AI is tackling longer tasks across nine different domains. Source: METR | 2025 was supposed to be the year of the agent, with AI finally becoming smart enough to take repetitive tasks off our plates. But we might have just jumped straight to the next stage. In some specialized subject areas, AI is going well beyond augmenting human work. Instead, it’s making discoveries that are already shaping the real world. Here are a couple of recent examples:
- Google’s Big Sleep, which was developed last year to spot software vulnerabilities, just made a big catch. It spotted a cybersecurity exploit right as it was about to be used in an attack, giving experts enough time to patch it. Google thinks it’s “the first time an AI agent has been used to directly foil efforts to exploit a vulnerability in the wild.”
- A “self-driving” lab at North Carolina State University recently went online. It uses machine learning to automate chemical experiments, collecting about 10x the data of a traditional lab. The new system could potentially speed up material discoveries from years to days across industries like clean energy and electronics.
â €But it goes beyond agents: Earlier this year, an analysis from the nonprofit research lab METR revealed that AI models are doubling the length of tasks they can handle roughly every seven months. Now, a new batch of research suggests a similar trend is taking place across fields like self-driving, robotics, math, and scientific reasoning. If anything, evidence points to the rate of progress speeding up, with a median doubling time of just 4.5 months across nine different benchmarks.