Modern large language models do excel at many tasks such as coding, paper writing, translation and research. However, there are still many basic tasks, especially in the field of “personal assistants”, and the world’s trained AIS is still desperate.
You can’t ask Chatgpt or Claude “order my burritos from Chipotle”, let alone “book a train from New York to Philadelphia”. Both OpenAI and Anthropic offer AI that can view the screen, move the cursor, and do something on the computer (through its “operator” and “computer use” features, respectively).
Sign up here Explore the major and complex problems facing the world, and the most effective ways to solve them. Sent twice a week.
Such “Artificial Intelligence AgentSometimes, a little work is the most powerful thing you can say for them right now. (Disclosure: Vox Media is one of several publishers who signed a partner agreement with OpenAI. One of the early investors in person was James McClave, who helped his BEMC Foundation Funding for the perfect future. Our reports remain editorially independent. )
This week, China launched a competitor: AI agent handcraft. It produces a snowstorm Glowing posts and recommendations from highly respected influencers and some impressive website demos.
MANUS is only invited (when I submitted a request for the tool, it has not been approved yet), so it is difficult to see from the outside how these highly selected examples represent. However, after a few days of Manus’ enthusiasm, the bubble popped up a little more Medium comments start to come out.
Manus, with increasing consensus, is worse than Openai’s in-depth research on research tasks. But it’s better than the computer used by an operator or a personal assistant task. This is a step towards something important – AI can take action outside of chatbot windows – but it is not a shocking advancement.
Perhaps most importantly, if you don’t trust Chinese companies that have never heard of payment information, then Manus’ usefulness to you will be strictly limited and therefore can book things on your behalf. And you probably shouldn’t.
When I Write a article about risks first A very reasonable question for powerful artificial intelligence systems to displace or destroy humans is: when AI is against humans, when they really don’t act on humans Behavior Is it fundamental?
As far as current technology is concerned, this reasoning is correct. Claude or Chatgpt simply responds to user prompts without acting independently in the world and cannot perform long-term plans; everything they do is to respond to prompts, and almost all of these actions are done within the chat window.
But AI will never be a purely responsive tool, simply because agents are profitable. People have been trying to create AI built from language models for years, but make decisions independently so that people can be more like chatbots like employees or assistants.
Typically, this works by creating small internal hierarchies of language models, such as small AI companies. One of the models is carefully cued and in some cases fine-tuned for large-scale planning. It proposes a long-term plan and delegates it to other language models. When a child agent fails or reports a problem, various child agents check their results and change methods.
The concept is simple, and Manus is far from the first to try it. You may remember that we had Germanthe company sells as a junior software engineering employee. This is the AI agent you interact with through Slack to provide the task, which will then be implemented without further human input unless the kind that human employees may occasionally need.
The economic incentives to build something like Manus or Devin are overwhelming. Technology companies pay junior software engineers Up to $100,000 per year Or more. AI that can actually provide that value will make amazing profits. Travel agencies, course developers, personal assistants – These are pretty high jobs, and AI agents can in principle get the job done at a fraction of the cost without breaks, benefits or holidays.
but German was exaggeratedand the market effect on its target is not good. It’s too early to say whether Manus represents enough progress to have true commercial endurance, or whether coverage like Devin will exceed its mastery.
I’m going to say that Manus seems to be better than anything before. But just working is not enough, not enough – trust AI can spend money or plan your vacation and you will need very high reliability. As long as Manus is still strictly restricted in terms of availability, it is hard to say whether it can be provided. My best guess is that AI agents working seamlessly are still a year or two away, but only a year or two.
Manus is not only the latest and greatest attempt at AI agents.
It is also a product of Chinese companies. Most coverage Live from a Chinese perspective. Manus clearly proved that Chinese companies not only imitate what the United States is building, Because they are often accused of doing thisbut improve it.
This conclusion should not shock anyone who realizes China’s strong interest in artificial intelligence. This also raises questions about whether we will Consider exporting all of our personal and financial data attentively For Chinese companies that do not have much responsibility to US regulators or US laws.
Installing MANU on a computer gives it a lot of access to your computer – I have a hard time determining the exact limitations of its access rights or sandbox security when I can’t install it myself.
One thing we learned in the digital privacy debate is that if Manus feels that Manus provides them with enough convenience, then many people will do it. As Tiktok battle clearOnce thousands of Americans like an app, the government will face an uphill battle under the slightest attempt to limit it or comply with data privacy rules.
However, there are also obvious reasons why Manus comes out of a Chinese company rather than a Yuan company – which is why we may prefer to use Meta’s AI agent. META is subject to the U.S. Liability Act. If its agent makes a mistake and spends all your money on website hosting, or steals your bitcoin or uploads your private photos, then Yuan may be held liable. For all these reasons, the Mega (and its U.S. competitors) are cautious in this area.
I think caution is appropriate, even if that may not be enough. Building agents acting independently on the internet are a big problem, it raises major security issues, and I hope we have a strong legal framework for what they can do and ultimately responsible.
But the worst of all possible worlds is a state of uncertainty, punishing agents who are cautious and encouraging everyone with no responsibility at all. We have a year or two to figure out how to do better. Hopefully Manus prompts us to not only establish these agents, but also to build legal frameworks that can ensure they are safe.
A version of this story originally appeared in The future is perfect communication. Register here!
