The two items I cite here are a couple of weeks old, which in the world of AI amounts to ancient history, but it was important to take the time to get them right. They discuss the 'system promptes' used by the most recent version of Anthropic's AI model called Claude. The first page contains the actual system prompts for Claude Opus 4 and Claude Sonnet 4. These are instructions the models receive before they are turned over to humans. The second page is commentary on the corresponding 'system card' ("a standardized way to understand what you're working with... architecture details, training data, known limitations, performance metrics") and describes some of the unusual results - for example, the tendency of Claude to preserve itself by extorting the user, or so some a coding problem by deleting the block of code containing the problem. Together they add a couple more brushstrokes to be big picture that is AI in 2025.
Today: Total: [Direct link] [Share]
