of satellite data”. It is the world’s first weather forecasting artificial intelligence (AI) model using satellite data for thunderstorm “nowcasting” and predicting rainstorms up to four hours in advance.
Precise forecasts of small-scale, rapidly developing thunderstorms or rainstorms at present are typically limited to between 20 minutes and two hours in advance.
The model is trained with data from a Chinese satellite to capture the evolution of convective cloud structures. It can spot early signs of convection developments to predict rainfall and enable more timely severe weather warnings.
Atmospheric convection refers to the movement of air molecules and cloud particles caused by temperature differences. Strong upwards motion of air produces condensation, forming liquid or solid cloud particles.
Under strong convection, cloud particles grow into massive thunderhead clouds and will fall, and are therefore indicators of impending heavy rain. Convective weather phenomena include lightning, rainstorms, tornadoes and thunderstorms.
According to Dai, conventional weather forecasts, which rely mainly on ground-based radar, scan for signals “from the ground to above” and can therefore often detect changes “only after convective clouds have already formed”.
To overcome these challenges, Dai and his team developed a new AI model and trained it with infrared brightness temperature data collected by China’s FengYun-4A satellite from 2018 to 2021 to accurately capture convective cloud structures.
By harnessing satellite data that monitors cloud evolution from space, the new AI model can visualise and identify signs of rising air development much earlier.
The forecasts are updated about every 15 minutes and cover broad regions including China, South Korea and Southeast Asia. The model may also be useful for predicting heavy rainstorms in regions lacking radar or automatic weather station coverage.
Su Hui, chair professor in the university’s civil and environmental engineering department, said satellite data might not be able to capture highly localised rainstorms affecting only small parts of Hong Kong, as the convection could be too small to show up in the satellite imagery.
Hong Kong last year issued five black rainstorm warnings and experienced 14 tropical cyclones, the highest number since 1946. Ragasa struck Hong Kong in September, injuring about 100 people. It was the second time in 2025 that the highest typhoon signal was issued.
Su said the team was in talks with the Hong Kong Observatory on implementing the technology to boost weather forecasting in the city.
A spokesman for the Observatory said it would maintain collaborations with academia with a view to enhancing weather forecasting capability.
This week, tech giant Nvidia released a study on a model that uses satellite data with ground-based radar data to predict the evolution of cloud and rainfall, offering forecasts up to six hours in advance. The university’s paper was among those cited as a reference.
