When decades become days: Dissecting AI 2027

Abstract

An analysis of the AI 2027 scenario that isolates the “active ingredients” behind its forecast of artificial superintelligence within a few years. Newman identifies four requirements. AI must massively accelerate its own development, reaching the scenario’s 2500-fold speed-up in AI R&D through better experiment selection, lower costs and smarter resource allocation — which would demand automating substantially more than 80% of research work. Today’s “jagged” capability profile must smooth out, so that models handle complex, subjective and long-horizon tasks such as large-scale coding projects and corporate strategy rather than only what is easy to benchmark. Progress must occur largely inside the lab, via simulation rather than the slow real-world feedback that has constrained technologies such as self-driving cars. And humans must build the first superhuman coder largely unaided, since the feedback loop cannot bootstrap itself. He adds a fifth, implicit requirement — that no unexpected obstacles arise — invoking Hofstadter’s law to argue that a forecast requiring simultaneous breakthroughs on several fronts is unlikely to run to schedule. – AI-generated abstract.