The new Claude AI can take control of your computer

The next wave of revolutionary AI models will soon arrive: “agent” style models that will be able to take over entire ongoing tasks and functions with complete autonomy. Anthropic's latest AI model gives us a sneak peek by taking over your entire computer.

If you've never come across the idea of ​​an AI agent before, or if you view large language models (LLMs) like Claude and GPT primarily as chat services, Sam Altman, CEO of OpenAI, might be able to help you put things into perspective. perspective. In the short video below, Altman presents the five levels of AI as his company sees things.

Firstly, there are chatbots, and many of us have learned about the incredible capabilities these bots offer in recent years. Then come the “reasoners”: Altman says OpenAI's latest o1 model is the first of them. The third level is “agents” – they will actually be AI machines that people trust to step in and take care of things for them, making their own decisions about how to get the job done.

AI agents will obtain your credit card and your permission to use it. They will have web access and the ability to interact with websites and tools on your behalf. You'll be able to give them a job, trust them to do it, and only contact you when necessary.

The fourth level, says Altman, would be the “innovators” capable of creating new knowledge, and the fifth level would be the “total organizations,” which operate almost without human intervention, a concept that would have been ridiculous to most people with so much. just a few years. years ago, but this seems inevitable now.

It can be said that there are examples of all five levels operating here and there around the world, and they have been around for many years, but in terms of universal availability around the world, none of the major AI companies have released anything that could be an agent called, to this day. Anthropic liberation.

As part of the launch of the new Claude 3.5 Haiku model and the updated Claude 3.5 Sonnet, the company released the following: “We are also introducing an innovative new capability in the public beta: PC use. Available today in an application programming interface (API), developers can tell the cloud to use computers like people do: by looking at the screen, moving the cursor, clicking buttons, and typing text. Watch the introductory video below.

“Using computers is a completely different approach to developing artificial intelligence,” the Anthropic team wrote. “So far, LLM developers have created tools that fit the model, producing customized environments where AI uses specially designed tools to complete various tasks. Now we can make the model tool-friendly: Claude can adapt to the computing environments we all use every day. Our goal is for Claude to take pre-existing pieces of software and simply use them as anyone would.”

Here is an example of an early use case: Anthropologist Pooja Rajan tells Claude that she would like to enjoy a long sunrise walk along the Golden Gate Bridge and asks if he can organize the logistics and set up an entry into the calendar to know what time you should do it. Leave Pete. Open your browser, find out sunrise times and trekking spots, find out travel times from Rajan, then open a calendar and make an entry accordingly.

MBAs like Claude have also become full-fledged programmers, but with this “computing advantage” comes the ability to not only create, edit, and debug code, but also to exit the browser window, run servers, and actually deploy code. :

It is important to note that this new feature is currently very early and limited. For starters, it's only available to developers who access Claude through the backend API, so non-financers can't jump in and start getting it to file our taxes.

You're also limited because you can only see what's happening on the screen as a series of screenshots, which you then use to determine how far to move the cursor and which buttons or keys to press. It is therefore useless in more visually dynamic applications, although Google Deepmind is already involved in the task of building artificial intelligence systems capable of gaming.

Surprisingly, sometimes he appears to get bored and surf the Internet, like in the video below, where he stops running the coding demo Anthropic was trying to record and goes off to enjoy some panoramic photos.

Which, apparently, is also pretty bad. In OSWorld's benchmark test, which evaluates a model's ability to use a computer, humans typically score between 70 and 75 percent, and Claude scored just 14.9 percent. But that's almost double the result of the second-best AI model in its class, and that's pretty much the start.

Of course, giving modern, widely accessible AI models so much access to computers poses security risks; in fact, Anthropic says that's why it released the computer feature in such a rudimentary format. As with OpenAI with GPT-4, the idea here is that opening the doors to the public will give Anthropic the ability to stay well ahead of security risks and escape attempts, so its security capabilities will improve. as the model's wobbly legs become stronger.

This way, Anthropic writes, “we can begin to address any security issues before the risks become too high, rather than adding computing capabilities for the first time to a riskier model.”

Claude switches between multiple apps and windows to get work done

anthropic

It's also certainly a unique opportunity for Anthropic to take on OpenAI to commercialize an important new modeling capability; OpenAI has been talking about agent-level AI for some time. It certainly has something similar, and many expect us to see the first GPT proxy prototypes in the coming weeks or months.

But for those of us trying to keep up with everything happening in this ridiculously fast-moving space, this seems like an important moment. Within a year, it is reasonable to expect that we will all have access to highly efficient client models that can take over computers and perform all kinds of tasks.

And this is another thumbs up moment for this crazy technology, because an AI agent can break down a task into hundreds of steps and then do it? This is starting to look more like an employee than a chatbot. The productivity gains could be huge and the job losses we are already seeing thanks to current AI models will accelerate.

Five or ten years from now, it's hard to see how these agents won't become our primary way of doing things in the digital world. Operating a computer, using a keyboard and mouse, searching for bits of information here to transfer there… how much of your day does this kind of busy work take up? How cool would it be to entrust these tasks to your trusted AI assistant? This is a great transformative moment.

Whereas I always find myself saying: Fasten your seatbelt folks, this train has no brakes.

Source: Anthropy

Related News

Why the Pelicot case changed the country

Why the Pelicot case changed the country

noviembre 22, 2024
Auto supplier Bosch plans to cut 5,550 jobs

Auto supplier Bosch plans to cut 5,550 jobs

noviembre 22, 2024