This illustration photograph taken on October 30, 2023
To somewhat limited fanfare, OpenAI has come out with a groundbreaking tool that can use a computer for you, and it’s been out most of this year, available to the company’s elite tier of “pro” users.
Operator is powered by something called a computer using agent or CUA that allows the model to “see” content on the Internet, and take actions that a human would take with a mouse and a keyboard.
That means the Operator can accomplish the full life cycle of a task, such as booking a reservation or signing someone up for something.
I thought it was time to address the remaining barriers to massive adoption of this agentic AI tool, as spring winds into summer.
The Reality around Operator
The first problem is cost. Operator now costs $200 a month, compared to $20 a month for all of the other things that ChatGPT can do.
It seems like the vast majority of average users are waiting for the cost to come down.
Read this review from Mike Todasco at Medium, and you’ll see that he just doesn’t feel like the technology is worth $200 a month:
“If this is the future, then I don’t think we need to worry about AI Agents taking our jobs,” Todasco writes. “Operator is a mess, and sure as heck not worth an extra $180/month. I spent several days trying to find any usefulness in it. But in the end, I had to hang up on this experiment.”
What if it was $40 a month? Where would daily user numbers be at right now?
Of course, we don’t really have anything to compare it to, since OpenAI has not released user numbers for Operator currently.
Vaguely Agentic Systems
You could also make the argument that Operator is vaguely genetic – that although it has the ability to use the Internet, it doesn’t have prebuilt task management tools in hand. It’s more of a do-it-yourself kind of task-based system.
The CUA is absolutely compelling as technology – it combines prior work on computer vision and tool use to offer the kind of environment that we’ve been waiting for a long time.
However, for most people, it’s still too expensive.
Access and Privacy Concerns
Right now, people are adding their own data to ChatGPT in granular ways, to come up with responses.
The same would apply to task-based systems. You’re going to have to decide how much data you trust Operator with in order for it to do its work.
You’ll have to also figure out how much you want to delegate, and what you want to keep for yourself.
And we’ll have to figure out, as users, how to deal with the hacker community, where black hats and bad actors will presumably be trying to get Operator to do things that you don’t want it to do.
However, there’s so much potential here that I thought it warranted a post today to talk about the potential moving forward, and that makes sense, since the company just recently unveiled Operator o3.
Cark Franzen at VentureBeat provides these potential use cases:
“Data engineers can delegate manual web interactions—such as data verification and scraping—with more confidence, freeing time for higher-level optimization work. Security professionals, meanwhile, gain a safer way to simulate user behavior in audits and incident response exercises, thanks to the model’s layered safety mechanisms.”
Over at Reddit, OpenAI did an AMA (Ask Me Anything) on Operator, where VP of Research Jerry Tworek said this:
“We… already have a product surface that can do things on your computer … we’re planning to make some improvements soon and it can become a very useful tool then.”
If you agree with this assessment, we are very close to enormous user bases playing around with the first over-the-counter agentic systems of their kind. We just need a little bit of a discount.
