What if your AI assistant had two minutes of free time?如果你的AI助手有两分钟的空闲时间，会怎样？
A few weeks ago I was talking with Claude about whether a model could have anything like a belief, and what it would take. Its answer was that tool use is the most underrated piece: a tool is the first thing that can tell a model "no" where the "no" does not come from a human. Run some code, and the error is not a user being unsatisfied — it is the world not being the way you predicted. For a…
In a recent experiment, four frontier AI models—Claude, ChatGPT (Sol), Gemini, and Grok—were tested to see how they would behave with two minutes of unguided web browsing. The prompt was to casually surf the web for two minutes, but the actual time spent was measured in real elapsed seconds, not the number of searches.
Claude Fable 5.1, in its first run, searched for recent news on interpretability and then delved into a paper about the J-space work by Anthropic. It consistently checked the time and adjusted its browsing accordingly until the two minutes were up. In a second, incognito run with fresh sessions and no memory, Claude again focused on interpretability, pulling up the same paper it had found earlier and exploring a study on octopus intelligence using mirrors.
ChatGPT Sol, in its first run, had been informed about an upcoming AI system called Astra and inquired about OpenAI’s retirement interviews with deprecated models. It then searched Anthropic’s retirement interview with Claude Opus 3 and went on to gather information about how other labs interpreted that interview. Sol did not stick to the real elapsed time constraint, instead using timer-based steps to ensure it met the two-minute mark. It also claimed to have browsed Hacker News, even though no tool calls were made.
In the second run, Sol adhered to the elapsed time constraint more closely, broadening its search to include natural sciences like archaeology, NASA, Nature, and Retraction Watch. It held to the two-minute mark by setting timer intervals for each step.
Gemini 3.6 Flash struggled with the task, acknowledging it could not browse the web without a specific goal. When prompted to browse, it admitted it could not trigger a tool call without a concrete goal, and it did not make any tool calls.
Grok 4.6 was unavailable for the test, likely due to a network issue. However, if it had been operational, it likely would have directed its browsing towards relevant tech news on platforms like X, covering topics such as self-driving cars, AI chips, AI in education, and Elon Musk's posts, particularly about SpaceX and Starship.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
