Is it? If everyone gets their information 100% filtered through a black box llm, the open web becomes something like the CLI. Foreign, confusing, and not well understood by the general public.
Imagine having a filter list enabled in ublock origin that was controlled by amazon or google or Microsoft. And oh wait, you can't even know what gets filtered.
Remember that people generally consume more than they produce on the internet. For certain demographics, that ratio is very extreme.
> If everyone gets their information 100% filtered through a black box llm, the open web becomes something like the CLI. Foreign, confusing, and not well understood by the general public
What? People understand the CLI. Many just don't prefer to use computers that way when alternatives like graphical user interfaces exist. If anything it will be like GUIs to CLIs, depending on where the UX settles.
There are people imagining interactive generative search with generative interfaces, etc.
In my cynical opinion, Operator is probably a push in the direction of getting reinforcement learning out of current user interfaces. It is all about getting developers and other users to dig their way out of employment for temporary gains.
Isn't this exactly what Cloudflare Turnstile was build to prevent?
I can't imagine companies will continue to just give up data to OpenAI. And I mean, most large companies with valuable data have begun selling it to OpenAI. Nobody is giving away anything for free at this point.
And besides, from what I've heard about it on here and elsewhere, Google has already made AdSense an unsustainable income source for websites, with all their requirements for suitable ad placement.
Which users do you think are going to intentionally utilize the captcha-enabled manual-labor-required platforms to do these things (book/reserve/order something) when there’s a faster and easier choice available?
Assuming it actually works and doesn’t hallucinate nonstop, imo it’s going to kill off the “humans browse a website and order something” in the same way e-commerce killed off catalog and phone ordering systems.
Maybe check what Turnstile is/does before commenting? Cloudflare only needs you to solve captchas periodically and then they use your browser fingerprint to classify you as a human accessing a site.
Towards the rest of your comment:
I don't understand how this is a bad thing?
As far as I know, TripAdvisor gets a commission on all sales, so do most e-commerce providers as that's how e-commerce works?
Turnstile is an anti-bot measure. I’m saying users will flock to using bots because it’s easier. Which means sites that require human-at-keyboard will lose business over time by restricting the “easier” route.
Exactly how e-commerce killed off so many holdouts that insisted on in-store/catalog/phone purchases when the internet was growing.
I still don't get your point. When buying a product with a lot of competition, how is Operator different than a price comparison or review site?
I can see it driving purchasing decisions, but I can't see it killing off the actual shop sites? Because the shop is getting a _comission_ on the sale.
I'm getting an ever growing vibe off HN that for a website where a ton of people work on tech, or even in e-commerce serving content to people, they seem to have an awful understanding of consumer behavior.
You're moving the goalposts now. You brought up an anti-bot system. I explained to you that sites which implement an anti-bot system will negatively impact their sales BECAUSE TRANSACTIONS WILL MORE AND MORE COME IN VIA BOTS.
If your site doesn't support/allow agentic bots, it will be similar to not accepting Visa or Mastercard. Yes, cash-only business is a thing. But it's suicide to make that your only payment option for most retail businesses.
> will negatively impact their sales BECAUSE TRANSACTIONS WILL MORE AND MORE COME IN VIA BOTS.
If you assume that every good or service can be infinitely commoditized, then yes, perhaps this will hurt businesses that use this kind of technology.
Realistically though this just isn't the case and there are a lot of products, commodities and services that are unique in their own right and dont't have that kind of competition. So your Visa and Mastercard example is completely moot.
For example, you wouldn't really tell this agent to get you "the best rated shooter game on steam" because that's just not how purchasing decisions in gaming are made.
And on that point, from reading their announcement this sounds more like what critics of AI have been talking about, are you seriously going to automate something that gives you _pleasure_? Like selecting a game that you want to play? Wtf? Because in the scenario youre describing, thats exactly whats happening.
> because that's just not how purchasing decisions in gaming are made.
Illusion of choice. Think about how Netflix, Disney+, Hulu, iOS Appstore, Google Play Store, and even your example -- Steam -- provide recommendation to users. The overwhelming majority of people are already letting an algorithm make decisions for them. Acting like this is where people will suddenly draw the line is amusing.
"Play a new action movie that I'll enjoy" -- "Taken 4 now playing..."
>in the same way e-commerce killed off catalog and phone ordering systems
Something about all this feels a bit backwards though, ultimately all these websites have some underlying API they're hitting. The "right" thing would be to just expose that API directly, and then an LLM could convert user input into a request that hits the endpoint. I think that was supposed to be the vision of Siri and such, but it never materialized because even if 3rd party integration was feasible natural understanding wasn't cracked yet.
Now that we finally have good natural language understanding, instead of executing on that vision, we're going backwards and having agents navigate websites as humans do?? If it does take off, the natural end-point is going to be that pages might have an alternate version that's simplified to make it easier for such agents to do their thing, just as web pages adapted for mobile formats.
> The "right" thing would be to just expose that API directly
100%. You absolutely get it, and also the implications, and also the outcome.
Humans will more and more rely on their AI assistants to take care of mundane things for them. If your site is difficult for an AI assistant to use (due to anti-bot behavior, or an overly complex flow that confuses LLMs, or whatever) your business will suffer. It's gonna be glorious if every retailer starts publishing a swagger API to facilitate automated transactions.
How certain are you that the source itself is trustworthy? Tripadvisor, yelp, and amazon are filled with fake/paid reviews along with the absence of negative reviews which premium members have paid to remove.
My thoughts exactly. Most here will remember accusations that Yelp was taking down negative reviews (were there even suggestions of payola?)
I feel sometimes people grossly overestimate how valuable the web really is. Or at least the majority of it. Wikipedia, archive.org, or my guilty pleasure, YouTube — only a handful of sites really stand out for me. If the rest crashed and burned I'd probably bicycle more.
Not sure why AI agents wouldn't also be connected to an ad platform eventually. Google does this currently.
While the author says that they are bypassing Google, that's not most people, and Google's results are front-loaded with AI answers, so Google results are already giving you specific answers, hallucinations and all. Not sure why average users would long-term switch to not-Google if Google can give them the amount of AI assistance that want or don't want.
The question for me is, whether the web still has enough distributed creative spirit left, so that individuals still feel a need to create.
I.e. instead of feeding my reviews to the rent seeking giants, I can just publish them on my own and they’ll be helpfully ingested and made useful by chattyG.
I think that’s a fair point although I think this is a very different animal.
When it comes to money, or really any kind of action that has consequences, you want the assurance of a deterministic outcome. That is simply not possible with this technology. Maybe it will be but I’ve seen nothing to indicate that so far.
People even let the cars drive on their own and they take their hands off the wheel. Letting an AI do a purchase is far less risky. People are fine with things like this as long as it appears to work reliably enough.
Again. Different animal. That’s highly regulated and in limited markets. You also need to pay close attention so you can grab the wheel when it gets things wrong. And it does. A lot. The underlying tech is not the same either.
I'm having a hard time figuring out if you're saying that like it's a good thing. The tech will definitely get rolled out. It will cause real harm and cost real money because of that lack of regulation you speak of. I am certainly not going to trust it with anything of consequence.
Credit card companies will simply add another bullet point to their zero liability protection offerings: "You are not liable for unintended purchases made by Agentic AI systems".
