Descriptions:
Nate Herk walks through Anthropic’s internal report ‘When AI Builds Itself,’ using its data to argue that AGI — defined as a model capable of tackling open-ended problems with no predetermined solution — has effectively arrived in practical terms. The numbers are striking: Claude’s success rate on open-ended coding tasks (where the engineer doesn’t know what the answer should look like) climbed from 26% to 76% in just six months. Maximum autonomous task duration has scaled from 4 minutes two years ago to 90 minutes a year ago to 12 hours today, with one internal model running uninterrupted for 16 straight hours. Anthropic engineers now ship eight times more code per day than in 2024, and on one optimization benchmark a current model achieved a 52x speedup where a year-old model managed only 3x.
Herk is careful to note that Anthropic never uses the word AGI in the report — that framing is his own interpretation of the data. The report itself lays out three forward scenarios: a productivity plateau, continued compounding with humans retaining directional control, and a fully self-improving loop bounded only by available compute. The most sobering section covers Anthropic’s own warning that rare misalignment in today’s models could compound multiplicatively as each generation trains its successor, growing harder to detect even as it intensifies.
For anyone trying to place the current moment on the AI capability curve, this video is a useful guided reading of one of the most data-rich internal disclosures Anthropic has made public.
📺 Source: Nate Herk | AI Automation · Published June 05, 2026
🏷️ Format: News Analysis
