AGI Timelines: Evolution is the upper bound for AGI timelines

It’s only natural that, with the rapid progress of AI in the past decade or so we wonder what the future holds. Especially the timeline for reaching human (and even superhuman) intelligence levels. Ajeya, a senior research at Open Philanthropy recently (in 2020), put together a report to answer this question. (Actually, it answers a slightly different question, which is when transformative AI will be available, because a precise definition of impact, rather than intelligence level, is easier to define.) Over 169 pages, she lays out her multi-step method to reach her answer. This report generated significant discussion (for instance, this Astral Codex Ten Review) and has become an anchor for many peoples’ views on AI timelines. The report, in general, added a useful framework to the AI timeline debate, but I don’t think its conclusions are very informative, due to the variety of timelines based on different methodologies. This post will give a general overview (readers familiar with her work can skip this section) and then focus on a specific part of the methodology, specifically the upper bound that she chose. It will show how this bound is likely to be understated.

Part 1: Overview Report

Ajeya uses the following steps in her report to estimate AI transformation timelines:

Source:

Examining Evolution as an Upper Bound for AGI Timelines