“Our analysis suggests that training runs of around 2e29 FLOP are likely possible by 2030 — a significant increase in scale over current models, similar to the size difference between GPT-2 and GPT-4.”
Claim is about technical feasibility of a ~2e29 FLOP training run by 2030; largest known runs to date remain orders of magnitude below this.
“On AI's default trajectory — what happens if current trends just keep going — by 2030 growth doubles (2% → 4%), with major AI-led scientific breakthroughs.”
Resolves against world real GDP growth around 2030; global growth was running near 3% when the forecast was made.
“2035: Nearly all cognitive work automated. Growth hits ~10%/yr. [The default trajectory of AI if current trends just keep going.]”
“2040: Full automation of society. [The default trajectory of AI if current trends just keep going; the pace depends on three cruxes: robotics, intelligence returns, R&D scaling.]”