This week a startup called Cognition AI caused a bit of a stir by releasing a demo showing an artificial intelligence program called Devin performing work usually done by well-paid software engineers. Chatbots like ChatGPT and Gemini can generate code, but Devin went further, planning how to solve a problem, writing the code, and then testing and implementing it. Devin’s creators brand it as an “AI software developer.” When asked to test how Meta’s open source language model Llama 2 performed when accessed via different companies hosting it, Devin generated a step-by-step plan for the project, generated code needed to access the APIs and run benchmarking tests, and created a website summarizing the results. It’s always hard to judge staged demos, but Cognition has shown Devin handling a wide range of impressive tasks. It wowed investors and engineers on X, receiving plenty of endorsements, and even inspired a few memes—including some predicting Devin will soon be responsible for a wave of tech industry layoffs. Devin is just the latest, most polished example of a trend I’ve been tracking for a while—the emergence of AI agents that instead of just providing answers or advice about a problem presented by a human can take action to solve it. A few months back I test drove Auto-GPT, an open source program that attempts to do useful chores by taking actions on a person’s computer and on the web. Recently I tested another program called vimGPT to see how the visual skills of new AI models can help these agents browse the web more efficiently. I was impressed by my experiments with those agents. Yet for now, just like the language models that power them, they make quite a few errors. And when a piece of software is taking actions, not just generating text, one mistake can mean total failure—and potentially costly or dangerous consequences. Narrowing the range of tasks an agent can do to, say, a specific set of software engineering chores seems like a clever way to reduce the error rate, but there are still many potential ways to fail. Not only startups are building AI agents. Earlier this week I wrote about an agent called SIMA, developed by Google DeepMind, which plays video games including the truly bonkers title Goat Simulator 3. SIMA learned from watching human players how to do more than 600 fairly complicated tasks such as chopping down a tree or shooting an asteroid. Most significantly, it can do many of these actions successfully even in an unfamiliar game. Google DeepMind calls it a “generalist.” I suspect that Google has hopes that these agents will eventually go to work outside of video games, perhaps helping use the web on a user’s behalf or operate software for them. But video games make a good sandbox for developing and testing agents, by providing complex environments in which they can be tested and improved. “Making them more precise is something that we’re actively working on,” Tim Harley, a research scientist at Google DeepMind, told me. “We’ve got various ideas.” You can expect a lot more news about AI agents in the coming months. Demis Hassabis, the CEO of Google DeepMind, recently told me that he plans to combine large language models with the work his company has previously done training AI programs to play video games to develop more capable and reliable agents. “This definitely is a huge area. We’re investing heavily in that direction, and I imagine others are as well.” Hassabis said. “It will be a step change in capabilities of these types of systems—when they start becoming more agent-like.”
