Computer Use 2025 Wrapped (www.onkernel.com)

Zengyi Qin, founder of OpenAGI, announced that their computer use model, Lux, has significantly outperformed other leading AI models, including Gemini and Claude, on the Online Mind2Web benchmark. This achievement emphasizes a pivotal shift in how AI can interface and understand human intent, potentially transforming user interaction with computers. Unlike previous models that struggled with causality during pre-training, Lux leverages a new approach by incorporating synthetic data and multi-turn reinforcement learning, allowing it to dynamically learn and adapt its behavior through trial and error rather than mere imitation. The implications for the AI/ML community are profound. By advancing computer use models, Lux enables broader applications, particularly in Robotic Process Automation (RPA) and software quality assurance, where traditional methods are limited. As these models become more adept at performing complex tasks interactively, reliance on DOM-based approaches is expected to wane. Looking towards 2026, key research challenges remain, notably the need for models to reflect on their mistakes and adapt in real-time, enhancing the robustness and versatility of AI systems in ever-evolving environments.