With Manus, AI experimentation has burst into the open
Watching the automatic hand of the Manus AI agent scroll through a dozen browser windows is unsettling. Give it a task that can be accomplished online, such as building up a promotional network of social-media accounts, researching and writing a strategy document, or booking tickets and hotels for a conference, and Manus will write a detailed plan, spin up a version of itself to browse the web, and give it its best shot.
Manus ai is a system built on top of existing models that can interact with the internet and perform a sequence of tasks without deferring to a human user for permission. Its makers, who are based in China, claim to have built the world’s first general AI agent that “turns your thoughts into actions”. Yet ai labs around the world have already been experimenting with this “agentic” approach in private. What makes Manus notable is not that it exists, but that it has been fully unleashed by its creators. A new age of experimentation is here, and it is happening not within labs, but out in the real world.
Spend more time using Manus and it becomes clear that it still has a lot further to go to become consistently useful. Confusing answers, frustrating delays and never-ending loops make the experience disappointing. In releasing it, its makers have obviously prized a job done first over a job done well.
This is in contrast to the approach of the big American labs. Partly because of concerns about the safety of their innovations, they have kept them under wraps, poking and prodding them until they hit a decent version 1.0. OpenAI waited nine months before fully releasing gpt-2 in 2019. Google’s Lamda chatbot was functioning internally in 2020, but the company sat on it for more than two years before releasing it as Bard.
Big labs have been cautious about agentic ai, too, and for good reason. Granting an agent the freedom to come up with its own ways of solving a problem, rather than relying on prompts from a human at every step, may also increase its potential to do harm. Anthropic and Google have demonstrated “computer use” features, for instance, yet neither has released them widely. And in assorted tests and developer previews, these systems are as limited by policy as technology, handing control back to the user at regular intervals or whenever a complex task needs to be finalised.
The existence of Manus makes this cautious approach harder to sustain, however. As the previously wide gap between big AI labs and upstarts narrows, the giants no longer have the luxury of taking their time. And that also means their approach to safety is no longer workable.
To some American observers, fixated on the idea that China might be stealing a march on the West, the fact that Manus is Chinese is especially threatening. But Manus’s success is nowhere near the scale of that of DeepSeek, a Chinese firm that stunned the world with its cheap AI model. Any company, be it American, Chinese or otherwise, could produce a similar agent, provided it used the right off-the-shelf components and had a large enough appetite for risk.
Fortunately, there is little sign yet that Manus has done anything dangerous. But safety can no longer be just a matter of big labs conducting large-scale testing before release. Instead, regulators and companies will need to monitor what is already used in the wild, rapidly respond to any harms they spot and, if necessary, pull misbehaving systems out of action entirely. Whether you like it or not, Manus shows that the future of ai development will play out in the open.
Post Comment