ChatGPT Atlas

So openAI's answer to Perplexity's Comet. I'm afraid this will be the future, as these AI-browsers do truly bring value. But they open up the gate for a single Big Tech Winner that truly knows everything about you, and can even control everything on your behalf.

I really hope open-source Browsers like Firefox follow up soon with better alternatives, like on-device LLMs to counteract the "all in the cloud" LLM approach. Of course that would require top-tier ML engineers who mostly all are pay-captured by Big Tech.

I’m not sure even if these browser automation brings much value other than financial analysts, at least for the moment

What value? I haven't used them myself, but from reviews I've seen on Youtube they appear to be flaky and not all that useful. It reminds me of when voice assistants like Siri came out, and it turned out that the only thing they were good for was setting timers, controlling music playback, and gimmicky stuff like that.

The only realistic competition I could see is Apple and they've invested in some small Safari AI features so far. They're also the only one I could trust in terms of privacy and keeping stuff secure and on device.

If the DeepSeek approach to training hyperscaler models "cheaply" after all the hype wears down works, we just need to follow in their footsteps and build open source alternatives to everything.

Frontier models take a lot of money and experimentation. But then people figure out how to train them and knowledge of those models and approaches leaks. Furthermore, we can make informed guesses. But best of all, we can exfiltrate the model's output and distill the model.

If we work together as an industry to open source everything, we can overcome this.

OpenAI has to 100x in five years or they're going to be in trouble.

Models are making it easy to replace SaaS, but also easy to replace other AI companies.

There may be no moat for any of this. The lead is only because they're a few generations ahead, running as fast as they can on the treadmill.

I don't think this hurts China at all.

I've found my browser works fine without LLMs.

>I really hope open-source Browsers like Firefox follow up soon with better alternatives, like on-device LLMs to counteract the "all in the cloud" LLM approach.

That's the first thing that came to mind. Every single action across every single website would be available to OpenAI with this browser. Even if I wanted to leverage something like this it'd have to be a fully local LLM interacting with a huge local DB of my info.

I was sceptical as well before trying it out, but it is unfortunately very practical to ask the AI sidebar about an information-dense website or even github codebase (just as dev examples).

There were of course many browser extensions that did this beforehand (and even better, by hyper-linking the exact text passage of answer segments), but the main differentiator is that most people don't use them/know about them, and this comes with a big tech nametag and it is free.

So the fear is a new Chrome, but "agentic"?

It's not an irrational fear, but the frightening bit depends on whether or not this actually takes off. I very much doubt it ever will. The browser ecosystem, despite being in desperate need of upheaval, is largely resiliant to it because things that work don't tend to get replaced unless they are broken to a point where even the most basic of users are inconvenienced. Or forced to change (due to vendor pressure). Oh, there's the rational fear.

Care to share some of those extensions? I feel like we're on the second/third wave of llm enhanced tools, and there's plenty of good stuff that got passed over from earlier waves of products/tool attempts.

Websites are terrible from a security, usability, accessibility, privacy, and mental health perspective. These tools could be used to fix all of those things. Instead they're just being used to do the same old junk, but like... faster.

I want an AI browser that digs into webpages, finds the information I want and presents it to me in a single consistent and beautiful UI with all of the hazards removed. Yes, I even want the stupid machine to filter content for me. If I tell it "no politics on Tuesdays" it should be able to go find the things I'm interested in, but remove the references to politics.

I understand that there are new risks to this approach, but it could be built with those things in mind. I'm aware that this would give a lot of power to the developers, but frankly trusting thousands/millions of individual weirdos on the open web hasn't turned out to be any better at this point and it's all become consolidated by near monopolies in user-hostile ways anyhow.

Indeed the "no-moat" (weight exfiltration of any openly accessible model) is something that makes me optimistic. That, and the fact that most tasks have "capacity thresholds" which on-device models will increasingly be able to saturate. One example is SQL query generation from text (example: duckdb-text2sql https://motherduck.com/blog/duckdb-text2sql-llm/).

