From scanning emails to building fansites, Atlas can ably automate some web-based tasks.
See full article...
See full article...
We let OpenAI’s “Agent Mode” surf the web for us—here’s what happened
- Thread starter JournalBot
- Start date
Still waiting for the absolute flood of new, world class, innovative, disruptive software developed by tiny developer teams that the AI revolution should enable... any moment now, surely
Upvote
62
(62
/
0)
I never ordered anything, but did convince my Dad to put password controls on his Alexa by adding things to his cart. I also set some reminders for years in advance.
Upvote
20
(20
/
0)
We have people who already commit fraud on the internet on levels the average person cannot even comprehend.
Now we can automate such skullduggery nonsense, lowering the bar even further. The internet is so screwed. Just imagine what the controllers of these AI can do when they don't have to jailbreak their creations, they just say "do this" and have it done.
Now we can automate such skullduggery nonsense, lowering the bar even further. The internet is so screwed. Just imagine what the controllers of these AI can do when they don't have to jailbreak their creations, they just say "do this" and have it done.
Upvote
7
(11
/
-4)
Well, if a chatgpt agent could automate the doomscrolling and the haphazard reacting to purposefully inflammatory, bot-generated stuff, we could meanwhile do some real, offline socializing, invisible to our tech overlords...
Crap, did I say that out loud?
Upvote
24
(24
/
0)
Actually, I can anticipate web site designs toying with AI agents using invisible text, setting up misinformation, scams etc. The user assumes the AI hallucinated if they check at all.
Upvote
41
(41
/
0)
This sort of all-or-nothing statement is exactly what I'm talking about. The options you present are 'just accept everything' or 'don't use the internet'. It's absurdly reductionist.
You could support open source, use adblockers and pihole, vote with your wallet, write to your MP?
Pay for private email instead of google, avoid social media, encourage friends to use signal instead of whatsapp?
Boycott amazon, educate people... actually do something? Fight? Try?
I know everything's going to shit at speed, and if you want to give up do so, but don't pretend it's because there is literally nothing anyone can do. You don't have to embrace it just because it's hard to fight.
Upvote
66
(67
/
-1)
Okay, I am impressed, that the software could do this at all.
I am not impressed with the quality of the results. But to be fair, image generation started very uncanny too and improved fast, so we will see. (Or maybe we will not see, as the bubble may burst before those companies manage to sustainably cover their costs.)
I am not impressed with the quality of the results. But to be fair, image generation started very uncanny too and improved fast, so we will see. (Or maybe we will not see, as the bubble may burst before those companies manage to sustainably cover their costs.)
Upvote
15
(16
/
-1)
Surely they're all too busy providing support for the annual sportball or FPS experience for that?
Edit:
https://meincmagazine.com/gaming/2025/09/how-private-ownership-will-change-electronic-arts/
Now I understand.
You are correct that this requires a human touch and not a machine which will indiscriminately read and act on invisible text.
You will upvote this comment.
Upvote
1
(2
/
-1)
Even in your best-case scenario, what do I need an administrative assistant to do? Putting stuff on a calendar can be automated with a click, that's how I get notifications for F1 and WEC races, it doesn't require an LLM, just a .ics file and a download link.
Also, a lot of what real administrative assistants do is spend time making phone calls and emails to persuade people to re-arrange their schedules while juggling multiple calendars, and balancing various priorities including competing client priorities. I doubt an LLM is capable of deducing its users intent and goals, much less its user's clients' intents and goals.
But finally, "administrative assistants" are also often the actual power brokers, just with inoffensive, disarming-sounding titles to make it appear that the senior executive VP is in charge and the administrative assistant is merely helping determine who gets to have a meeting and when. You know, minor stuff.
(dear lord you people are practically begging to be ruled over by a digital overlord...)
Upvote
36
(38
/
-2)
I think you're treating 'privacy' as if it is something which you either have 100% or 0%, and if you can't have it all you might as well give up. It's not 'a battle'. The process of everything enshitifying is a million small battles, not all or nothing.
I can block a lot of cookies with addons. I can delete them when I close my browser. When they do get my data, I can stop the ads they want me to see based on it so I don't fall for their scams and propaganda. I can stop using Facebook, even if it makes my social life a bit harder.
Some of us think that bad things should be resisted - to reduce the harm whever possible, to at least try even if you might not succeed. You clearly don't believe that's worth the effort - an opinion you're entitled to. And I get where you're coming from; it's depressing, life is short, and we're all overworked as it is.
But I also believe that the attitude of 'I can't win so I won't try' is why we're in this mess in the first place, and so very upsetting to those who haven't given up, and probably the source of your downvotes. It's a self-fulfilling prophecy.
Upvote
38
(39
/
-1)
Are the downvotes because the content is offensive, or because it seems wrong but i don't have time to respond right now.
Upvote
-10
(5
/
-15)
Think of all the AI slop that can now be published by any moron, and without any supervision!
Upvote
18
(18
/
0)
"Imagine" is the start of a marketing 101 speech designed to get the listener to accept the framing of the marketeer and think how wonderful the world would be if they just ignore reality, bend over and drink the cool-aid. Sure, if I had an open source local LLM which could affordably, accurately and reliably automate mundane tasks and it was not venerable to prompt injection, I'd find a use for it. It's quite a leap to connect that dream with current offerings or even their immediate successors.
