3 months ago
- OpenAI researchers are automating themselves away by improving AI, but LLMs remain imprecise and non-deterministic.
- Beagle SCM uses Anthropic's Fable to spot issues in code and make fixes, but LLMs like Claude can be clumsy and error-prone, such as committing wrong directories.
- To handle expensive, slow, and clumsy LLMs, use fast, deterministic tools and integrate them into a formal workflow with self-correction.
- Make tools and processes malleable to automate frequent LLM actions and verify failures, allowing LLMs to automate themselves away in favor of reliable tools.
- Beagle SCM enables LLMs to script routines in JavaScript, with heavy lifting in C, and uses a tooling and workflow layer in JavaScript for tasks like tokenizing files and checking git history.