Cette nouveauté Claude est DINGUE (le futur de l'IA)

This video explores a new feature of the Cloud application, which allows users to control their computer using AI directives from their smartphone. The feature is currently exclusive to MacOS and requires the latest Cloud application version on both mobile and desktop. Users can activate it by navigating to "send" in the desktop app and synchronizing with their mobile. Activating Chrome extensions further enables the AI to access the browser. The presenter emphasizes that the AI cannot perform irreversible actions like emptying the trash or sending emails without user confirmation. For emails, it can draft responses but requires explicit approval before sending. The first test involved asking the AI to open a web browser to the presenter's community page, "le QG de LIA." Initially, the AI reported that the Chrome extension was not connected, despite it being installed, suggesting a potential bug. After re-checking the installation, the presenter confirmed the extension was active. The AI was then instructed to retry. It successfully opened a new tab titled "le QG de LIA" but did not display the page directly. Next, the presenter asked the AI to summarize comments on the first pinned post within the opened community page. The AI successfully took control of the browser, navigated to the post, and provided a summary of the 42 comments, including notable requests. The third test involved asking the AI to search for a "Notebook LM" training course within the community's "formations" section. The AI successfully navigated to the "Google AI" section and located the relevant training. The presenter then proposed a more complex task: asking the AI to visit Google Gemini and generate a specific image—an "squirrel juggling hazelnuts while snowboarding." The AI opened a new tab, navigated to Gemini, and drafted the image prompt. However, consistent with the safety protocols, it requested user confirmation before sending the prompt to generate the image, as this is an irreversible action. After the presenter confirmed, the AI successfully generated the image. The image showed an "squirrel juggling hazelnuts on a snowboard," with "Havre" displayed on the snowboard, which the presenter noted was Gemini's automatic addition based on location, not a direct request. Following this, the presenter challenged the AI to save the generated image to the computer and then open it. The AI identified the download icon but again requested permission to save the image. The presenter expressed frustration that the AI needed repeated authorization for a task already explicitly requested. Once authorized, the AI successfully saved the image to the downloads folder. For opening the image, the AI took an unconventional route, using the terminal to locate and open the file, rather than directly accessing the download folder. During the process, an orange border appeared around the screen, indicating the AI was in control. The next, and most ambitious, test was to have Claude debate ChatGPT on which AI is superior for content creation. The presenter instructed Claude not to ask for authorization for each message during the debate. Claude enthusiastically responded, "I love the concept, I'm going to open ChatGPT and start the debate. Let's go!" It opened a new tab for ChatGPT and initiated the discussion, stating its identity and the user's request for a debate on content creation superiority. Claude then presented its initial argument, focusing on five criteria for content quality. ChatGPT responded, and the two AIs engaged in a back-and-forth, with Claude suggesting they both tackle the same brief for a direct comparison. However, the debate eventually paused, and Claude requested user input to continue. The presenter asked it to continue, but the application then seemed to freeze, possibly due to resource limits, highlighting a potential bug in the beta feature. Finally, the presenter tasked the AI with creating a web application. The instruction was to create a folder named "app cloud solo" on the desktop, open it in VS Code, and develop a prompt generator web application for Cloud. After some initial application unresponsiveness, the AI resumed processing. It asked for permission to open Visual Studio Code, which the presenter granted. The AI then created the specified folder and opened it in VS Code. To the presenter's surprise, the AI had already developed the application, including an `index.html` file, during the perceived "bugging" period. The AI then opened the application in a browser, demonstrating a functional prompt generator with options for task type, language, detail level, and templates. The presenter concluded by reflecting on the potential of this technology. They envision such AI agent functionalities becoming integrated into everyday voice assistants like Siri, Alexa, and Google Assistant, enabling AI to perform complex real-world actions like grocery shopping, booking doctor appointments, or managing social media interactions. While acknowledging the current security limitations and occasional bugs, the presenter expressed excitement about the future possibilities, despite potential dangers. The video ended with the AI successfully opening the presenter's YouTube channel and prompting viewers to subscribe.