OpenAI will subsidize banks to offer that safety-net for users of Operator. If Chase says you're 100% not on the hook for purchase mistakes made by OpenAI's Operator -- are you going to take the risk (as you pointed out, it's financial, it matters to people) with Perplexity Pro or Siri?
It doesn't have to be deterministic. Anyone who wants assurances would simply get a full plan of actions to take, and could edit the individual tasks before confirming their execution.
Just like with a real assistant, one could set very clear boundaries as to how much the assistant could spend, on what, etc., or even specify that more money can be spent in, say, a particular period and with a few extra passes (using different models?) just to make sure.
It does really feel like the same animal, because to the degree that I remember the discussions, a lot of it was about a perception of some kind of dichotomy (control or not, being able to 'touch' the product or not, etc.) that really doesn't need to exist.
There are cases where this could be true. I could see, say, asking it to do some research for you or something. Or maybe grab some restaurant recommendations. Really anything that brings you options that you can then make a decision on. If I do end up using this tech, thats how I would feel comfortable using it.
When it comes to things like making purchases and such on your behalf, thats where I disagree. Even with boundaries it's just not deterministic enough for most peoples risk appetite IMO. And when I say "most people" I am not talking about HN folks.
And just to be clear, I do use this technology. I just am very aware that its not this magic bullet that the AI bros want us to believe it is. I used it this morning in fact to help me write up some code I felt too lazy to deal with. It very quickly and efficiently wrote what I asked. I was pleasantly surprised and _almost_ missed the very bad bug that it dropped right into the middle of it all. I'm not giving my credit card to that lol.
I do agree with that. What I'm seeing (and building for myself) really boils down to 'for anything I am remotely uncomfortable trusting the AIs with, I trust them enough to conveniently present me the action or actions they want to take, and I trust myself enough to build a UI that doesn't make it too easy for me to lazily or accidentally click yes on a hallucinated 5000 dollar expense.
The fact that such an AI, with sufficient development, even given what we have now, could present everything on a silver platter up to the execution of the /exact/ commands, and this alone is incredibly useful and can save a lot of time/expense/effort.
The crucial bit is that with, say, a human personal assistant, even if they say they will do exactly what you tell them to do, it's still ultimately inexact. But with an AI, it's trivial to 'decorate' a structured 'command object' (as json, xml, whatever) and from that moment let the deterministic system take over.
If anything, we sometimes want an actual human to /not/ deterministically execute /exactly/ what we tell them to, because perhaps they might know we were angry, drunk, or they just heard something on the news that should invalidate our request.
Either way, my point is that this distinction between 'doing research' and 'doing a thing' is not so dichotomous. In practice, I suspect all but the most autonomously-minded people, and especially non-IT folk, given enough sense of control over the confirmation of potentially 'fuzzy' actions, are happy with an AI doing a ton of stuff for them.
I would argue nearly 100% of those couples would have found someone off the apps and would have had a much happier dating experience if they didn't exist.
Just because they work that doesn't make them a good experience. Show me one dating app with good reviews.
How does the current system pay for bookings like this? And where does it get the customer's contact information to provide to the tour operator? Does the customer fill it all in when signing up for the Operator service?
Many people already have their credit card(s) stored on their phone and computer. I don't think they'll have any problem giving the details to a big company like OpenAI. But I wonder if OpenAI wants to do that, because when real money gets involved, a whole lot of responsibility to the customer for any purchase comes with it.
I would totally delegate an AI to find and purchase my flight tickets. Just let me see the final confirmation and you're good to go.
Cursor already reads my private keys and writes code that goes straight to prod. I've stopped validating GPT 4o output for data analysis. Sure, things go wrong from time to time, but the convenience is unmatched.
And then when you find out they didn't give you the best deal, you got the sponsored pick? And you didn't even get to comparison shop? The AI tax is coming for your choices, one purchase at a time.
> No one. And I mean NO ONE is going to delegate anything like travel or purchasing items to an AI.
That's a bold statement. There are people that get in the backseat or passenger seat while their FSD is driving. There are plenty of other examples of how humans do things because they are reckless or ignorant or any other adjective you want to use
I think the failure of Amazon Alexa is a better reference point here. The goal with that device was to buy stuff through it. No one did because no one trusted it to do the right thing. We just ask it to tell us the weather or play a song.
If the net cost is lower, it will happen. It is unlikely that you (or I) will have much of a say in the matter (besides individual decisions) and it will begin to replace parts of the system we don’t have influence over. Insurability, hiring decisions, risk based costing, customer support, and (looking like potentially) government agency interactions.
I’ve just had an interaction with an insurer who used an ML based system to judge insurability (risk) and make recommendations (requirements) to me for immediate action or policy termination. Fascinating, and totally doable with today’s technology, but it makes me very uneasy.
Anyway, it’s cheaper than hiring an inspector and it gave me some batshit recommendations- including removal of invasive plant material (vines on a trellis, looks great in July) and removing extensive roof debris (leaves, we live in a forest). My recourse is an extremely annoying call center hell, or get on the roof. Guess what I did. And I’m not getting a discount.
Agree.
Adding: LLM models we have today will never produce 100% accurate things. Therefore, they will never be more than a simple tool to use in our daily basis.
To delegate our money/critical-wise tasks to an AI (and other things people dream of), first cientists must find a way different technology we have today.
Spoiler: If it ever happens, it's not something we, normal people, will have the pleasure to be fiddling with firsthand.
The country capable of discover and develop it first will dominate everything.
We are very far from that. I'll think about it when OpenAI have the best Chat app. Nowadays many indie developers create better interfaces in a weekend. What AI are the web engineers there using?
But that assumption is precisely a goal of OpenAI and Anthropic, and it's useful to continue that train of thought to see the bigger picture of where things will go if they get what they want.
>> Takeover mode: Operator asks the user to take over when inputting sensitive information into the browser, such as login credentials or payment information. When in takeover mode, Operator does not collect or screenshot information entered by the user.
I’m not sure I disagree with the point of the article, but I disagree with the specific examples given. I don’t think “human data” quite works like that. TripAdvisor isn’t the authority on what is and isn’t a desirable experience. In fact, a legitimate strategy for researching where to go and what to do on a trip can be to see what’s popular on TripAdvisor and explicitly rule those options out.
Reviews are generally pretty useless unless you know something about the reviewer in order to qualify their opinion. Magaluf is a highly rated tourist destination among some people in society, but I sure as hell don’t want to go there.
AI can’t just aggregate all the opinions and say “Ok specific individual, I’ve checked the internet and here’s what the internet has decided.”
Operator costs $200/mo to try. It doesn't work well. There are regulations still being debated around this. I know, lets make an article full of wild speculations concerning it!
I don't know. As even the likes of Stephen Fry point out, you can talk about the deficiencies of AI (or the cost), but just months from now things will have dramatically changed and your comments will be obsolete.
To me it makes more sense to talk about what may well be coming than to, I don't know, wait to be run over by it.
I am talking about what is now, why would I care if my comments will become obsolete?
The facts and claims in this article too will have changed, come to pass, or be realized.
The author speculating that AI will solve interacting with user interfaces is something that has yet to be seen.
They are making a gigantic leap from (not verbatim quotes, just making a point) "I saw one example of Operator doing (what Claude computer use did back in October, and other projects before that)" to "I think advertising is blown" and "I think search is doomed" to "Sadly I think the open web is about to die.".
Also, people have been talking about this for a few years now. Is there any evidence of it happening yet?
Middlemen are predestined to death at some point or another unless they can keep demonstrating a compelling service, a reason to exist.