Try it yourself and ignore the hype. For certain things it can be really useful, just keep in mind that it's early days.

I recently used Comet to find out of print movies that were never released on DVD/Bluray, then find them on ebay, then find the best value, then provide me with a list to order. It felt like magic watching it work, and saved me many hours of either doing it myself or scripting it.

I did have to repeatedly break it into ever smaller tasks to get everything to fit within the context windows, but still... it might have been janky but it was janky magic.

Downloaded Comet last week and I was wondering why anthropic/openAI didn't have one. It's a no brainer. Google's hegemony is really going to take a hit, wondering if hardware-software monopolies like Apple are also going to get hit at some point.

At what point does the gimmick become a product feature that people use and come to expect? I set alarms/timers and control music with Siri every day. Siri still sucks for more than that, and I wish it were good for more, but I really do like and use those features.

Yeah, but if I could tell my web browser to read my comment history to figure out what I engage with on HN and to have it read HN for me and write comments for me then I would get all this time back and have more time to do chores around the house!

I share many of your ideas, and I think the best solution would be:

1. a pure data API web (like the original semantic-web idea)

2. open-source browsers which can query for information using on-device LLMs and display it to the user in any UI way they want.

I think 1. will happen, since all search engines will use AI results, with no click through to the original data-owner (website). So there is no more financial incentive to keep a UI website. The question is if the "data API web" will be decentralised or under the control of a few big players that already mined the web.

2. will hopefully happen if on-device models become more capable, the question is by then whether most people are already defaulted to AI browsers from big tech (since they have the money to burn-cash using cloud LLM services to capture market share before on-device LLMs are good enough). The only way to prevent this is user-education and mistrust verus Big Tech, which is what already befell Microsoft's Recall (besides a terrible security architecure).

Surely this is something you can do with simple searches...? Unless you're reaching the level of trying to buy dozens or hundreds of movies.

https://xkcd.com/1205/

This thing is going to be the only way people use their computer in 18 months. Google is dead as a company.

If they get a decent audio interface and get this on phones, apple is in trouble.

> Significant losses: Internal documents revealed that Amazon's devices division lost over $25 billion between 2017 and 2021. A separate report estimated the Alexa division alone lost around $10 billion in 2022.

I would definitely prefer these to be browser plugins that have clear sandboxing instead of owning my entire browser. That said, I do like Comet.

Curious as to what companies you think DO have a moat? I would say their moat is the same as Facebook's. Not permanent (see TikTok) but also super strong.

> So openAI's answer to Perplexity's Comet.

OpenAI probably barely knows or cares that Perplexity/Comet exists.

But why? The AI can go off and search and I can do other stuff while it does.

The point is the multiply how much you can get done, simple searches still require me to be present and to do the work of compiling the list myself, this type of busy work seems much better suited to tools like this that take a sentence or 2 to kick off

This xkcd comic doesn't apply anymore due to AI making generating automation code trivial.

I'm mostly just curious how that experience differs from using chatgpt directly and having it run searches and present the results?

> But they open up the gate for a single Big Tech Winner that truly knows everything about you, and can even control everything on your behalf.

Do not want, I want none of it and no part of it.

I'll use Lynx before I use that.

AI is already infesting search results directly (til I adblocked it), writing the crap on whatever page I just landed on and led me to turn unhook up to "just show the damn video" on YT.

I've yet to see a single use of AI that in any way improves my life and I'm supposed to hand companies who are already too powerful even more of my life/data for that.

I'll pass.

From my point of view it's become very tiresome pretending the emperor is wearing clothes or at least not pointing that out.

> I want an AI browser that digs into webpages, finds the information I want and presents it to me in a single consistent and beautiful UI with all of the hazards removed.

