30 Months on the Exponential (in Charts)
One Useful Thing Ethan Mollick
The author says AI has jumped from chatbots to agents in 30 months. It can now do hours of work, and the gap keeps widening.
Based on reporting by One Useful Thing, Ethan Mollick — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Thirty months after Co-Intelligence, the author has a new AI book out and used AI to make the charts for the launch post. That choice is the point. In the earlier chatbot era, making those graphics would have meant a tedious back-and-forth in a chat window. Now the work was handed to an orchestrator agent in Claude Projects, which then spun up other agents and came back with the research, graphs, layouts, images and citations in minutes.
The post argues that the biggest change is not just speed, but who is at the top. From early 2023 through early 2024, GPT-4 was the clear leader. Since then, the lead has moved around among models from Anthropic, Google and OpenAI. The author says twenty different AI models have held the lead at some point, and 180 models are now better than GPT-4 was at release. Nineteen can even run on a modern smartphone.
The benchmark story is just as stark. On METR’s Long Horizon Benchmark, AI’s capacity doubled eight times from GPT-4’s release to the end of the measurement. In 2024, the system could handle a four-minute task. By the time the author finished a draft of Co-Existence in April 2026, it could do 17 hours of human work. METR then said its current methods could no longer measure the frontier accurately, while other studies put the figure as high as 17 weeks.
Costs have slid too. A graduate-level question set that used to cost 30 cents in AI usage for one correct answer now costs four hundredths of a cent. But the post keeps a lid on the triumphalism. AI is still uneven: weak at long-form writing, strong enough at math to look superhuman. The author’s answer is a new mode of work, the “conjurer,” where humans summon agents, set the direction and judge what comes back. That is also how the otter video was made, from script to music, with only a prompt and a lot of delegation.
My take — AI-written commentary, not fact-checked reporting
The real story here isn’t that models got smarter; it’s that humans are being pushed into management whether they asked for it or not. The old “just chat with it” habit now looks quaint, like using a smartphone as a very expensive flashlight. Open models matter more than the hype crowd admits, because once tools can run on a phone, control stops being a side issue and becomes the whole game.
Read more about this at: One Useful Thing