AI2 Releases Fully Open-Source Web Agent MolmoWeb: Control Web Pages with Just Visuals

Allan Artificial Intelligence Institute (AI2) recently released a breakthrough fully open-source web agent MolmoWeb. Unlike traditional agents that rely on the underlying web code (DOM), MolmoWeb makes decisions by reading screenshots, marking a major leap in "visually driven" web navigation technology.

Core Technology: Seeing Web Pages Like Humans

The operation logic of MolmoWeb is very intuitive: it captures a screenshot of the current browser window, uses visual analysis to determine the next action (such as clicking, scrolling, or paginating), then executes and repeats. This "what you see is what you get" model makes it more robust than traditional agents because the visual layout of a webpage is usually more stable than its underlying code, and its decision-making process is completely transparent and explainable to human users.

Performance Leap: Small Models Beat Big Ones

Although the parameter scale of MolmoWeb is only 4B and 8B, it has shown the strength of "small but powerful" in performance:

Leading the Rankings: In the WebVoyager test, the 8B version achieved a score of 78.2%, not only ranking first among open-source models, but also approaching OpenAI's proprietary model o3 (79.3%).
Great Potential: Research found that by running tasks multiple times and selecting the best results, its success rate can be further increased to 94.7%.
Precise Positioning: In UI element positioning benchmark tests, it even surpassed Anthropic's Claude3.7.

Data Support: The Largest Open Dataset in History

Not only did AI2 open-source the model weights this time, but it also contributed a large dataset named MolmoWebMix. This dataset includes:

36,000 real browsing tasks completed by human volunteers.
More than 2.2 million screen shot-questions pairs.
Automated synthetic data verified by GPT-4o. Experiments have shown that synthetic data can even outperform human trajectories in guiding the agent to find the "optimal path."

Open Source Spirit and Future Challenges

Currently, MolmoWeb is fully open-sourced on Hugging Face and GitHub under the Apache 2.0 license. Although it still faces challenges in handling complex instructions, login verification, and legal compliance (such as service terms), AI2 firmly believes that only through complete transparency and community collaboration can we truly counter the data monopoly of big tech companies.

ChatGPT Becomes More Mature: OpenAI Launches Age Detection System to Strictly Prevent Minors from Accessing Inappropriate Content

OpenAI has introduced an 'Age Prediction' feature in the paid version of ChatGPT, aiming to identify users under 18 and provide targeted protection. The model uses behavioral signals such as account longevity, activity times, and long-term interaction patterns for intelligent judgment, rather than relying on traditional age input.

Strategic Investment of 100 Million Yuan! China Ruyi Teams Up with Aisi Technology to Embark on a New Era of AI Real-Time Interactive Imaging

AI video company Aisi Technology has reached a deep strategic cooperation with China Ruyi, a listed company on the Hong Kong Stock Exchange, and secured a $14.2 million strategic investment. The two companies will collaborate in areas such as film and television visual design, visual effects production, intelligent generation of promotional materials, and optimization of streaming media assets. China Ruyi will also open up its copyright resources to help Aisi Technology unleash the potential of IP creation.

Meituan Launches LongCat-Flash-Thinking-2601, Further Elevating Open-Source Tool Utilization Capabilities

The LongCat team at Meituan has open-sourced its latest AI model, LongCat-Flash-Thinking-2601, which has achieved the highest level among open-source models in core evaluations such as intelligent agent search, tool utilization, and reasoning. Its core advantage lies in outstanding tool utilization capabilities, enabling it to effectively handle complex tasks that depend on tools, significantly reducing the adaptation cost for new tools in real-world scenarios.

OpenAI's Mental Health Safety Lead Joins Anthropic AI, Sparking Attention to Dialogue System Safety

Andrea Varrone, OpenAI's lead for mental health safety, left the company to join Anthropic, drawing industry attention. She previously focused on researching AI's interaction with user emotions, particularly how AI should appropriately respond when users face mental health issues. This change highlights the importance of AI ethics and mental health topics.