The browser you're looking for already exists :) (partially) its called arc browser on mobile and specifically their browse for me feature

My brain can't work like that. When I'm pursuing things, I always await until the blocking operation is done, something about uninterrupted train of thought and avoiding context switching.

Think about this way: They subsidized the hardware and hardware development too much and created a messy third-party ecosystem rather than focusing these 25 bn USD on developing an AI chat (OpenAI spent less in total).

Could you please suggest some of those extensions? I've been looking for an agentic tool like that, but isolated as an extension.

We have to ask when is Chrome bringing this functionality? With Gemini

TikTok, Facebook, Instagram, etc. have network effect stickiness. YouTube has a vast library that grows in value over time and has been able to lose money for long enough to make the experience unassailable, etc. (Look at Vimeo.)

Maybe LLMs get that by knowing our entire past? But I find that creepier than useful. Right now ChatGPT is at the top of the world, but I don't see it becoming the new unrivaled Google Search. There's just too many people building it, and once OpenAI starts monetizing and "enshittifying" it, other offerings will become more compelling.

AI models put a large swath of mostly tech companies at risk. Including the old business models of titan products like Google Search.

Image/video/world models specifically do this more to legacy media incumbents than LLMs do to complex business processes. We see orders of magnitude savings with marketing, design, film, and possibly in the future game design. LLMs, on the other hand, aren't great at getting your taxes or complex business logic right.

> It's a no brainer. Google's hegemony is really going to take a hit

Wonder which company has the best in class browser today, along with a really really good model, an in-house chip, datacenter infra, and most importantly, is cash flow positive?

Can't be far off.

Definitely the feature but I'm sure Gemini is seconds away (figuratively) from invading Chrome and if it has an agent mode itself, it will eat everybody's lunch in the browser space.

With LLMs we're all becoming managers. Good news is we'll get more done. Bad news is that we'll have to get way better at persisting mid-state process status (I sometimes ask my LLM "could you summarize what we were talking about and why"), tracking outstanding tasks (linear for our agents) and jumping between contexts.

I am also finding work is becoming more tiring. As I'm able to delegate all the rote stuff I feel like decision fatigue is hitting harder/faster as all I spend my time doing is making the harder judgement decisions that the LLMs don't do well enough yet.

Particularly tough in generalist roles where you're doing a little bit of a wide range of things. In a week I might need to research AI tools and leadership principles, come up with facilitation exercises, envision sponsorship models, create decks, write copy, build and filter ICP lists, automate outreach, create articles, do taxes, find speakers, select a vendor for incorporation, find a tool for creating and maintaining logos, fonts and design systems and think deeply about how CTOs should engage with AI strategically. I'm usually burned pretty hard by Friday night :(

This message is so alien to me... I fail to see the real world value of any task you described. Aren't we all fooling ourselves at this point?

How is that different from Google Chrome? Not debating that it's scary to have all that context collected in one place, just saying it's already been status quo on the web for years.

everything AI is doable manually. the point is AI saves us mental toil (maybe).

When ChatGPT was first released I quickly lost count of the number of HN comments predicting Google had about 6-12 months before their search monopoly and ad revenue evaporated.

Alexa came out in 2013, long before transformers (2017) or chatgpt (2022). The state of the art back then was closer to CleverBot than anything useful.

We will likely have decent standalone voice assistants at some point soon but Alexa and Siri were way too early for that.

Arc doesn’t have that functionality last I checked - are you thinking of Dia?

If the movies were never released on DVD/Bluray, what is it that you are searching for on ebay? VHS?

You're discounting the mental effort that goes into describing what you want to an LLM and then verifying the output, which is only worth it if it actually saves you time. Doing a simple search for a few DVDs does not seem worth the effort to me.

Yes, there is a on-device Gemini model for Chrome. (Chrome Built-in AI), they have been testing it for over a year (been participating) and there were hackathons around it (ongoing until Nov 1 https://googlechromeai2025.devpost.com).

