Author uses agentic AI to chart 30 months of exponential AI progress in new book

Ethan Mollick used agentic AI to produce charts and visuals for his new book, Co-Existence, demonstrating how far AI tools have advanced since his previous work. He delegated research, graphing, layout, images, and citations to an orchestrator agent in Claude Projects (Opus 5.5), completing in minutes what would have taken hours manually.

He reports that AI capability on the METR Long Horizon Benchmark doubled eight times from GPT-4’s release to early 2026, rising from tasks measured in minutes to at least 17 hours of human work per prompt. METR discontinued the benchmark after concluding its methods could no longer accurately measure the scale of progress. Other studies suggest capability may now reach 17 weeks or more per prompt.

Mollick notes that over 180 AI models now outperform the original GPT-4, with 19 of those models capable of running on a modern smartphone. He highlights that the identity of the leading model has shifted among OpenAI, Anthropic, and Google roughly 20 times during this period.

The author revisited his long-running “Otter on an Airplane Using Wifi” prompt to test image and video generation. With Opus 5.5, he generated a video from script to final cut—including concept, images, music, and technical assembly—without additional human input, a capability unavailable 30 months prior.

Mollick argues that the shift from chatbots to agentic systems is changing how humans work with AI. He introduces the concept of a “conjurer,” who delegates tasks to AI agents, evaluates outputs, and iterates with broad guidance rather than step-by-step oversight.

Despite rapid gains, frontier AI remains uneven: long-form writing remains weak while math problem-solving has gone from poor to superhuman. Mollick emphasizes that even where AI excels, small gaps create large needs for human judgment, taste, and deep knowledge.