in the Toolformer paper, it seems what they are doing is : * take a performant language model (copied from some other team e.g. GPT-J) * show the machinery th

in the Toolformer paper, it seems what they are doing is : * take a performant language model (copied from some other team e.g. GPT-J) * show the machinery that it can learn new tokens using one or more tools from a provided set of tools ... e.g. WikiSearch tool * demonstrate that the sequence of characters in the full API call content, has some effect.. e.g. no reply, useless content, or content that helps predict another token. Save that complete set of characters in the API call as an entry * run a learning session with tool calls to APIs, improve the model for resolving known or new tokens (queries with answers) * show the model that it can try new combinations itself (!) * let the machinery try API calls itself to resolve tokens * minimize loss functions for API results comments - this is strikingly different than some RDF hard-wired data store.. it is using huge numbers of failed attempts to find results that work. The results that work are complete API calls as type string. It does not seem to care about "learning" the contents of the API calls, just that it is methodical about remembering API calls that work, and retrying those not a specialist, feedback welcome