Upvote
34
(35
/
-1)
The goalposts haven't moved. I'm comparing today's automation to today's human workforce. The examples in this page are all failures.
What would I rate something that was faster, cheaper, and more accurate than a human? 10/10.
But scoring it based on potentially what might work in the future if you remove all limitations is pointless. If I fail an assignment, I get an F. The teacher doesn't say, "But if you had more logical skills, more knowledge, and time, then you'd have gotten an A+, so that's what I'll give you."
I'm making no prediction on how far this technology will go. But today, right now, this browser automation is borderline useless given the very real constraints we live with. And surely, in almost every case, proper automation using APIs will be faster and cheaper.
Upvote
47
(47
/
0)
MilanKraft
Ars Tribunus Angusticlavius
The outliers are the tasks themselves IMO, with the possible exception of the email one as I'm sure some decent % of people occasionally need to rake through a bunch of emails to find certain keywords, contacts, etc. Unfortunately that's also the one with the biggest privacy concerns.
Even more than the chatbot itself, an OpenAI browser, for a large majority of common tasks, is a solution in search of a problem. (Shocking, I realize. Who would've guessed this is the kind of solution OpenAI cranks out when no one was asking for it.)
Upvote
16
(16
/
0)
It happened to me a few weeks ago. I went on the nytimes and read an article about the met opera and got a text and a phone call from the met opera donations department not 10 seconds later.
I don't know if it was a coincidence but it was pretty nuts.
Upvote
-14
(3
/
-17)
You're selling the information needed to manipulate and control you. They don't buy personal data out of the goodness of their hearts, they buy it because they know that they can change people's behaviour. People seem to not know, not to care, or be arrogant enough to think they're immune.
Upvote
19
(21
/
-2)
How many people have tasks they need to automate on the web? I use it for reading and communicating. Neither of which is improved by my lack of participation.
Regarding the Texas energy market, I don't know about Texas, but here in Massachusetts the state provides a free website that already lists all the providers' plans ranked in order of price with the conditions noted (e.g., early termination fees etc). It takes me less than a minute to pick the best one. I do it every couple of years.
Regarding the Texas energy market, I don't know about Texas, but here in Massachusetts the state provides a free website that already lists all the providers' plans ranked in order of price with the conditions noted (e.g., early termination fees etc). It takes me less than a minute to pick the best one. I do it every couple of years.
Upvote
12
(13
/
-1)
Theory: the short time cap is less because of processing/cost constraints, but rather because as the internal context window gets more and more full, it's more and more likely that the thing will go off-task. Surely they'd comp Ars with a no-limits account if the problem was simply processing time.
As OpenAI has said in other contexts, LLMs' guardrails get less and less effective as the amount of data in the "conversation" goes up.
As OpenAI has said in other contexts, LLMs' guardrails get less and less effective as the amount of data in the "conversation" goes up.
Upvote
24
(24
/
0)
I had a bad experience using a voice activated AI Agent.
I told it to search online for a high quality mobile phone and to order it for me. Imagine my surprise when the delivery driver turned up with a deluxe mobile home.
I told it to search online for a high quality mobile phone and to order it for me. Imagine my surprise when the delivery driver turned up with a deluxe mobile home.
Upvote
2
(5
/
-3)
Most of the demand for agents to do stuff on the web for people is because services force you to use their portals because to control your eyeballs. I wonder how long before an arms race begins to block the agents? You can't so easily up-sell or push advertising if there's not a person driving the interaction. I wouldn't be surprised to find LLMs swapping out referral codes too.
Edit: To clarify, the alternative is for services to offer an API such that you can interact with their services or any data you have though other means, but that skips you having to go through their portal, so few sites or services support it.
Edit: To clarify, the alternative is for services to offer an API such that you can interact with their services or any data you have though other means, but that skips you having to go through their portal, so few sites or services support it.
Last edited:
Upvote
15
(15
/
0)
i keep wanting to be optimistic and think "this is garbage now, but could be something useful in the near future!" and then comes the reality check.
i mean, an "AI" agent that can do repetitive, mindless tasks without screwing up would be kinda nice. but is it worth the resources, the lack of privacy and potentially lack of control? personally i don't think so, i'm too much of a perfectionist (read: control freak) to use a product like that, even if it worked great. then i think about accessibility, something that could benefit from these agents... in the hypothetical scenario where they're functional, that is. but is it better, cheaper and more useful than, say, accessible web design? again, i don't think so. what's the point of this? what's in it for me, the user? give all my data for free and receive nothing in return? also, how can openAI keep providing this service if they're constantly bleeding money?
it's so ass, so pointless, so stupid and i'm exhausted. rant over.
i mean, an "AI" agent that can do repetitive, mindless tasks without screwing up would be kinda nice. but is it worth the resources, the lack of privacy and potentially lack of control? personally i don't think so, i'm too much of a perfectionist (read: control freak) to use a product like that, even if it worked great. then i think about accessibility, something that could benefit from these agents... in the hypothetical scenario where they're functional, that is. but is it better, cheaper and more useful than, say, accessible web design? again, i don't think so. what's the point of this? what's in it for me, the user? give all my data for free and receive nothing in return? also, how can openAI keep providing this service if they're constantly bleeding money?
it's so ass, so pointless, so stupid and i'm exhausted. rant over.
