OpenAI Unveils AI Agent To Automate Web Browsing Tasks

The rumors are true: OpenAI today launched Operator, an AI agent capable of performing web-based tasks through its own browser, as a research preview for U.S. subscribers of its $200 monthly ChatGPT Pro tier. The agent uses GPT-4's vision capabilities and reinforcement learning to interact with websites through mouse and keyboard actions without requiring API integration, OpenAI said in a blog post.

Operator can self-correct and defer to users for sensitive information though there are some limitations with complex interfaces. OpenAI said it's partnering with DoorDash, Instacart, OpenTable and others to develop real-world applications, with plans to expand access to Plus, Team and Enterprise users.

Full circle?

[ugen]

Wait, so we built this web thing with interfaces designed for humans to now build a computer system that will interact with services using these human interfaces? When you think the insane inefficiency has nowhere to go - they "do something like that and totally redeem themselves!"

Re:

[KlomDark]

It's typical, like the abbreviation for World Wide Web (3 syllables) being WWW (9 syllables). It's all silliness.

Re:

[zephvark]

It's pronounced "dub dub dub", ya old coot. Now get off your lawn!

Re:

[i kan reed]

Rendered irrelevant by the fact that it's now the Five-or-six-cloud-providers'-data-centers wide web.

http://foscpdcww.example.com/ [example.com]

Thatâ(TM)s even more stupid than RPA

[pcias1]

Spent years on system integrations, API development just to see the âzRPA revolutionâ and costly consultants to get enlightened

Re:

[gtall]

You are not properly considering the beauty in this idea. Imagine you are an ad-slinger wanting to get your blobs in front of eyeballs. Now you get them in front of a bag of bits, and no one really knows how they will react. If done properly, the bot will ignore the ads and concentrate on whatever silly thing it thinks it is doing. If done improperly, just use your imagination.

So the ad-slingers must up their game to get the ads into such a state that the bot will somehow use them. The bot on the other hand

Re:

[Tony Isaac]

There are many, many web sites that don't have any kind of API built, and will never build one. AND web sites that do offer APIs, almost always require authentication and agreeing to API terms of service and/or payment. What alternative do you propose for such sites?

It's more efficient to use an AI bot, than it is to write an API for a web site that has no incentive to build an API.

Re:

[Darinbob]

Keeping up with tiktok is hard, AI can you do this for me while I go text my friends? No wait, also text my friends, I need more sleep.

$200 seems steep.

[Tablizer]

There are other web-scraping/automating apps or sites for less that use pre-AI OCR. One should probably try these first before spending more for AI. The OCR-based tools are time-tested and mature. Whippersnappers often don't look around enough at older tools. "New = Better" as a default is often wrong. (Disclaimer: best tool fit often depends on specific need.)

Thanks to over-use of JavaScript and other web bloaters, HTML analysis alone is often not enough to automate web tasks. BeenThereDoneThat.

Re:

[rsilvergun]

I think the question's going to be are they better at pretending to be regular users and evading bot bans. I suspect they will be because they will follow a lot of the patterns of regular human being would. Nowadays modern antibot tools will do things like look for consistent typing cadence it looks bot like or how you mouse around the screen. An AI trained on actual users browsing would be much better at faking that.

I expect we are going to have a wave of scams, extremist political views being pushed a

Re:

[Tablizer]

So it's really a bot-detection-evasion tool but they are not saying that publicly to avoid not getting sued?

Re:

[Tablizer]

Correction: "to avoid getting sued". I pulled a double-negative. (I need a bot that can catch such typos. Reddit is spoiling me, I can fix typos there.)

Oh goodie More bots

[rsilvergun]

I'm sure this can't possibly be abused for political purposes. Or for just plain old Astro turfing.

Re:

[Tablizer]

In the 90's it was "on the Internet, nobody knows if you're a dog".

Now it's "a bot".

Re:

[paiute]

On the net, you are assumed to be a bot until proven otherwise. And there's no way to prove otherwise.

CAPTCHA and ads

[ukoda]

I wonder how this goes with CAPTCHAs and ads?

If it can bypass CAPTCHAs then maybe we will can get rid of those for humans too? I'm tired of clicking on traffic lights.

Will it count as an ad view if it hits your typical ad infested page? Will some sites try and stop it like those stupid sites that think they are actually important enough that I would disable my ad blocker for them?

Is it just me or are times getting weirder?

Re:

[BishopBerkeley]

Such incredibly intriguing questions. Maybe the human intervention will actually train ChatGPT on Captchas. As for the ads, yeah, Google and company will probably try to argue that the bot counts as much as the human.

Re:

[gweihir]

As for the ads, yeah, Google and company will probably try to argue that the bot counts as much as the human.

Or they will simply claim they cannot tell a bot from a human, but that some fabricated statistics say most are real clicks.
Hmm. If this thing can buy stuff, they may also try to claim that ads are effective on it?

Re:

[gweihir]

Hmm. If this thing can bypass CAPCHAS, filter ads out and, even more important, can effectively filter SPAM, it might be worth the $200 per month. But somehow I do not expect it to live up to the promises. Like at all. No idea what makes me think that.

Agent for my Desktop?

[felixrising]

Where is the agent for my Desktop that can help me do things locally?! That is what I want, and also probably the beginning of the end of many jobs....???

Re:

[timeOday]

Have you taken a look at what Microsoft is building into office / desktop?

Re:

[Turkinolith]

Sadly, it's lacking so far. One of the first things I tried to do with copilot was a very "Siri" style: Look at my emails, find anything from and set up a reminder each day for it.

All it did was say how it couldn't do that and offer some suggestions of how to go about doing it myself.

It's FSD Supervised for Spending Money

[rabun_bike]

So instead of crashing my car it may crash my credit card. Interesting.

Is this another step toward human eradication via.

[ZERO1ZERO]

Is this another step toward human eradication via AI overlords. Now that AI can just do all this asked of it, what's to stop the AI just asking itself to do these things for itself?

The paper clip maximiser problem is getting ever closer.

In theory, you could point an AI overlord at the web, ask it to do whatever and off it trots.

Remember - on the internet, no one knows you are a dog, or in this case an AI overlord coming for you like a heavy chonk hellbent on maximising paperclip production, or world do

Doom Scroll For me

[Turkinolith]

Great, now I can have an AI Agent doomscroll for me while I go outside and touch grass.

What would you say...you do here?

[thesjaakspoiler]

My AI agent does all my work.

distracted

[drolli]

can it simulate a distracted user?