In a demonstration of the ChatGPT agent this week, Neel Ajjarapu, OpenAI’s product manager for the software, gave the chatbot a detailed prompt: Browse Etsy for vintage-style lamps that are under a couple hundred dollars and available with free shipping, then put the best-looking items in his online shopping cart and provide a URL for each one.

OpenAI has also experimented with using the tool to make presentations and PowerPoints, Ajjarapu said, though he cautioned it’s more for making “very early rough drafts” of presentations people can then refine. Microsoft, the company that makes PowerPoint, also offers AI tools to help professionals draft presentations.

Ajjarapu said the AI model that powers the tool uses a computer and web browser to complete assignments. It can also take in feedback from the user while a task is underway and alter its approach, he said. While users are accustomed to chatting in nearly real time with ChatGPT, it can take much longer — several minutes at least — for the chatbot to complete agent-like tasks.

AI agents present new safety and security risks, given the potential for AI to make mistakes or be misused by bad actors. The company said ChatGPT agent is meant to turn down some tasks, including those related to finances or legal advice. There are also a number of actions the tool will seek permission for before carrying out, including making purchases, the company said. For some tasks, such as writing emails, the service will require a user to supervise it.

As with the launches of Operator and Deep Research, the company acknowledged its latest agent effort still needs work.

“It is far from perfect,” said OpenAI Chief Product Officer Kevin Weil during the demonstration. “But I think if we had gone back six months ago or 12 months ago and said this was going to be possible today, we would have been pretty excited about it.”

Source link