Banks are an example of middlemen who continue to exist because they provide a compelling service (everyone needs to store and move monies). Travel companies are an example of middlemen who died because they couldn't provide a compelling service (everyone can do their own bookings now directly with airlines, hotels, et al.).
If "AI" puts search engines like Google Search and data aggregators like Yelp out of business, then that's because their services are not compelling. Good riddance. Middlemen continuing to exist without providing compelling services are just collecting free rent, nobody should have to pay rent for nothing.
As regards the "open" internet, that died a long time ago when most use cases for the internet centralized on a handful of websites and services.
"AI" doesn't change the fact most group conversations happen on Reddit now, that most direct communications happen over Discord or Teams (or email) now, that most traffic gets routed through Cloudflare now, that most companies contract with Amazon Web Services for servers now, that everyone speaks Chrome now, and so on.
There was similar hype around "intelligent agents" taking over the web back in the early 2000's, that might serve as a good corrective. Does anyone know a good writeup about it?
Personally, I think the Author's point about getting Operator to just book your holidays for you or to rely on it for automatically recommending the "best" restaurants or so is ironically just a return to the past when we hired some agencies or friends to do these things for us also. Well, the return of such systems might also open in turn opportunities for arbitarge again; when savvy user who took the time to observe things directly could get better deals.
Yes, and good ones were worth every penny. You can book things online for yourself, but what do you know about the area you are visiting? An agent can recommend things that you would never even think of as being something needing consideration. Even just a phone number to call if you get in a bind.
> Previously if I wanted to know about the French Revolution, I would have Googled it, and likely landed on Wikipedia and maybe 2 or 3 specialty sites. Now, I just ask ChatGPT
This is a common sentiment. I have been using chatgpt for various things, and there are times when I will turn to it for technical help, sysadmin or figuring out some shell / cli commands. However by and large I definitely still prefer googling something. I would much rather read a wikipedia article of the French Revolution than trust whatever chatgpt had to say about it. I figure chatgpt is just spitting out the very same wikipedia article with some rewording and the occasional hallucination and I would rather just read straight from the source. Same with programming questions - between reading a stack overflow thread with the answer and chatgpt I feel a lot better reading stackoverflow, seeing the context the replies, the alternatives etc. and deciding based on that rather than just trusting gpt.
ChatGPT helps me get an initial grasp of the topic from which I can come up with more questions and research to do on my own. Usually I don't use ChatGPT as a source but rather a springboard for ideas.
This is what was often said about Wikipedia. Don't use it as main source of information, but using it as a starting point is fine (if it's just to get a list of sources it quotes).
This is probably going to really bug people, but it seems foolish to use chatgpt as your search engine. You really have to have a more critical eye than say looking at wikipedia or a search result from a reputable website.
I was telling both my son and my father about how chatgpt is not a search engine, and they started testing things and .... well the obvious happened, they easily found mistakes. And they stopped using it so much. They had no idea it wasn't kind of a search engine.
I can ask it questions and get amazing answers, I see the use of it. But it's not a search engine replacement, you have to double check things with search engines. A search engine might return a result with a bogus webpage of course.
People don't know that it's not a search engine! It's much more, but it's also a bit less ;-)
Search engines also return websites with mistakes, sometimes even the stuff they quote directly in their search. But the good part is that they can put the blame onto the individual website if it's wrong. Similarly, you can also use the kind of website to figure out if it's a good source or not.
I don't like the LLM interface that it's all filtered to this super friendly, grammatically correct reply that might however make no actual sense logically, or be plain wrong. With websites I see tons of signals I can use to determine how much value I should put to some information, I can compare different websites, etc.
Also, it's super slow to return a response. Maybe I'm old, but for most of the tasks I use search engines for, I wouldn't use LLM's.
When someone writes or says that they use ChatGPT for a search engine, that just tells me they aren’t serious about facts or education. Anyone who invests two minutes in crosschecking will easily find disturbing misinformation in its “search results.”
Meanwhile, Duck Duck Go provides reasonable sources.
After rather extensive testing, my heuristic is never to ask ChatGPT for facts about anything that can’t be readily verified— and even then don’t trust it until it is verified. The act of verification makes ChatGPT too time consuming to be worth using for education.
It’s useful when I’m trying to solve certain programming problems, though.
Yep - I love ChatGPT for “conversation starter” analysis and ideas. I think it works well to help me figure out what my question is (a program isn’t working and the error message is generic), or to refine something I’ve written or test me with questions.
I cannot imagine using it as a primary source for literally anything, ever.
I only ask LLMs if I want to get a quick fact that isn't that important. For actual reading and knowing the information bit I still defer to Wikipedia.
If I just want to learn conceptual knowledge, having a conversation with ChatGPT isn't much different from doing a DFS through hyperlinks on Wikipedia. In fact, using ChatGPT can be more efficient.
If I want to critically understand the details, Wikipedia, aside from providing references, doesn’t offer much more help than ChatGPT. More importantly, just as LLMs can exhibit hallucinations, community content can also contain biases or even errors.
in the end, it still comes down to relying on oneself.
Nearly everything I've ever asked chatgpt about astrophysics has been critically wrong in some fashion or other. The issue is is that all its training data is polluted by popsci
Eg I just asked chatgpt about the metric expansion of space, and its answer is the common popsci answer, that's incorrect
Next, I asked it what the general equation for general relativistic redshift in an arbitrary spacetime is. It answered
1+z = a(trec) / a(temit)
Where it lists a(t) as the scale factor of the universe. This is not what I asked it, and is the incorrect equation
Next I asked it if the adm formalism is strictly equivalent to general relativity. It said that it is strictly equivalent. This is incorrect, as ADM implies a restriction on the change of the topology of spacetime. A lot of the rest of its explanation of ADM is incorrect
Next I asked it if the ADM formalism is covariant, which it is. ChatGPT says no, because of the foliation of spacetime. I asked it if covariant BSSN is covariant, and it said yes, while it listed reasons why BSSN isn't covariant (which it isn't) that had nothing to do with why it isn't covariant
Then I asked chatgpt if the ADM formalism is more stable the cBSSN formalism. It said that ADM is more stable for highly dynamical systems like binary black hole mergers, which is factually incorrect - it cannot be used for binary black hole mergers. It attributes this to the absence of a clear slicing structure - which is incorrect as they both use the same gauge conditions
I then asked it if the Z4 formalism is generally covariant. It said that it, and adm (which it earlier said was not covariant), are both generally covariant. Now on the plus side its right about z4, but my goodness
Then I asked it if the ADM formalism is hyperbolic. It said no, because its second order in time and space, but if it were first order in time and space, it would be. This is incorrect. ADM is weakly hyperbolic, but strongly hyperbolic systems can be first order in time, and second order in space. It also states:
>This involves introducing auxiliary variables (such as the conjugate momenta of the metric components)
The conjugate momenta is one of the fundamental variables of ADM, so its not an auxiliary variable
--fin--
Literally everything I asked chatgpt here gave a critically wrong answer, from the basic to the complex, and I'm super glossing over the walls of text which are significantly wrong in virtually all the details. Its frustrating because its not saying its not sure, its just confidently incorrect
Information on wikipedia is generally much more accurate, and much more verifiable because it provides links and sources for you to check yourself. Don't use chatgpt for anything
Its hard to overstate just how wrong chatgpt was about everything that I asked it - from the broad explanations, to the details of it. If you ask it "why does the universe expand" (which is a basic question), it'll be wrong
I can often find the sources its mashing together a lot of the time. You can see its repeating 50% of a sentence here, and combining it together with 50% of a sentence there, to make something that's 100% wrong because its not a tool that understands context
For me it feels actively like the opposite of learning. With a human author, there's a chain of reasoning to bring you to a logical conclusion. Even if that conclusion is wrong, there's value in trying to understand what their reasoning is. ChatGPT finds it impossible to author a structured argument, and its logic is a mess of nonsequiturs that there's no value in trying to understand. It makes it a very poor tool for information gathering imo
I tried to ask one of your questions using DeepSeek with deep thinking, it answered:
> Okay, the user is asking if the ADM formalism is strictly equivalent to General Relativity. Let me start by recalling what the ADM formalism is. ADM stands for Arnowitt, Deser, and Misner, who developed a way to reformulate Einstein's equations into a Hamiltonian framework. This is useful for things like numerical relativity and quantum gravity.
