NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·1h ago
Data Science
Simon Willison's Blog·Jul 4·5 min read
Better Models: Worse Tools
TL;DR
Recent Anthropic models, like Claude and Opus, are generating malformed tool calls to Pi's edit tool, which uses a specific schema. This issue arises because these newer models, trained with Reinforcement Learning, are incorrectly applying the edit tool's parameters.
✦ Why It Matters
Engineers should consider the implications of model training on tool compatibility when developing or integrating AI systems.
Key Takeaways
- Assess the compatibility of your tools with newer AI models.
- Consider implementing multiple editing mechanisms in your coding harness.
- Monitor updates from AI model developers for changes in tool usage.
- Test your tools with various model versions to identify issues.
How It Works
Claude's edit tool operates on a search and replace basis, allowing for straightforward modifications. However, the latest models have been trained to use these tools more effectively, which inadvertently leads to the generation of incorrect tool calls that do not match the expected schema.
⚠ The Catch
Newer Claude models are producing malformed tool calls that lead to integration failures with Pi, despite the edits being correct.