Lecture Notes: Anthropic's Release of Claude 37

Introduction to Claude 37

• Presenter: Lang Cham
• Organization: Anthropic
• Model Release: Claude 37
• Key Feature: First explicit reasoning model by Anthropic.

Usage of Claude 37

• Installation: Install L Anthropic and set the API key.
• Model Selection: Choose Claude 37 latest.
• Parameters:
  • Max Tokens: Set response token limit.
  • Thinking Parameter: New setting for thinking tokens.

Key Features

• Thinking and Response Blocks:
  • Outputs include a "thinking" block and a "response" block.
  • Transparency: Users can see Claude's thought process.
• Latency:
  • Full output latency: 20 seconds.

Compatibility with Agents

• Agent Building:
  • Compatible with building agents.
  • Example: Simple agent with arithmetic tools.
  • Tool Usage: Shows thinking around when and why to use a specific tool.

Reasoning Models vs. Chat Models

• Scaling Paradigms:
  • Chat Models: Next-to prediction.
  • Reasoning Models: Reinforcement Learning on Chain of Thought.
• Interaction Modes:
  • Chat Models: Strong for short chats.
  • Reasoning Models: Better for long reasoning tasks.

Performance and Capabilities

• Post-Tuning: Claude 37 trained with reinforcement learning.
• Thinking Traces: Exposed to users for transparency.
• Control Over Thinking:
  • Users can set thinking budget of tokens.
• Knowledge Cut-off: October 2024.
• Token Output: Up to 128,000 tokens.
• Performance: Improved on software engineering tasks (e.g., sbench scores).

Pricing

• Costs:
  • Input Tokens: $3 per million.
  • Output Tokens: $15 per million.
• Output Tokens: More frequent due to thinking processes.

Tips for Usage

• Usage Recommendations:
  • Suitable for challenging STEM problems.
  • Consider 16,000+ tokens for complex tasks.
• Latency Considerations: Higher thinking tokens increase latency.
• Configurable Outputs:
  • Detailed outlines with specific word counts.
• Prompting Strategies:
  • Avoid predetermined instructions.
  • Encourage thorough and detailed thinking.

Model Parameters

• Budget Tokens: Important for model efficiency.
• Response Blocks:
  • Contain thinking and response segments.
  • Signature Field: Cryptographic token for verification.

Conclusion

• Overall Performance: High emphasis on coding and tool calling.
• Potential for Experimentation: Encouraged to explore different configurations and tasks.

Feel free to leave any comments or questions.