The open-source community is actively discussing Cerebellum, a new AI-driven web automation system that uses Claude 3.5 Sonnet for browser interaction. While the tool shows promise, community members have raised important questions about data collection, privacy, and local model support.
Data Collection and Privacy Concerns
A significant discussion point in the community centers around Cerebellum's future plans for crowdsourcing browser sessions. According to Han Wang's response, the project aims to implement functionality for converting BrowserStep[] into a portable file format and creating .jsonl files compatible with the transformers library. However, community members have raised valid concerns about personally identifiable information (PII) in authenticated use cases, with the project currently lacking built-in anonymization features.
Local Model Integration Challenges
The possibility of using local models with Cerebellum has emerged as another hot topic. The project maintainer clarified that current local models lack the necessary strong segmentation capabilities for x and y coordinates, making them unsuitable for the task. While some community members suggested potential workarounds using Selenium's WebDriver protocol and DOM modifications, these solutions remain experimental and would require further development.
Technical Implementation
Cerebellum currently relies on Selenium WebDriver v4 for browser interaction, with screenshots being transferred as images through the WebDriver protocol. A community member suggested a potential approach involving DOM modification before screenshot capture, though this would require careful implementation to maintain functionality.
Future Development Path
The project's roadmap indicates plans to address these community concerns, with goals including:
- Creating functionality to save browsing sessions as training datasets
- Developing and integrating a local model with required capabilities
- Improving system performance and user experience
- Expanding LLM support beyond Claude 3.5 Sonnet
Conclusion
While Cerebellum represents an innovative approach to AI-driven web automation, the community discussion highlights important considerations around data privacy and local model support that will need to be addressed as the project evolves. The active engagement between developers and users suggests a collaborative path forward in resolving these challenges.