Google DeepMind has launched a new AI model, Gemini 2.5 Pro, designed to interact with web browsers like a human. This innovative model can perform tasks such as clicking, typing, and scrolling based on natural language prompts. For example, users can ask it to “Open Wikipedia and summarize the history of Atlantis,” and the model autonomously fetches the URL, analyzes the site, and describes its actions in real-time.
The Gemini model outperforms similar AI tools from OpenAI and Anthropic in accuracy and speed. It is available via the Gemini API and Vertex AI, with a demo on Browserbase. This model incorporates safety controls to prevent unintended actions, like bypassing CAPTCHAs or compromising data security.
Despite its advancements, Google acknowledges limitations, including potential hallucinations akin to those in other foundation models. Users are advised that it may require confirmation for sensitive tasks, ensuring a balance between utility and safety.
Source link
