The 200 Concrete Open Problems in Mechanistic Interpretability is a list of 200 concrete research questions in neural net interpretability, proposed in December 2022 by Neel Nanda. (A centralized table of all problems is available on this Google Sheet and this Coda document.) As of market creation (July 20, 2023), only 12 problems are listed as having existing work.
This market resolves PROB to the fraction of all problems that are listed as having existing work on or dating from before January 1, 2025. I plan to use the Coda document to resolve this market (unless it goes down or becomes obviously untrustworthy, in which case I'll try the Google Sheet instead). If there's no way at all that I can find out what problems were solved by 2025, or if human civilization collapses before then, then this market resolves N/A.
To make New Years' Day 2025 more interesting, this market closes exactly 42 minutes after midnight EST.
1,000🏅 Top traders
| # | Name | Total profit |
|---|---|---|
| 1 | Ṁ24 | |
| 2 | Ṁ10 | |
| 3 | Ṁ0 |
After a bit of coding, I found out that there are 29 solved problems. Of these, 28 are listed as solved in the Google Sheet (as you can readily find out by doing a temporary sort), and only one other is listed as solved in the Coda document but not in the Google Sheet (namely problem 6.2). So I'm resolving to 29/200 = 14.5%, which I'm rounding down to 14% since that also matches what the Google Sheet says.