Draft report on AI timelines

Hi all, I’ve been work­ing on some AI fore­cast­ing re­search and have pre­pared a draft re­port on timelines to trans­for­ma­tive AI. I would love feed­back from this com­mu­nity, so I’ve made the re­port vie­w­able in a Google Drive folder here.

With that said, most of my fo­cus so far has been on the high-level struc­ture of the frame­work, so the par­tic­u­lar quan­ti­ta­tive es­ti­mates are very much in flux and many in­put pa­ram­e­ters aren’t pinned down well—I wrote the bulk of this re­port be­fore July and have re­ceived feed­back since then that I haven’t fully in­cor­po­rated yet. I’d pre­fer if peo­ple didn’t share it widely in a low-band­width way (e.g., just post­ing key graph­ics on Face­book or Twit­ter) since the con­clu­sions don’t re­flect Open Phil’s “in­sti­tu­tional view” yet, and there may well be some er­rors in the re­port.

The re­port in­cludes a quan­ti­ta­tive model writ­ten in Python. Ought has worked with me to in­te­grate their fore­cast­ing plat­form Elicit into the model so that you can see other peo­ple’s fore­casts for var­i­ous pa­ram­e­ters. If you have ques­tions or feed­back about the Elicit in­te­gra­tion, feel free to reach out to elicit@ought.org.

Look­ing for­ward to hear­ing peo­ple’s thoughts!