Developing a computer use model

Oct 22, 2024

Claude 3.5 Sonnet can now interact with a computer like a human—moving a cursor, clicking, and typing via a virtual keyboard—by interpreting screenshots. This is a significant AI breakthrough currently in public beta.

Key Capabilities & Research Insights

• Foundation: Built on previous work in tool use and multimodality.
• Core Skill: Trained to interpret screen images and count pixels accurately to position the cursor. This pixel-counting skill was critical and non-trivial.
• Generalization: Surprisingly, Claude generalized from training on just a few simple, offline applications (e.g., calculator, text editor) to use other software.
• Performance: State-of-the-art for human-like computer use models.
  • OSWorld Benchmark: Claude scores 14.9%, significantly higher than the next-best model (7.7%), but still far below human performance (70-75%).
• Behavior: Can break down written prompts into logical steps, self-correct, and retry tasks when encountering obstacles.

Safety Considerations

• Risk Assessment: Computer use primarily lowers the barrier for applying existing AI skills, not fundamentally increasing them. The model remains at AI Safety Level 2 (no new catastrophic risks identified).
• Strategic Timing: Introducing this capability now, while models are at Safety Level 2, allows for grappling with safety issues before higher-stakes models emerge.
• Key Concerns & Mitigations:
  • Prompt Injection: Malicious instructions in screenshots could override user intent. Developers must take precautions (guidance provided in reference implementation).
  • Intentional Misuse: Classifiers and other methods are in place to flag abuse.
  • Election Safeguards: High alert for misuse undermining electoral trust. Systems monitor and nudge Claude away from activities like generating social media content, registering domains, or interacting with government websites.

Current Limitations & Future Direction

• Current State: Slow, error-prone, and limited. Cannot perform actions like dragging or zooming. The "flipbook" view (screenshots) can miss short-lived events.
• Amusing Errors: During demos, Claude accidentally stopped a screen recording and took a break to browse photos of Yellowstone.
• Future Goal: Make computer use faster, more reliable, and easier to implement. The paradigm shift is from making tools fit the model to making the model fit existing tools.
• Call to Action: Developers in the public beta are invited to provide feedback via a dedicated form to improve usefulness and safety.

"We think this skill—which is currently in public beta—represents a significant breakthrough in AI progress."