Mollick charts 30 months of AI progress and the shift to 'conjurer' work

Ethan Mollick marks the 30 months between his two books, Co-Intelligence and the new Co-Existence, with a set of charts that try to make the pace of AI progress legible. GPT-4 held the top spot alone from early 2023 to early 2024. Since then, he writes, about 20 different models have traded the lead, all from Anthropic, Google, or OpenAI, and roughly 180 models now beat GPT-4's launch quality, with 19 small enough to run on a phone. On METR's benchmark, the length of task AI can finish at 50% success doubled eight times, from about four minutes in 2024 to roughly 17 hours by April 2026, at which point METR said it could no longer measure the top. The cost of a correct answer on a graduate-level question-answering test fell from about 30 cents to 0.04 cents.

Progress stays jagged, which is Mollick's real point. Math went from weak to superhuman while long-form writing is still mediocre, and small remaining gaps create large needs for a person in the loop. His framework adds a third mode of working alongside AI. Centaur work splits tasks between human and machine and cyborg work blends them, but conjurer work means summoning an agent, handing off a whole job, and judging the result. He calls it the most powerful mode and the riskiest, because you lose sight of what the system actually did. His long-running "Otter Test" for generative media has now largely fallen: Claude Opus 5.5 produced a short video with script, images, and music from almost no guidance.

Why it matters

If you are deciding how to fold AI into real work, the durable signal here is not any single model but the move toward delegating entire tasks to agents and the loss of visibility that comes with it. The specific leaderboard will churn within weeks, so the skill worth building is judging when to hand a job to an agent and when to keep it yourself.