Intelligence
What the first three levels already give you: a model that answers well. Necessary, and on its own not enough — answering well is a skill exercised inside what is already known.
Intelligence Inspiration Innovation
Discovery is a search problem.
A research lab at the fourth level of machine intelligence — the level where a model has to produce a claim nobody has made yet, and be right.
The name is a sequence, not a slogan.
I³ is three steps in the order they have to happen. Most of the field is working on the first. We are interested in what it costs to reach the third.
Inspiration
The conjecture intelligence alone does not produce — a guess nobody in the training data made. Cheap to generate and very expensive to generate well, because most guesses are not worth an experiment.
Innovation
The small fraction of those guesses that survive being tested. A claim counts when something outside the model can reject it and does not.
Where the ladder stops being a ladder
OpenAI's five levels describe capability, not chronology. Levels one through three are continuous with each other. Level four is a change of kind: the first level where the model has to be right about something no one can check by remembering.
- 01ChatbotsLanguage you can talk to
- 02ReasonersHard problems at the level of a specialist
- 03AgentsSystems that hold a goal and act on it over days
- 04InnovatorsKnowledge that was not in the training set, or in anyone's head
- 05OrganizationsWork at the scale of an institution
OpenRecursive
Our discovery loop, in the open. Every run is published, including the 608 out of 612 that went nowhere.
Propose Test Update
We are hiring people who have been wrong in public.
Research scientists, systems engineers, and experimentalists who have run a real instrument. You do not need a background in machine learning — you need to have spent a long time on a search that did not work and to know precisely why.