So, first, I need to confirm whether ADM is just a reformulation of GR or if it changes the theory. From what I remember, ADM splits spacetime into space and time, using a 3+1 decomposition. They introduce variables like the lapse function and shift vector to handle the foliation of spacetime into spatial hypersurfaces.
Wait, does this decomposition change the physics? Or is it just a different mathematical approach? I think it's the latter. The Einstein equations in ADM form should be equivalent to the original ones, just expressed in terms of different variables. But I should check if there are any subtleties, like constraints or boundary conditions that might affect equivalence.
Also, the user might be confused because sometimes when you change the formalism, especially in constrained systems, there can be issues. For example, in ADM, you have the Hamiltonian and momentum constraints. These need to be satisfied at all times, and numerical implementations have to ensure that. But theoretically, as long as the constraints are satisfied, the ADM equations should be equivalent.
Another point is the initial value problem. ADM is used for that, which suggests it's a way to set up initial data consistent with GR. If the equivalence wasn't strict, then initial data in ADM wouldn't correspond properly to GR solutions. But I think they do, as long as the constraints are met.
However, I should mention that while the formalism is equivalent, practical implementations (like in numerical relativity) might face issues like constraint violation over time, which aren't a problem with the theory itself but with the method of solving the equations.
Also, maybe the user is coming from a background where they heard about different formulations of GR, like tetrad formalism or others, and is wondering if ADM is another such equivalent approach. So, confirming that ADM is a Hamiltonian formulation that doesn't alter the physical content of GR, just the mathematical structure.
Need to make sure to explain that equivalence here means the solutions to the equations are the same, and any differences are in the mathematical handling, not the physical predictions. Also, perhaps touch upon the applications where ADM is useful, to highlight why this reformulation is valuable even though it's equivalent.
I should also check if there's any gauge freedom or choice in the ADM formalism that might lead to different interpretations. For example, the lapse and shift are gauge variables, but that's part of the coordinate freedom already present in GR. So, choosing different lapses and shifts corresponds to different coordinate choices in standard GR, which doesn't affect the physical equivalence.
In summary, the ADM formalism is a reformulation that breaks down Einstein's equations into a set of constrained evolutionary equations. The physical content remains the same as GR, so they are strictly equivalent in terms of physical predictions, but the mathematical approach is different, suited for certain types of problems like numerical simulations or canonical quantization.
---
It's far far beyond my capability to validate if it's right, but I think it would be helpful if the LLM can explain how it thinks?
I agree LLM often being confidently incorrect, it often gives me code that completely un-compilable, users need to review its content very carefully, and we need to do that for content from other sources also
What I want to explain is that LLMs are helpful at many scenarios(not as hardcore as astrophysics, for example, there...
So what it misses is that the ADM formalism is a subset of GR, in that it can only simulate a subset of the full equations for GR. It also gives a lot of reasoning there that is quite faulty with respect to answer the question of if its strictly equivalent
>Wait, does this decomposition change the physics? Or is it just a different mathematical approach? I think it's the latter
This for example is faulty reasoning, because its not an either or situation
>I should check if there are any subtleties, like constraints or boundary conditions that might affect equivalence
This is also faulty reasoning, because it assumes that non equivalence must mean that they produce difference physics
>Another point is the initial value problem. ADM is used for that, which suggests it's a way to set up initial data consistent with GR. If the equivalence wasn't strict, then initial data in ADM wouldn't correspond properly to GR solutions. But I think they do, as long as the constraints are met.
Notice how its trying to sneak a fast one past you. Its saying that because ADM with valid initial conditions maps to GR, therefore GR maps to ADM - which is a logical fallacy! It makes this mistake several times in its chain of reasoning
The issue is its not doing any real reasoning, and isn't capable of actually making deductions. Its about as productive as just.. throwing all the pieces of information on the floor, and picking them up in a random order
This isn't really getting into the weeds with astrophysics, its ADM 101, and deepseek is just.. lying right to your face with confidence, presented under the guise of a chain of reasoning
If I want to know about the French Revolution it’s still Google and Wikipedia. If I want a list of famous characters involved in the French Revolution then it’s definitely chatGPT.
I would explicitly go for private search instead of LLM
I do translations of long podcasts from time to time and it's never accurate and often produces hallucination(I know enough in both languages just using LLM to speed up
Long before OpenAI killed the open web, advertising suffocated it. Advertising is the reason why it would even occur to anyone that it would be more fun to just have a robot book them a vacation than to deal with the misery that is every internet booking website.
Something like Operator may well be the death of the ad-funded web—as TFA observes, who is going to see the ads? But the reason why a tool like Operator has why sort of appeal is because advertising already killed the open web and replaced it with a lifeless corpse.
Even if there were no ads, having a robot book your trips would still be nicer as it is still work to do. We used to have travel agents do all the booking instead and merely got rid of them for cost reasons.
Why trust a robot that may or may not get your request right if you just go to the source and book it yourself? The experience booking anything with a travel agent was always horrible with so much back and forth, do we really want that?
Eh, I disagree. The actual booking process itself is both pretty painless and very much not something I trust to be automated. Enter my card (better yet, pick a saved one), confirm the dates and price, click confirm. I can't think of anything in there that needs to be automated more than it is.
And choosing something to do should be both fun and something you'd want to do yourself—it only isn't fun because businesses aren't incentivized to make it fun, they're incentivized to try to redirect your attention away from what you actually want and instead draw you to their various sponsors.
Oh, the stories that could be told about travel agents and the trips they booked!
Delegation sounds great and sometimes is, but be careful what you wish for, especially when the your delegate is actually working for somebody else, as will be the case for most LLM agents and was the case for most travel agents.
My thoughts exactly. The AI gatekeepers will also become gatekeepers of consumer choices in a similar way as Google is now. That goes for both sides of google's business: search and ads. If you're asking for "the hest car brand", who is to say the AI cannot he swayed by some sweet advertising dollars? Of it books you a guided tour directly, you'll not even notice of one tour guide is preferred.
That may sound pessimistic, bit honestly I have zero confidence that in Trumps's USA, OpenAI will have to give a flying fuck about users' best interests or honest advertising or any of that stuff...
This makes me wonder. OpenAI isn't the only company offering computer use. The list of companies and models that do this will only grow.
Meaning advertiser's will have to be selective which company they pay to get the most exposure to their target human customers via the agents. Will we see affiliate programs for AI agents which in turn promote products to the users? and we end up with the same shit show we have today?
