Lecture Notes on Anthropic's Claude 37

Introduction to Claude 37

• Claude 37 Release: An exciting development from Anthropic as it's their first explicit reasoning model.
• Installation & Setup
  • Install L Anthropic
  • Ensure API key is set
  • Select Claude 37 latest as the model

New Features

• Max Tokens: Similar to previous models but with new additions
• Thinking Parameter: Set a budget for thinking tokens, which plays a role in the model's reasoning capacity

Model Output

• Response Format
  • Outputs a list with a thinking block and a response block
  • The thinking block shows Claude's thought process
• Latency & Output Viewing
  • Examples can be seen in tools like lsmith where latency is noted
  • Full output includes both thinking and text blocks

Compatibility & Use Cases

• Agent Building
  • Example: Simple agent with arithmetic tools using Lang graph for a react-style tool calling
  • The model shows reasoning in the choice of tools
• Tool Use Visibility: Offers transparency in why and when a tool is used

Model Characteristics

• Reasoning Models vs. Chat Models
  • Different scaling paradigms: Chat models use next prediction, reasoning models use RL on chain of thought
  • System types: System One vs. System Two thinking
  • Interaction modes differ; reasoning models are better for longer, reasoning-heavy tasks

Training and Performance

• Reinforcement Learning: Post-tuning with reinforcement learning is used
• Reasoning Traces: Exposes reasoning traces to the user
• Control: Users can set how long the model can think
• Knowledge Cut-off: October 2024, allowing output of up to 128,000 tokens
• Performance: Strong for software engineering tasks, verified by sbench
  • Example: Claude 37 outperforms Claude 35 (49% vs 62% in some metrics)
• Coding Emphasis: Supports coding tasks, surpasses Claude 35 Sonet

Pricing

• Cost Comparison
  • Input tokens: $3 per million tokens
  • Output tokens: $15 per million tokens
  • Note: Claude 37 output tokens are frequently larger due to thinking

Usage Tips

• STEM Problems: Recommended for challenging STEM problems
• Token Budgeting
  • Complex tasks: Consider over 16,000 tokens
  • Simpler tasks: 4,000 to 8,000 tokens

Latency Considerations

• Control Over Outputs
  • Request detailed outlines with specific word counts
  • Ability to index paragraphs to the outline

Prompting Guidelines

• General Instructions: Avoid step-by-step instructions
• Encouragement: Encourage general problem-solving and detailed consideration

Parameters and Example Usage

• Budget Tokens: Essential parameter
• Thinking & Response Blocks: Easy segmentation of outputs
• Signature Field: A cryptographic token verifying the thinking block

Conclusion

• Powerful Model: Offers high configurability and performance
• Experimentation Encouraged: Ideal for tool calling and extensive reasoning tasks
• Feedback: Encouragement to leave comments for further discussion