Summary
- OpenAI's new tool, Operator, can execute autonomous online tasks.
- Operator requires human confirmation for actions like purchases and has safeguards against external tampering and limitations like managing calendars.
- Rollout begins with US-based Pro subscribers, future availability for others pending improvements.
OpenAI, the creators of ChatGPT and Dall-E, released a new product today. Called Operator, the "agent" is reportedly able to perform repetitive online tasks, and can even interact with website interfaces.
OpenAI claims Operator is a solution for dull or time-consuming tasks, and has a significant level of autonomy. In the tool's introductory video, Operator responds to a prompt asking it to find a recipe in a website, then purchase the ingredients for it. It then starts performing the tasks on-screen, describing its steps and requesting additional input when needed.
Google Gemini Advanced vs ChatGPT Plus: Which is better?
Both services are great and cost the same, but which is better: Gemini Advanced or ChatGPT Plus?
Should you allow an AI to use your computer?
Would you show ChatGPT your browsing history?
The idea of letting an AI tool use your computer raises all sorts of concerns. Even for simple tasks, like the one in the example video, it can present risks. Imagine Operator buying 1,000 cans of sauce for the pasta, instead of just one, due to AI hallucination.
According to OpenAI, Operator features "advanced reasoning", and also requires user confirmation for actions like purchases. Login credentials can only be provided by a human, too.
In other situations, like online banking or dealing with sensitive information, OpenAI claims Operator either refuses the task outright or requires human supervision. How well that stands is yet to be seen, considering ChatGPT has been tricked in the past to provide dangerous information — like, as reported by TechCrunch, a bomb recipe.
The company states that Operator has a number of safeguards against external tampering, too. Lastly, since all activity happens on-screen, the user can simply call off the workflow.
ChatGPT and Gemini are cool, but they're not where the future of AI is heading
The future of AI is almost certainly going to be models trained for specific tasks, rather than general-purpose LLMs.
OpenAI's Operator will have a staged rollout
US-based Pro users to get it first
Right now, Operator has significant limitations, like creating slideshows or managing calendars — Microsoft's Copilot office tools seem to have an edge there. OpenAI's plan is to perform a staged rollout, beginning with US-based ChatGPT Pro subscribers, who can already use the tool.
Customers on other plans and living in other countries will be able to use Operator "in the future". OpenAI gave no timeline for wider availability, however, stating that it's still trying to improve the tool's "safety and usability at scale".
I led ChatGPT-4o through a dungeon, and here's how it went
After the release of the smartest AI known to man, I just had to know; how good is it at D&D?