Resolves according to my best estimate of revenue in inflation-adjusted 2022 USD on January 1st, 2030. I will include and try my best to estimate returns for difficult and illegible cases.

Mechanistic interpretability is here meant to mean 'any work that is about mechanistically understanding in fine detail what the computations in a neural network are doing'. One test of this is if after the work, someone is able to erase and rewrite large parts of the model while maintaining performance, although that is a high bar that doesn't include all work I would consider mechanistic interpretability. I will determine what work is/isn't mechanistic interpretability subjectively.

Examples of possible sources of revenue:
- interpretability consultants producing evidence that an AI system is not discriminating based on race for regulators of e.g. insurance.
- interpretability is used to specify what humans want from an AI system in a way that improves revenue of an AI lab.
- interpretability is used to understand what an AI system has learned about some problem/system so that a company can directly use or sell the knowledge.