OpenAGI emerges from stealth with an AI agent that it claims crushes OpenAI (venturebeat.com)

OpenAGI, a stealth startup founded by MIT researcher Zengyi Qin, has unveiled its powerful AI model, Lux, claiming it outperforms established systems from OpenAI and Anthropic in controlling computers while significantly reducing costs. Lux boasts an impressive 83.6% success rate on the Online-Mind2Web benchmark, which evaluates AI agents in real-world scenarios, far surpassing OpenAI's Operator (61.3%) and Anthropic's Claude Computer Use (56.3%). This model is designed to interpret screenshots and execute tasks across various desktop applications, marking a departure from competitors that focus mainly on browser-based tasks. The significance of Lux lies not just in its superior performance but also in its innovative training methodology, called "Agentic Active Pre-training." Unlike traditional language models that learn from text, Lux is trained on visual data, allowing it to autonomously control software by learning action sequences. OpenAGI also emphasizes safety mechanisms to mitigate risks associated with potentially harmful commands. As the race for capable computer-use agents intensifies, OpenAGI aims to establish itself as an independent player, challenging the industry giants and potentially reshaping user expectations regarding AI's capabilities in productivity and task execution.