| Mobile Web

Search results for IEEE

AI & Enterprise

KAIST develops tech to cut sensory confusion and hallucinations in multimodal AI

Researchers in South Korea have developed techniques to reduce cases in which AI confuses multiple sensory inputs or generates nonexistent information. KAIST said a team led by professor Yong-man Ro developed two methods to improve sensor understanding in multimodal large language models and reduce confusion between visual and audio data. The approaches are designed to work with less data or be applied without retraining. KAIST said the technology could be used in autonomous driving, rescue robots, drones and medical AI.

AI & Enterprise

AI agents can use up to 136.5 times more energy than conventional AI, KAIST study finds

KAIST said a research team led by Yoo Min-soo (유민수) systematically analysed, for the first time, the compute resources and power consumption used by AI agents in real service environments. The team defined AI agents as a new form of workload that data centre servers and GPUs must process continuously. It found AI agents repeatedly call large language models, increasing computation and response time. A 70 billion-parameter agent consumed 348.41Wh per query, 136.5 times more than simple generative AI Q&A.

AI & Enterprise

KAIST develops active aerodynamics technology for EVs to improve braking and cornering

Researchers at the Korea Advanced Institute of Science and Technology (KAIST) have developed active aerodynamics technology that adjusts airflow around a vehicle in real time to improve performance and safety for high-performance electric vehicles. KAIST said the team built a multi active aerodynamics system integrating several front and rear devices, backed by Hyundai Motor support, and validated it on an actual vehicle. Tests at a circuit confirmed shorter lap times, better braking and cornering, and improved stability.