Paul’s second AI book, Co-Existence, was published today, roughly 30 months after his first book, Co-Intelligence, which appeared in early 2024. He uses the interval between the books as a way to illustrate how much AI has advanced, and presents several charts that were produced with AI assistance.
Making the charts as a signal of progress
Paul notes that the very method used to produce the charts signals how far AI has come. In the chatbot era, when Co-Intelligence was written, creating such visuals would have required step-by-step interaction with a model in a chat window. In the current agentic AI era, he delegated the work. He used Opus 5.5 inside Claude Projects, where a single “orchestrator” agent talks to the user, spins off other agents as needed, assigns tasks and reports back.
With only a couple of sentences of guidance, the AI completed in minutes what would have taken him hours: all the research, graphing, layouts, images and citations. Because he did not have to do the labor himself, he asked the AI for multiple options, selected some, and gave loose feedback like “make it better” on others. The AI even caught a math error; the author had considered titling the post “28 months on the exponential.”
The end of a single best model
He recalls that from early 2023 through early 2024 there was a single model that was clearly the best in the world: GPT-4 from OpenAI. GPT-4 is the model many people remember from the first years of AI use and is central to ChatGPT’s popularity. Since that period, however, twenty different models have held the lead at various times, all from Anthropic, Google or OpenAI.
Far more models and mobile-capable systems
An even larger change is the breadth of capability: there are now 180 models that outperform GPT-4 as it was at release. Many of these are open-weights models that anyone can use or modify. Nineteen of those models can run on a modern smartphone.
Exponential gains on the METR Long Horizon Benchmark
The presented data comes from the METR Long Horizon Benchmark, which measured how much human work an AI could perform with at least a 50% chance of success from a single prompt. Performance on this benchmark rose in a remarkably regular way: from the release of GPT-4 to the end of METR’s measurement, the amount of work AI could do doubled eight times. In 2024 an AI could perform about a four-minute task; by the time Paul submitted his Co-Existence draft in April 2026, an AI could perform roughly 17 hours of human work from a single prompt.
METR’s measurement then stopped: the benchmark’s creators concluded they could no longer accurately measure AI capabilities with their existing methods. Other studies suggest the actual amount might be much larger — for example 17 weeks of work or more.
Falling costs for hard benchmarks
As capabilities rose, costs fell. Paul’s old favorite benchmark, the Google Proof Question and Answer Exam — which gives problems that routinely challenge graduate students — used to cost about $0.30 in AI usage to obtain a single correct answer. Today, producing the same result costs roughly 0.04 cents.
How humans work with AI is changing
Despite large gains, capability is still uneven: some areas remain weak (long-form writing is still notably poor), while others, such as solving math problems, improved from terrible to superhuman within the same period. Even where AI performs very well, small gaps create large needs for human oversight — at least for now.
In Co-Intelligence Paul described two modes of working with AI: centaur work (tasks split between AI and humans) and cyborg work (deeply blended collaboration). Both keep humans in the loop, but a person can only participate in so many loops simultaneously. With real agents now available, he introduces a third mode: the conjurer. Conjurers summon agents when needed, assign work, and evaluate returns. The charts in this post were made in that conjurer style: agents were summoned, given general direction, and repeatedly told to “make it better.” This is the most powerful use pattern so far, but it raises new risks because we lose visibility into exactly what the agents are doing to produce results.
The Otter test and AI-generated video
Paul has been testing image and video generation tools for years with the prompt “Otter on an Airplane Using Wifi.” He shows what the best image generator, Midjourney, produced 30 months ago. For the current test he asked Opus 5.5 (including some typos in the prompt, reflecting that prompting matters less as models improve): “I have a post showing the what happened over the past 30 months of AI in honor of my new book (published 30 months after the origina). i want you to make the ultimate otter on an airplane using wifi video. just something fantastic, i know you are creative, give it a go.” Using subscriptions to various online services including Google and ElevenLabs, Opus 5.5 produced a short video: script, images, concept, music and the technical production were all handled by the AI. The video has some flaws, but the notable point is that the AI created the entire piece without human help.
Why this matters now
Because of these gains, AI is an increasingly large societal issue. Paul presents a timeline of achievements and controversies since Co-Intelligence, while acknowledging he omitted hundreds of important events — even with AI’s help it is hard to keep up.
The pace of advancement challenges authors, so Co-Existence does not attempt to chronicle every model release or news item. Instead it focuses on trends: AI is improving exponentially but unevenly, and usage is shifting from chatbots to agents that do real work. The book examines what that means for people who must coexist with these systems: where humans retain advantage (Paul argues this comes down to deep knowledge, wide knowledge, taste and agency), when to hand work to AI and when to avoid it. Understanding these trends makes the present and near future less surprising.
Co-Existence is out today; Paul also notes that Opus 5.5 redesigned the book’s website after he instructed it to “make it better.”