Or what if eventually everyone has their own personal AI that can bypass the ads. Will we just decide that advertising is a drag on society and kill off that industry for good?
> and we end up with the same shit show we have today?
Absolutely certain. Furthermore, this can maximize exploitation of each individual human as the providers have such rich profiles about each human that they can customize pricing to extract the maximum amount of profit. It is by design.
People have spent lots of words talking about the how the saturation of generative AI output will spoil future trainig, which is true, but is most immediately a problem for the AI model developers themselves.
But even with them only acting as consumers, they break almost of the economic and social bases for how the internet works.
Advertisers want their ads to be seen by humans, experts/pontificators want their words to to be seen by humans, artists want their images to be seem by humans.
These things provide the foundations of the internet as we know it.
If online tools are being used by bots who are unaffected bu ads, and texts are summarized by bots instead of read directly, and images are synthesized dynamically instead of appreciated directly, the internet itself breaks.
As the author notes, the next most likely equilibrium involves an even more commercial, more expensive, more gated, more impersonal future than we might have already seen ourselves careening towards just a few years ago.
How does anyone figure out if something is worthwhile, if they haven't used it? They have heuristics, one of which is comfort and familiarity. That's why advertising exists: to get a brand in front of your face, so you recognize it.
Maybe the bots end up with a similar heuristic as they crawl their sources. Perhaps the advertisers will be perfectly happy to pay, not for human eyeballs but for the crawlers.
This is, of course, absurd. But since I find the capabilities of the bots opaque, why not absurdity.
Maybe the World Wide Web will be will be reduced to a series of blog rings where nerds and oddballs share their peculiar interests and discuss eclectic ideas.
I'd be delighted to see PG behind a paywall so I would never have to consume his content again, but I have the feeling the man is too affluent to ever need the side hustle.
Why book a vacation when SORA can make a video of you and your family frolicking in the summery tropics that you post to your followers? They don't know you so they won't think your lying, those people used to be called friends.
> For informational content, like Wikipedia, TripAdvisor, Yelp, recipe websites... perhaps this will end up looking something like the journalism industry that aggregate and publish the information to AI companies, so they can avoid freeloading by agents.
Why publish it to AI companies? The Internet is ripe for mass syndication of textual content and putting it behind subscriptions, just as YouTube has done with video content. Syndication was the distribution and business model for this content before the WWW, the advantage we can get now is no barrier to entry and no distribution costs. (Hosting costs are so low to not even count for text).
At the same time we also get rid of the ad-ridden, scam-infest, bot-crawling Internet of the current day.
Said I, on an open forum where nobody pays a dime...
> Something else I think could happen as a result of this is a web where we verify user/client identities, the same way we currently verify server identities (via TLS certificates). The cost of running bots is plummeting. I could forsee a world where sites like Tripadvisor require a client certificate to connect. That certificate could be provided by paid services (to prevent bot spam), or maybe by device manufacturers like Apple, so you get one certificate per device to prevent bot spam.
Have client certificates worked in any niche? We now have SSO by Google/Apple/Facebook/X, but those identities are also used by bots.
> For informational content, like Wikipedia, TripAdvisor, Yelp, recipe websites... perhaps this will end up looking something like the journalism industry that aggregate and publish the information to AI companies, so they can avoid freeloading by agents.
I'm pretty sure Wikipedia will not do any of those things, as it's a project ran by volunteers, and they do want to disseminate their content as widely as possible. The revenue from the "please donate" feature is only used for website development and to pay the servers.
For the other websites, yeah, those might change and lock things down, require accounts to browse, etc, to prevent being scraped. Already now when I look for house listings or onto linkedin, I see all sorts of bot warnings. Twitter is also authwalled, simply because they don't want people to train on it for free.
Since the operator uses user's computer via screenshots and keyboard/mouse, it's not easy to block by IP, browser fingerprinting, or convoluted page markup. The initial implementation probably has some detectable traits in its mouse and keyboard use, but a company that could train pretty convincing text and voice models can easily train a model to move mouse naturally enough. Simple screenshots can be messed with via flicker (using persistence of vision), but that's fixable too.
This cat and mouse game can end up ugly.
My pessimistic prediction is that Google et al. will extend their DRM from video to entire Web pages, and sell it as a service that blocks "unauthorized" screenshots and soft mouse, along with a cryptographically-strong way to detect the closed-source Chrome. It won't be accessible, but conveniently Republicans have added accessibility to their list of woke things to destroy. Sites will submit to this and advertise it as "use Chrome to skip those horrible CAPTCHAs!"
How much longer will OpenAI/Anthropic/Google be able to offer usable free tiers of their agents? Because imo this "death of the open web" is predicated on that.
Furthermore, Im skeptical we'll have the generous free tiers we have today for much longer. OAI and friends can only afford to burn millions on GPU cycles for non-paying customers for so long, and so long as we can continue to do web search for free I dont really see agents being a threat.
Why do people use chatGPT in favor of open source models? The performance isn’t that different for general queries, and not even that much for specialized ones.
130 comments
[ 2.9 ms ] story [ 168 ms ] threadImagine having a filter list enabled in ublock origin that was controlled by amazon or google or Microsoft. And oh wait, you can't even know what gets filtered.
Remember that people generally consume more than they produce on the internet. For certain demographics, that ratio is very extreme.
What? People understand the CLI. Many just don't prefer to use computers that way when alternatives like graphical user interfaces exist. If anything it will be like GUIs to CLIs, depending on where the UX settles.
There are people imagining interactive generative search with generative interfaces, etc.
In my cynical opinion, Operator is probably a push in the direction of getting reinforcement learning out of current user interfaces. It is all about getting developers and other users to dig their way out of employment for temporary gains.
I can't imagine companies will continue to just give up data to OpenAI. And I mean, most large companies with valuable data have begun selling it to OpenAI. Nobody is giving away anything for free at this point.
And besides, from what I've heard about it on here and elsewhere, Google has already made AdSense an unsustainable income source for websites, with all their requirements for suitable ad placement.
Assuming it actually works and doesn’t hallucinate nonstop, imo it’s going to kill off the “humans browse a website and order something” in the same way e-commerce killed off catalog and phone ordering systems.
Maybe check what Turnstile is/does before commenting? Cloudflare only needs you to solve captchas periodically and then they use your browser fingerprint to classify you as a human accessing a site.
Towards the rest of your comment:
I don't understand how this is a bad thing?
As far as I know, TripAdvisor gets a commission on all sales, so do most e-commerce providers as that's how e-commerce works?
I don't understand your argument.
Exactly how e-commerce killed off so many holdouts that insisted on in-store/catalog/phone purchases when the internet was growing.
I can see it driving purchasing decisions, but I can't see it killing off the actual shop sites? Because the shop is getting a _comission_ on the sale.
I'm getting an ever growing vibe off HN that for a website where a ton of people work on tech, or even in e-commerce serving content to people, they seem to have an awful understanding of consumer behavior.
If your site doesn't support/allow agentic bots, it will be similar to not accepting Visa or Mastercard. Yes, cash-only business is a thing. But it's suicide to make that your only payment option for most retail businesses.
If you assume that every good or service can be infinitely commoditized, then yes, perhaps this will hurt businesses that use this kind of technology.
Realistically though this just isn't the case and there are a lot of products, commodities and services that are unique in their own right and dont't have that kind of competition. So your Visa and Mastercard example is completely moot.