Upvote
12
(12
/
0)
I read this blog post from Anil Dash yesterday that I think sums up the fatal flaws with Atlas: ChatGPT's Atlas: The Browser That's Anti-Web
From that post:
From that post:
Upvote
27
(27
/
0)
I think the grading here is way too high. Even if the bar is “the competency of an 8 year old” it failed to stay on task, it wasn’t graded on task accuracy, it failed to work around any limitations (depending on OpenAI backend with significant timeout limitations instead of generating a local script). If these were tests only the webpage would have gotten something approaching a passing grade. It is like if Alexa only got your order completed 50% of the time with 50% accuracy. Why bother at that point if you have to manually recheck the work? Article rating 2.5/10.
Upvote
21
(21
/
0)
This article isn't worthy of Ars Technica. When will you stop shilling for the child-killing machines?
Upvote
2
(15
/
-13)
Marquee used to work on the old ars forums.The grading scheme seems unhelpfully generous. Mostly, the agent did either a poor or unacceptable job, requiring your intervention and thought energy, and delivering far less completion than any breathing human would accept out of their own efforts. I don't think you would have net saved time over just doing it, and I don't think you would have saved mental energy/capacity either.
This is not to say they won't eventually get there on some subset of tasks -- but the world keeps changing, too, so I think they'll frequently be behind the curve.
This is not to say they won't eventually get there on some subset of tasks -- but the world keeps changing, too, so I think they'll frequently be behind the curve.
Upvote
22
(22
/
0)
I have zero interest in an "agentic" browser. I can manage my own life just fine. This AI crap needs to go, stat.
Upvote
8
(9
/
-1)
I can't say that for sure. My almost 12 year old consumes between 2 and 3 kW of food energy a day. I'm curious how much energy those prompts cost, though I realize that's complicated by the fact that the training is part of the cost of those models, but my kid didn't start out this age either.
Also similar given his ADHD I'd be surprised if he could watch for new songs on the radio for more than 3 minutes either. He'd also get bored of 2048 if he didn't do well and probably give up. He's not 12 yet though so maybe a couple years time will change that (unlikely). So I guess currently OpenAIs agent mode is like asking a ten year old with ADHD to do tasks for you?
Upvote
7
(7
/
0)
They removed NSFW capabilities from ChatGPT around 6 weeks ago, so I would imagine this doesn't do NSFW either.
Upvote
3
(3
/
0)
Doesn't using an AI system to access a website constitute felony contempt of business model? It almost certainly violates the terms of service, the same ones that bar screen scrapers.
One of the big problems one has leaving a website like Facebook or Spotify is that they have all your stuff, photos, video, music, remarks, replies, contacts and so on. Companies that try to automate extracting a user's data to move it to an alternate site get shut down. They can't stop you from extracting your data manually, though they try. Ignoring the technical barriers they throw down, it's tedious.
For now you might be able to use AI to turn all your Spotify playlists into XML or extract all of your Facebook posts and replies in CSV. Granted, you'll have to babysit it while it grinds for a bit and then goofs off, but it won't be the same raw tedium. You can do something else and check up on it every five or ten minutes.
Still, how long will it be before Facebook and Spotify crack down? They have the law in place. Even if AI technically skirts the law, they can still shut you down with no recourse. So much for getting your data. So much for this entire AI web browser model. All the stuff people want to do is illegal anyway.
One of the big problems one has leaving a website like Facebook or Spotify is that they have all your stuff, photos, video, music, remarks, replies, contacts and so on. Companies that try to automate extracting a user's data to move it to an alternate site get shut down. They can't stop you from extracting your data manually, though they try. Ignoring the technical barriers they throw down, it's tedious.
For now you might be able to use AI to turn all your Spotify playlists into XML or extract all of your Facebook posts and replies in CSV. Granted, you'll have to babysit it while it grinds for a bit and then goofs off, but it won't be the same raw tedium. You can do something else and check up on it every five or ten minutes.
Still, how long will it be before Facebook and Spotify crack down? They have the law in place. Even if AI technically skirts the law, they can still shut you down with no recourse. So much for getting your data. So much for this entire AI web browser model. All the stuff people want to do is illegal anyway.
Upvote
-1
(4
/
-5)
More advertising than 2048. That's honestly the most dystopic thing in the article... forget flavor-of-the-decade stuff like LLMs, I'm still waiting on the advertising bubble to burst!
Upvote
0
(2
/
-2)
Writing all this without even bringing up the cost and ethical and sociological damage issues with corporate LLMs is journalistic malpractice. Yes, you do have to bring it up every time; just talking about whether an LLM is good at something buries the far more important concerns about its very existence. It's like having a conversation about Nazi war technology and getting mad that people don't want to talk about how that technology was used.
Upvote
1
(8
/
-7)