In general, I like the idea, but I'm afraid the final implementation will still auto-decide between using the local (for cost saving) and cloud LLM. And potentially outsourcing all your browser-usage LLMs calls to Big Tech like Google is a no-go .

Try search on a browser extension store for "AI Chat website" or similar. Also you can check out https://github.com/nanobrowser/nanobrowser.

You’d be surprised.

I consider myself extremely competent at getting niche results. Couldn’t for the life of me find a certain after market part for a home appliance.

I go ask Claude to find it and it comes back with exactly what I need. One of its queries hit a website with a poorly labeled product that it was able to figure out was exactly what I needed. That product was nested so deeply in results that I would have never found it on my own.

I'm on Macos and Chrome has a Gemini button that functions very similar to Grox on X.. It's useful and unobtrusive.

I also use Comet for specific financial research, also very useful.

I have edge with Copilot. Never use it. Microsoft is gross.

I have an invite to something called Strawberry but at this point, I'm tapped out. Sorry OpenAI, you're late to the game for me.

Maybe next week.

You really think chrome has such a big moat?

You think it's not happening? Do you still use google? I don't personally

VHS, 16mm film, etc. I enjoy owning films that are basically unobtanium.

> hundreds of movies.

That was the goal, yes. In the end I only actually found about 10 I didn't already own, but the AI had to wade through a few hundred to find them.

It wasn't a few. There are literally thousands of films that never made the leap from VHS to DVD. My starting list was only a few hundred, but still it saved me hours.

I just found it. Its called Arc Search and it seems to be really cool.

I think this is the natural endpoint - local models doing something like what Atlas and Codex are doing, acting like a firewall between the user and web. You don't need to wade through the crap online yourself, the AI extracts the useful signal for you, acts like a memory layer for your values and preferences. Don't like the feed ranking - use your agent to extract, filter and rerank by your criteria. Not a big fan of dark UI patterns? Not a problem anymore, the UI can be regenerated. Need a stronger model? Sure, your agent can delegate.

I see this as a big unbundling, since your agent has your ear now, not Google, not social networks, they lose their entry point status and don't control by ranking, filtering and UI what I see or what I can do. They can spread out searches to specialized engines, replace Google for search, walk above all social networks and centralize your activities so you don't have to follow each one individually. A wrapper or cocoon for the user, taking the ad-block and anti-virus role, protecting your privacy and carefully reducing your exposure to information leaks.

All of this only works if you can host your model. But this is where the trend is going, we can already see decent small models, maybe before 2030 we will be running powerful local models on efficient local chips.

Hey, I like your thinking, I wrote about this a bit in another comment here

https://news.ycombinator.com/item?id=45663857

> There may be no moat for any of this.

It's what I was thinking looking at the launch video. Is there a moat? No, there is none. The LLM itself is fungible, what matters is the agentic and memory layer on top. That can be reconstructed easily. All you need to do is export your data from old providers to bootstrap your system in another place. I actually did that, exported from reddit, hn, youtube, chatgpt, claude and gemini - about 15 years worth of content, now sitting on my laptop in a RAG system.

At minimum all you need is a config file, like CLAUDE.md containing the absolute minimum information you need to set your preferences and values. That would be even more portable, you can simply paste text in any LLM to configure it before use. Exporting all your data is the maximalist take on the problem of managing your online identity.

It's more about the fact that if both Chrome and Atlas have AI, then which will people choose?

We are on exactly the same page. I wonder if anyone is working on something like this? It would be incredibly meaningful work, I think.

This is actually along the lines of what I'm working on in my free time at the moment. I am working to extend a local model's memory to allow smaller self-hosted models become a better solution than paying someone else.

Once this is working better, it will allow to extend the abilities of local models without running into the massive issues with context limitations I personally was hitting for self hosted.

Very interesting and I would also love to export my own data! Care to share/elaborate a little further how you went into doing this, especially with platforms that do not offer the option to export your own data.

yes a 100ms search still beats a prompt that takes several seconds to return a response everytime

It depends on what I'm looking for, but by and large yes. It's still much faster.

