Infinite Tool Use
a year ago
- #AI Architecture
- #LLM
- #Tool Use
- LLMs should exclusively use tools for output, externalizing intelligence to domain-specific programs.
- Infinite tool use enables multi-resolution text generation, editing, and selective forgetting.
- Current LLMs generate text forward-only, making multi-scale generation and error correction difficult.
- Tool use allows for explicit editing, backtracking, and interleaving abstraction levels in text generation.
- 3D generation benefits from tool use by enabling gradual object creation and human-like workflows.
- Video understanding with tools allows LLMs to re-watch parts and manage notes efficiently.
- Infinite tool use enhances AI safety by making outputs more faithful and legible.
- Training LLMs with infinite tool use involves reinforcement learning and managing unbounded context length.
- Architectures for infinite tool use should have constant inference budgets and support forgetfulness.
- Tool-use paradigm is growing but needs scaling to longer contexts and broader model interactions.