Honor’s New AI Agent Can Learn and Perceive Your Display

It selected a restaurant, however then could not full the method because the spot it selected required a bank card to verify a reservation, at which level the person needed to take over. You might be versatile in your question—in one other instance, asking it to ebook a “extremely rated” restaurant meant it will have a look at critiques with excessive scores, although the agent does not do any extra analysis than that. It is not cross-referencing OpenTable critiques with knowledge from different components of the online, particularly since all of this knowledge is processed on machine and is not despatched to the cloud.

This sort of agentic synthetic intelligence is the present buzzword within the tech sphere. My colleague Will Knight lately examined an AI assistant that would browse the online and carry out duties on-line. Google late final yr unveiled its Gemini 2 AI mannequin educated to take actions in your behalf. It additionally renews the concept of a generative person interface for smartphones—at MWC 2024, we noticed a couple of firms engaged on methods to work together with apps with out utilizing apps in any respect, as a substitute leaning on AI assistants to generate a person interface as you issued a command.

Honor’s strategy feels considerably like what Rabbit—of the notorious Rabbit R1—is doing with Train Mode, the place you practice its assistant manually to finish a process. There is no must entry an app’s Utility Programming Interface (API), which is the standard means apps or providers talk with one another. The agent memorizes the method, permitting you to then concern the command and have it execute the duty.

However Honor says its self-reliant AI execution mannequin is not educated to observe strict steps—it is able to multimodal display screen context recognition to carry out duties autonomously. As an alternative of getting to coach the assistant to study each single a part of the OpenTable app, it’s able to understanding the semantic parts of the person interface and can follow-through with a multi-step course of to execute your request. Honor highlighted that this course of was less expensive: “In contrast to opponents corresponding to Apple, Samsung, and Google, which depend on exterior APIs—leading to increased operational prices—Honor’s AI Agent independently manages a variety of duties.”

{Photograph}: Julian Chokkattu