For example, you wouldn't really tell this agent to get you "the best rated shooter game on steam" because that's just not how purchasing decisions in gaming are made.
And on that point, from reading their announcement this sounds more like what critics of AI have been talking about, are you seriously going to automate something that gives you _pleasure_? Like selecting a game that you want to play? Wtf? Because in the scenario youre describing, thats exactly whats happening.
Illusion of choice. Think about how Netflix, Disney+, Hulu, iOS Appstore, Google Play Store, and even your example -- Steam -- provide recommendation to users. The overwhelming majority of people are already letting an algorithm make decisions for them. Acting like this is where people will suddenly draw the line is amusing.
"Play a new action movie that I'll enjoy" -- "Taken 4 now playing..."
Something about all this feels a bit backwards though, ultimately all these websites have some underlying API they're hitting. The "right" thing would be to just expose that API directly, and then an LLM could convert user input into a request that hits the endpoint. I think that was supposed to be the vision of Siri and such, but it never materialized because even if 3rd party integration was feasible natural understanding wasn't cracked yet.
Now that we finally have good natural language understanding, instead of executing on that vision, we're going backwards and having agents navigate websites as humans do?? If it does take off, the natural end-point is going to be that pages might have an alternate version that's simplified to make it easier for such agents to do their thing, just as web pages adapted for mobile formats.
100%. You absolutely get it, and also the implications, and also the outcome.
Humans will more and more rely on their AI assistants to take care of mundane things for them. If your site is difficult for an AI assistant to use (due to anti-bot behavior, or an overly complex flow that confuses LLMs, or whatever) your business will suffer. It's gonna be glorious if every retailer starts publishing a swagger API to facilitate automated transactions.
For example in the TripAdvisor example, would you really trust that OpenAI has gotten you the best answer? Personally I wouldn’t.
I feel sometimes people grossly overestimate how valuable the web really is. Or at least the majority of it. Wikipedia, archive.org, or my guilty pleasure, YouTube — only a handful of sites really stand out for me. If the rest crashed and burned I'd probably bicycle more.
While the author says that they are bypassing Google, that's not most people, and Google's results are front-loaded with AI answers, so Google results are already giving you specific answers, hallucinations and all. Not sure why average users would long-term switch to not-Google if Google can give them the amount of AI assistance that want or don't want.
I.e. instead of feeding my reviews to the rent seeking giants, I can just publish them on my own and they’ll be helpfully ingested and made useful by chattyG.
The dis-aggregation of the aggregators?
enenenshittification.
No one. And I mean NO ONE is going to delegate anything like travel or purchasing items to an AI. They just get it wrong too much.
When it comes to money, or really any kind of action that has consequences, you want the assurance of a deterministic outcome. That is simply not possible with this technology. Maybe it will be but I’ve seen nothing to indicate that so far.
OpenAI will subsidize banks to offer that safety-net for users of Operator. If Chase says you're 100% not on the hook for purchase mistakes made by OpenAI's Operator -- are you going to take the risk (as you pointed out, it's financial, it matters to people) with Perplexity Pro or Siri?
Just like with a real assistant, one could set very clear boundaries as to how much the assistant could spend, on what, etc., or even specify that more money can be spent in, say, a particular period and with a few extra passes (using different models?) just to make sure.
It does really feel like the same animal, because to the degree that I remember the discussions, a lot of it was about a perception of some kind of dichotomy (control or not, being able to 'touch' the product or not, etc.) that really doesn't need to exist.
There are cases where this could be true. I could see, say, asking it to do some research for you or something. Or maybe grab some restaurant recommendations. Really anything that brings you options that you can then make a decision on. If I do end up using this tech, thats how I would feel comfortable using it.
When it comes to things like making purchases and such on your behalf, thats where I disagree. Even with boundaries it's just not deterministic enough for most peoples risk appetite IMO. And when I say "most people" I am not talking about HN folks.
And just to be clear, I do use this technology. I just am very aware that its not this magic bullet that the AI bros want us to believe it is. I used it this morning in fact to help me write up some code I felt too lazy to deal with. It very quickly and efficiently wrote what I asked. I was pleasantly surprised and _almost_ missed the very bad bug that it dropped right into the middle of it all. I'm not giving my credit card to that lol.
The fact that such an AI, with sufficient development, even given what we have now, could present everything on a silver platter up to the execution of the /exact/ commands, and this alone is incredibly useful and can save a lot of time/expense/effort.
The crucial bit is that with, say, a human personal assistant, even if they say they will do exactly what you tell them to do, it's still ultimately inexact. But with an AI, it's trivial to 'decorate' a structured 'command object' (as json, xml, whatever) and from that moment let the deterministic system take over.
If anything, we sometimes want an actual human to /not/ deterministically execute /exactly/ what we tell them to, because perhaps they might know we were angry, drunk, or they just heard something on the news that should invalidate our request.
Either way, my point is that this distinction between 'doing research' and 'doing a thing' is not so dichotomous. In practice, I suspect all but the most autonomously-minded people, and especially non-IT folk, given enough sense of control over the confirmation of potentially 'fuzzy' actions, are happy with an AI doing a ton of stuff for them.
Just because they work that doesn't make them a good experience. Show me one dating app with good reviews.
(That seems like a lot of trust to me!)
Cursor already reads my private keys and writes code that goes straight to prod. I've stopped validating GPT 4o output for data analysis. Sure, things go wrong from time to time, but the convenience is unmatched.
That's a bold statement. There are people that get in the backseat or passenger seat while their FSD is driving. There are plenty of other examples of how humans do things because they are reckless or ignorant or any other adjective you want to use
I’ve just had an interaction with an insurer who used an ML based system to judge insurability (risk) and make recommendations (requirements) to me for immediate action or policy termination. Fascinating, and totally doable with today’s technology, but it makes me very uneasy.
Anyway, it’s cheaper than hiring an inspector and it gave me some batshit recommendations- including removal of invasive plant material (vines on a trellis, looks great in July) and removing extensive roof debris (leaves, we live in a forest). My recourse is an extremely annoying call center hell, or get on the roof. Guess what I did. And I’m not getting a discount.
To delegate our money/critical-wise tasks to an AI (and other things people dream of), first cientists must find a way different technology we have today.
Spoiler: If it ever happens, it's not something we, normal people, will have the pleasure to be fiddling with firsthand. The country capable of discover and develop it first will dominate everything.
But that assumption is precisely a goal of OpenAI and Anthropic, and it's useful to continue that train of thought to see the bigger picture of where things will go if they get what they want.
The article did a very astute job of that.
>> Takeover mode: Operator asks the user to take over when inputting sensitive information into the browser, such as login credentials or payment information. When in takeover mode, Operator does not collect or screenshot information entered by the user.
Reviews are generally pretty useless unless you know something about the reviewer in order to qualify their opinion. Magaluf is a highly rated tourist destination among some people in society, but I sure as hell don’t want to go there.
AI can’t just aggregate all the opinions and say “Ok specific individual, I’ve checked the internet and here’s what the internet has decided.”
To me it makes more sense to talk about what may well be coming than to, I don't know, wait to be run over by it.
The facts and claims in this article too will have changed, come to pass, or be realized.
The author speculating that AI will solve interacting with user interfaces is something that has yet to be seen.
