Ineffable Intelligence is a name that would look excellent at the bottom of a moat slide. Fortunately, the proposition behind the name is more specific than the name.
In April, Sequoia announced its partnership with David Silver’s London-based company. The investor described an ambition to build a system that learns through experience using reinforcement learning, rather than depending on pretraining in the familiar way. The announcement also acknowledged uncertainty about the timeline. This is a proposed research direction, not a declaration that the system has already achieved its goal.1
The name is not the method
Unfamiliar research should not be dismissed merely because it is difficult to explain in a paragraph. Many important ideas require new concepts, and many early descriptions sound incomplete because the work is incomplete.
But difficulty of explanation can become a social advantage. If the audience is afraid of appearing insufficiently visionary, ambiguity receives more respect than it has earned. The person asking a basic question risks becoming the person who does not understand the future.
A healthy research culture needs the opposite arrangement. Basic questions should remain available, especially when the stakes are large. What is the learning signal? What environment supplies experience? How will progress be measured? What result would weaken the hypothesis?
Experience must come from somewhere
The appeal of learning from experience is easy to grasp. A system that can discover rather than only reproduce patterns in existing material offers a powerful possibility. But the word “experience” does not mean “free.”
An environment must provide opportunities to act and feedback about what happened. The quality of that environment matters. So does the relationship between success inside it and usefulness outside it. A system can become excellent at the game it is given while leaving open the question of which game the customer needed solved.
The experiment has to show how learning transfers beyond the conditions that produced it. A good result in one setting gives the next experiment somewhere to start.
The distinction protects ambitious research as much as it protects the audience. If a hypothesis is presented as a hypothesis, a partial result can be informative. If it is presented as destiny, every revision has to be managed as a communications incident.
The investment memo’s favorite tense
Venture writing is unusually fluent in the future perfect. By the time the reader arrives, the company will have transformed the field. The difficult intervening work appears as an expanse of grammatical confidence.
Write down what the investment case assumes the system will learn. Then mark which of those assumptions has experimental support. The unmarked portion is where conviction begins billing by the hour.
A strong technical team can be evidence of ability to conduct the research. It is not a guarantee of a particular result. A promising research direction can justify experimentation. It does not determine the timetable for commercialization. A new approach can be important even if it does not replace every existing approach.
A useful form of reverence
Look for the task, the conditions, the comparison, and the failures. Those details make it possible to learn something from a result without first choosing a camp.
Before admiring the moat, ask what it keeps out.