The loop is the floor.
Everything worth adding attaches to it instead of replacing it. Nobody knows yet how high that goes.
score 0.93 +0.22
- A question arrives.
- A priming step reaches into your own store and pulls out a handful, before the agent has seen anything.
- An agent decides what to do with what it has. If it needs something it does not have, it reaches out for it the same way, and whatever comes back lands in front of it and it decides again. It goes round as often as it needs to.
- Everything either of them reaches for sits outside the machine: the one store they share, and the open internet. Neither is touched, moved or rewritten.
- Several candidate answers come out, not one, and get ordered best first. Depending on who taught you, that step is called reranking, critique or evaluation. It is one step under three names.
- One answer leaves.
- The dashed frame is the part the optimizer is allowed to change. The slider marks are the four settings inside it that somebody once guessed: how much is pulled up front, how it is asked, how many candidates come back, and what counts as better.
- any algorithm
- self-improvement
- synthetic data
- topology mapping
- bigger networks
- web search
- not thought of yet
Previously: The science was done. The tool was missing., on the gap between a result that works and a thing anybody can run.