They are making a gigantic leap from (not verbatim quotes, just making a point) "I saw one example of Operator doing (what Claude computer use did back in October, and other projects before that)" to "I think advertising is blown" and "I think search is doomed" to "Sadly I think the open web is about to die.".
Also, people have been talking about this for a few years now. Is there any evidence of it happening yet?
Banks are an example of middlemen who continue to exist because they provide a compelling service (everyone needs to store and move monies). Travel companies are an example of middlemen who died because they couldn't provide a compelling service (everyone can do their own bookings now directly with airlines, hotels, et al.).
If "AI" puts search engines like Google Search and data aggregators like Yelp out of business, then that's because their services are not compelling. Good riddance. Middlemen continuing to exist without providing compelling services are just collecting free rent, nobody should have to pay rent for nothing.
As regards the "open" internet, that died a long time ago when most use cases for the internet centralized on a handful of websites and services.
"AI" doesn't change the fact most group conversations happen on Reddit now, that most direct communications happen over Discord or Teams (or email) now, that most traffic gets routed through Cloudflare now, that most companies contract with Amazon Web Services for servers now, that everyone speaks Chrome now, and so on.
This is a common sentiment. I have been using chatgpt for various things, and there are times when I will turn to it for technical help, sysadmin or figuring out some shell / cli commands. However by and large I definitely still prefer googling something. I would much rather read a wikipedia article of the French Revolution than trust whatever chatgpt had to say about it. I figure chatgpt is just spitting out the very same wikipedia article with some rewording and the occasional hallucination and I would rather just read straight from the source. Same with programming questions - between reading a stack overflow thread with the answer and chatgpt I feel a lot better reading stackoverflow, seeing the context the replies, the alternatives etc. and deciding based on that rather than just trusting gpt.
Whatever particular aspect of the summary intrigues me is what I'll then go look at the original source for.
I was telling both my son and my father about how chatgpt is not a search engine, and they started testing things and .... well the obvious happened, they easily found mistakes. And they stopped using it so much. They had no idea it wasn't kind of a search engine.
I can ask it questions and get amazing answers, I see the use of it. But it's not a search engine replacement, you have to double check things with search engines. A search engine might return a result with a bogus webpage of course.
People don't know that it's not a search engine! It's much more, but it's also a bit less ;-)
I don't like the LLM interface that it's all filtered to this super friendly, grammatically correct reply that might however make no actual sense logically, or be plain wrong. With websites I see tons of signals I can use to determine how much value I should put to some information, I can compare different websites, etc.
Also, it's super slow to return a response. Maybe I'm old, but for most of the tasks I use search engines for, I wouldn't use LLM's.
Meanwhile, Duck Duck Go provides reasonable sources.
After rather extensive testing, my heuristic is never to ask ChatGPT for facts about anything that can’t be readily verified— and even then don’t trust it until it is verified. The act of verification makes ChatGPT too time consuming to be worth using for education.
It’s useful when I’m trying to solve certain programming problems, though.
I cannot imagine using it as a primary source for literally anything, ever.
If I just want to learn conceptual knowledge, having a conversation with ChatGPT isn't much different from doing a DFS through hyperlinks on Wikipedia. In fact, using ChatGPT can be more efficient.
If I want to critically understand the details, Wikipedia, aside from providing references, doesn’t offer much more help than ChatGPT. More importantly, just as LLMs can exhibit hallucinations, community content can also contain biases or even errors.
in the end, it still comes down to relying on oneself.
Eg I just asked chatgpt about the metric expansion of space, and its answer is the common popsci answer, that's incorrect
Next, I asked it what the general equation for general relativistic redshift in an arbitrary spacetime is. It answered
1+z = a(trec) / a(temit)
Where it lists a(t) as the scale factor of the universe. This is not what I asked it, and is the incorrect equation
Next I asked it if the adm formalism is strictly equivalent to general relativity. It said that it is strictly equivalent. This is incorrect, as ADM implies a restriction on the change of the topology of spacetime. A lot of the rest of its explanation of ADM is incorrect
Next I asked it if the ADM formalism is covariant, which it is. ChatGPT says no, because of the foliation of spacetime. I asked it if covariant BSSN is covariant, and it said yes, while it listed reasons why BSSN isn't covariant (which it isn't) that had nothing to do with why it isn't covariant
Then I asked chatgpt if the ADM formalism is more stable the cBSSN formalism. It said that ADM is more stable for highly dynamical systems like binary black hole mergers, which is factually incorrect - it cannot be used for binary black hole mergers. It attributes this to the absence of a clear slicing structure - which is incorrect as they both use the same gauge conditions
I then asked it if the Z4 formalism is generally covariant. It said that it, and adm (which it earlier said was not covariant), are both generally covariant. Now on the plus side its right about z4, but my goodness
Then I asked it if the ADM formalism is hyperbolic. It said no, because its second order in time and space, but if it were first order in time and space, it would be. This is incorrect. ADM is weakly hyperbolic, but strongly hyperbolic systems can be first order in time, and second order in space. It also states:
>This involves introducing auxiliary variables (such as the conjugate momenta of the metric components)
The conjugate momenta is one of the fundamental variables of ADM, so its not an auxiliary variable
--fin--
Literally everything I asked chatgpt here gave a critically wrong answer, from the basic to the complex, and I'm super glossing over the walls of text which are significantly wrong in virtually all the details. Its frustrating because its not saying its not sure, its just confidently incorrect
Information on wikipedia is generally much more accurate, and much more verifiable because it provides links and sources for you to check yourself. Don't use chatgpt for anything
For every piece of content like this generated by an LLM, the first thing I'll do is validate it.
By "conceptual knowledge", I mean something like "why the moon keeps one side towards the Earth" and "what is tidal locking."
I can often find the sources its mashing together a lot of the time. You can see its repeating 50% of a sentence here, and combining it together with 50% of a sentence there, to make something that's 100% wrong because its not a tool that understands context
For me it feels actively like the opposite of learning. With a human author, there's a chain of reasoning to bring you to a logical conclusion. Even if that conclusion is wrong, there's value in trying to understand what their reasoning is. ChatGPT finds it impossible to author a structured argument, and its logic is a mess of nonsequiturs that there's no value in trying to understand. It makes it a very poor tool for information gathering imo
> Okay, the user is asking if the ADM formalism is strictly equivalent to General Relativity. Let me start by recalling what the ADM formalism is. ADM stands for Arnowitt, Deser, and Misner, who developed a way to reformulate Einstein's equations into a Hamiltonian framework. This is useful for things like numerical relativity and quantum gravity.
So, first, I need to confirm whether ADM is just a reformulation of GR or if it changes the theory. From what I remember, ADM splits spacetime into space and time, using a 3+1 decomposition. They introduce variables like the lapse function and shift vector to handle the foliation of spacetime into spatial hypersurfaces.
Wait, does this decomposition change the physics? Or is it just a different mathematical approach? I think it's the latter. The Einstein equations in ADM form should be equivalent to the original ones, just expressed in terms of different variables. But I should check if there are any subtleties, like constraints or boundary conditions that might affect equivalence.
Also, the user might be confused because sometimes when you change the formalism, especially in constrained systems, there can be issues. For example, in ADM, you have the Hamiltonian and momentum constraints. These need to be satisfied at all times, and numerical implementations have to ensure that. But theoretically, as long as the constraints are satisfied, the ADM equations should be equivalent.