I feel the same way.

Yet people are using AI for therapy, listening to AI-written novels read by AI, gladly watching AI-made slop. There seems to be real actual demand for it. Feels like total insanity to me; but here we are and facts are facts.

I guess whether the emperor is wearing clothes or not never actually mattered.

> the only thing they were good for was setting timers, controlling music playback, and gimmicky stuff like that.

That's _still_ the only thing they're good at...

Firefox has a page summary using a local LLM, not the same as the full data mining the "Atlas" thing does though.

https://blog.mozilla.org/en/mozilla/ai/ai-tech/ai-link-previ...

> acting like a firewall between the user and web.

A firewall which can set the user on fire.

https://guard.io/labs/scamlexity-we-put-agentic-ai-browsers-...

I'm convinced it's some kind of SV-induced "culture of hyper-productivity", the only apparent goal is to be _more_ productive. Endlessly. There's no goal, or end-state, or reason, other than to just be productive. Downtime? Wasted productivity time! Hobbies? Wasted time unless you can monetise them! Get an AI to do it, and be _even more_ productive!

> I'm afraid this will be the future, as these AI-browsers do truly bring value. But they open up the gate for a single Big Tech Winner that truly knows everything about you, and can even control everything on your behalf.

Oooh boy, thanks, but no thanks. I don't want a BitTech to know everything about me and my browsing habit.

Yet I wonder how these browser behave when you are visiting your Friday night porn site?

Also, your comment made me think about the fact that free AI is dead in a near feature because it probably is economically unsustainable. Consequently, pay to browse might around the corner. Or ..., like for the for all those social networks, we all be the product. Is AI powered browsers the dead of our freedom ?

I saw someone post a Gemini summary from a Google search as their interpretation of a legal question (not a lawyer) and, when they were called on it, scoffed that they hadn’t used AI. People don’t even know when they’re relying on AI at this point. Which isn’t great given the current state of things.

Who needs browsers at all? We have libraries and the Dewey Decimal system

Similar experience. I searched all over for the install manual of my HVAC system (~25 years old at this point). Multiple days turned up nothing. Chatgpt gave me a link to a pdf that Im pretty sure was supposed to be private in about 30 seconds.

> But they open up the gate for a single Big Tech Winner that truly knows everything about you, and can even control everything on your behalf.

Hasn't this gate been open since Chrome conquered the browser market years ago?

Is it me or has there been an enormous uptick in usage of the word "truly" lately? It's starting to give me AI smell anytime I see it.

> these AI-browsers do truly bring value

Do they? Truly?

> a single Big Tech Winner that truly knows everything about you

Again?

What a scam that was! Boy, this Dewey guy really cleaned up on that deal.

https://amphetamem.es/meme?id=seinfeld_03_05_67&timestamp=0%...

I am actually. Let's further discuss: send me an email at this temporary address cdvcrg8zu@mozmail.com to keep in touch.

That's a sweet anecdote. Fortunately, their quarterly numbers say otherwise.

Likewise, let's further discuss: you can send me an email at this temporary address cdvcrg8zu@mozmail.com.

I shot you an email from my vanity domain. Feel free to edit/delete this message.

I hate to tell you, but Siri and other voice assistance are not good at controlling music playback. I have been waiting for that specific feature to get “good” for over a decade now.

The closest one came to handling controlling music playback well in anyway was Cortana but even Cortana didn’t do the things that I needed to do with my voice while controlling music play.

The biggest Used case for me was always hey Cortana hey Siri add a specific song to now playing and play it next. No matter what on any operating system. The voice assistant delete the entire queue and then probably plays the wrong song. Not only will they play the wrong song, but they will play the entire album from the wrong song if it is available so now I’ve gone from a playlist that I was building and listening to on loop for potentially days to some album that I don’t wanna listen to because it could not just add a song next.