The Shape of the Thing

Ethan Mollick, writing on One Useful Thing, argues that AI has entered a new phase in which users assign work to AI systems rather than prompting them back and forth. He dates the shift to late 2025 and names Claude Code, OpenAI's Codex, and OpenClaw as examples of agents that can be given work — sometimes hours of human work — and return useful results in minutes. He describes the change as an era of managing AIs rather than working with them, succeeding the co-intelligence model he associated with ChatGPT's introduction.

Mollick attributes the shift to what he calls rapid exponential improvement in AI abilities. To illustrate, he revisits his Otter Test, which asks image models to depict an otter on a plane using wifi, and says progress from 2022 to 2025 was rapid and remarkable. He reports that video has become the new frontier, presenting a first-result video from an advanced model by TikTok maker Bytedance, which he notes is still unreleased in the US. The prompt was for a documentary about how otters view the Otter Test; he says the output was nearly perfect aside from a single pronunciation mistake.

He also points to benchmarks. He cites the METR Long Tasks graph, which measures how much human work an AI can complete autonomously with some reliability, noting it has critics and that METR itself has pointed out potential issues, but says most AI ability graphs show the same curve. As examples, he cites Google-Proof Q&A, where graduate students using Google score 34% outside their field and about 70% inside it while the best AIs score 94%; GDPval, where industry experts judge AI against experienced humans and the latest AIs reach or exceed parity with top performers 82% of the time; Humanity's Last Exam; and puzzle-solving. He says each shows similar rapid gains with few signs of slowdown until the top score is reached.