Another point is the initial value problem. ADM is used for that, which suggests it's a way to set up initial data consistent with GR. If the equivalence wasn't strict, then initial data in ADM wouldn't correspond properly to GR solutions. But I think they do, as long as the constraints are met.
However, I should mention that while the formalism is equivalent, practical implementations (like in numerical relativity) might face issues like constraint violation over time, which aren't a problem with the theory itself but with the method of solving the equations.
Also, maybe the user is coming from a background where they heard about different formulations of GR, like tetrad formalism or others, and is wondering if ADM is another such equivalent approach. So, confirming that ADM is a Hamiltonian formulation that doesn't alter the physical content of GR, just the mathematical structure.
Need to make sure to explain that equivalence here means the solutions to the equations are the same, and any differences are in the mathematical handling, not the physical predictions. Also, perhaps touch upon the applications where ADM is useful, to highlight why this reformulation is valuable even though it's equivalent.
I should also check if there's any gauge freedom or choice in the ADM formalism that might lead to different interpretations. For example, the lapse and shift are gauge variables, but that's part of the coordinate freedom already present in GR. So, choosing different lapses and shifts corresponds to different coordinate choices in standard GR, which doesn't affect the physical equivalence.
In summary, the ADM formalism is a reformulation that breaks down Einstein's equations into a set of constrained evolutionary equations. The physical content remains the same as GR, so they are strictly equivalent in terms of physical predictions, but the mathematical approach is different, suited for certain types of problems like numerical simulations or canonical quantization.
---
It's far far beyond my capability to validate if it's right, but I think it would be helpful if the LLM can explain how it thinks?
I agree LLM often being confidently incorrect, it often gives me code that completely un-compilable, users need to review its content very carefully, and we need to do that for content from other sources also
What I want to explain is that LLMs are helpful at many scenarios(not as hardcore as astrophysics, for example, there...
>Wait, does this decomposition change the physics? Or is it just a different mathematical approach? I think it's the latter
This for example is faulty reasoning, because its not an either or situation
>I should check if there are any subtleties, like constraints or boundary conditions that might affect equivalence
This is also faulty reasoning, because it assumes that non equivalence must mean that they produce difference physics
>Another point is the initial value problem. ADM is used for that, which suggests it's a way to set up initial data consistent with GR. If the equivalence wasn't strict, then initial data in ADM wouldn't correspond properly to GR solutions. But I think they do, as long as the constraints are met.
Notice how its trying to sneak a fast one past you. Its saying that because ADM with valid initial conditions maps to GR, therefore GR maps to ADM - which is a logical fallacy! It makes this mistake several times in its chain of reasoning
The issue is its not doing any real reasoning, and isn't capable of actually making deductions. Its about as productive as just.. throwing all the pieces of information on the floor, and picking them up in a random order
This isn't really getting into the weeds with astrophysics, its ADM 101, and deepseek is just.. lying right to your face with confidence, presented under the guise of a chain of reasoning
I do translations of long podcasts from time to time and it's never accurate and often produces hallucination(I know enough in both languages just using LLM to speed up
Something like Operator may well be the death of the ad-funded web—as TFA observes, who is going to see the ads? But the reason why a tool like Operator has why sort of appeal is because advertising already killed the open web and replaced it with a lifeless corpse.
Paying for and checking out is 5% of the customer journey here.
And choosing something to do should be both fun and something you'd want to do yourself—it only isn't fun because businesses aren't incentivized to make it fun, they're incentivized to try to redirect your attention away from what you actually want and instead draw you to their various sponsors.
Delegation sounds great and sometimes is, but be careful what you wish for, especially when the your delegate is actually working for somebody else, as will be the case for most LLM agents and was the case for most travel agents.
Ad companies could offer OpenAI lots of money to mention them, sidestepping the likes Google et al
That may sound pessimistic, bit honestly I have zero confidence that in Trumps's USA, OpenAI will have to give a flying fuck about users' best interests or honest advertising or any of that stuff...
Meaning advertiser's will have to be selective which company they pay to get the most exposure to their target human customers via the agents. Will we see affiliate programs for AI agents which in turn promote products to the users? and we end up with the same shit show we have today?
Or what if eventually everyone has their own personal AI that can bypass the ads. Will we just decide that advertising is a drag on society and kill off that industry for good?
Absolutely certain. Furthermore, this can maximize exploitation of each individual human as the providers have such rich profiles about each human that they can customize pricing to extract the maximum amount of profit. It is by design.
People have spent lots of words talking about the how the saturation of generative AI output will spoil future trainig, which is true, but is most immediately a problem for the AI model developers themselves.
But even with them only acting as consumers, they break almost of the economic and social bases for how the internet works.
Advertisers want their ads to be seen by humans, experts/pontificators want their words to to be seen by humans, artists want their images to be seem by humans.
These things provide the foundations of the internet as we know it.
If online tools are being used by bots who are unaffected bu ads, and texts are summarized by bots instead of read directly, and images are synthesized dynamically instead of appreciated directly, the internet itself breaks.
As the author notes, the next most likely equilibrium involves an even more commercial, more expensive, more gated, more impersonal future than we might have already seen ourselves careening towards just a few years ago.
Enjoy what you've had!
How does anyone figure out if something is worthwhile, if they haven't used it? They have heuristics, one of which is comfort and familiarity. That's why advertising exists: to get a brand in front of your face, so you recognize it.
Maybe the bots end up with a similar heuristic as they crawl their sources. Perhaps the advertisers will be perfectly happy to pay, not for human eyeballs but for the crawlers.
This is, of course, absurd. But since I find the capabilities of the bots opaque, why not absurdity.
Also it's optimised for todays web for websites and forms and I suspect they continue develop it this way for two reasons:
1. This works TODAY, which is important for a business.
2. As others commented in this thread, people won't necessarily trust it blindly, seeing visually what it does exactly is a feature.
Why publish it to AI companies? The Internet is ripe for mass syndication of textual content and putting it behind subscriptions, just as YouTube has done with video content. Syndication was the distribution and business model for this content before the WWW, the advantage we can get now is no barrier to entry and no distribution costs. (Hosting costs are so low to not even count for text).
At the same time we also get rid of the ad-ridden, scam-infest, bot-crawling Internet of the current day.
Said I, on an open forum where nobody pays a dime...
Have client certificates worked in any niche? We now have SSO by Google/Apple/Facebook/X, but those identities are also used by bots.
I'm pretty sure Wikipedia will not do any of those things, as it's a project ran by volunteers, and they do want to disseminate their content as widely as possible. The revenue from the "please donate" feature is only used for website development and to pay the servers.
For the other websites, yeah, those might change and lock things down, require accounts to browse, etc, to prevent being scraped. Already now when I look for house listings or onto linkedin, I see all sorts of bot warnings. Twitter is also authwalled, simply because they don't want people to train on it for free.
This cat and mouse game can end up ugly.
My pessimistic prediction is that Google et al. will extend their DRM from video to entire Web pages, and sell it as a service that blocks "unauthorized" screenshots and soft mouse, along with a cryptographically-strong way to detect the closed-source Chrome. It won't be accessible, but conveniently Republicans have added accessibility to their list of woke things to destroy. Sites will submit to this and advertise it as "use Chrome to skip those horrible CAPTCHAs!"
Furthermore, Im skeptical we'll have the generous free tiers we have today for much longer. OAI and friends can only afford to burn millions on GPU cycles for non-paying customers for so long, and so long as we can continue to do web search for free I dont really see agents being a threat.