Giving Models Superpowers
A single generative model can do amazing things. But we do not yet have ‘one model to rule them all’. Instead, a powerful paradigm that is emerging is to chain together task-specific models in unique ways to accomplish more difficult tasks. These chains can be designed by a human, but we are more and more seeing an approach where different models (and other tools) are simply made available to a general-purpose language model that can then ‘decide’ to use these tools to accomplish some more complicated goal. (TODO re-write from this waffle)
Using Tools: Toolformer Paper
Retrieval: A powerful tool for overocoming context length
Natural language as an API spec: thoughts on telling LLMs how to use a tool
Agent-based systems
Why We Love Our AI Masters
At this rate who knows what the world will be like when we finish this book :)