OpenAI Announces SearchGPT
chatgpt.com
chatgpt.com
[0] https://www.ifixit.com/Guide/Data+Connector++Replacement/147...
[1] https://chatgpt.com/share/e52dc4dd-77e6-48a5-a7ca-77e3dfa39e...
Not really, from the example you provided I think it's pretty clear that a custom system prompt was used and ChatGPT is using its own creativity, it doesn't have anything in common with the iFixit guide.
Yeah, and it was suddenly made just the day before Google announced their models... and do you remember Sora? That was in February.
They know how to release to get press coverage.
Maybe it's released to paid users only? The only thing missing from the original announcement is voice interrupting the ongoing response (well and the Scarlett Johansson voice)
Don't get me wrong, search has become extremely problematic, ... but how much effort does it take really? Compared to writing a letter, reading a map, walking half a mile etc?
The google results for "music festivals in boone north carolina august" are completely adequate. If not then you search again, sure. What makes that "a lot of effort" compared to "asking follow-up questions like you would in a conversation"?
According to Sequoia Capitol there is a $600B hole in this sector at the moment, which continues to grow.[1] They need to invent something akin to the global smartphone market, over the next 2-4 years. Some thing new that solves significant problems for people that aren't already solved for free.
Whether this gets AI out of being a money pit remains to be seen.
I feel like Google used to be pretty great at answering questions until they started pushing ads more aggressively into their results, and then intentionally gimping their results (for instance if you put quotes around a word Google will happily ignore it, whereas this used to actually be an enforced rule).
I'm not sure AI is the solution to this problem at all. I think it's about misaligned incentives and trying to shoehorn a new cash cow into the mix.
Misaligned incentives will of course make either side worse but search engines ~2010 were (and now still are) far from being good solutions for all types of queries, they were just closer to being good than anything else we had for them.
Depending on the query, substantial effort.
I know I've taken to just asking Bing (read: ChatGPT) if I just want something answered (FSVO answered) because sometimes I just don't feel like speaking Search Engine across multiple attempts to try and find something.
Many days or weeks infact
The provided examples don't illustrate the problem properly. Let's say you're new to a problem domain but don't know it's a highly specialized domain. So you don't know the keywords to search but have a vague sense of what needs to be accomplished.
If the new tool can make sense of your vague problem and point you in right direction, it would be great.
Don't know if this is actually a $600B sized problem.
As far as I can see there's no sending people away from SearchGPT, it just gives answers. I can't see any reason to allow AI crawlers on my sites, all they do is crawl my site and slow things down. I'm glad that most of them seem to respect robots.txt.
My impression from the demo is it has a perplexity-like result, with the answer and references to where each part comes from.
All of that disappears when a bot devours whatever it assesses to be your "content" and then serves it up as a QA response, stripped of any of the surrounding context.
You know why people like these chat systems? Because it straight up saves time. When a system is made it to indexable, "context dependent", and "creating a certain experience" it just begs to be summarized and made to be something you can use. That interpretable work is.... Pointlessly difficult.
A good example: discord. A vast number of communities are designed to be "experiences" where you have to pour hours of your time to adapt to their little fiefdoms if you wanted to obtain any useful information in the form of important information on a topic. Try doing this in any serious fashion and you will quickly be wasting more of your time than you want.
Yeah so maybe chatgpt gives you the occasional incorrect fact. I haven't had that happen in any way, shape, or form. Furthermore: just be critical of your information. Not hard, and they are already working on fixing that.
Especially for people that are bonafide adults time is worth more than "the pride of human work".
Why are you not just asking chatgpt "what's the latest tech news"?
Could it be that there's something else you get from this site other than just it's content being easily searchable in someone elses database?
Note however that HN is not gatekeeping any useful information that may be produced during conversations here; in fact, it's all indexed and searchable.
I'm not normally a purist on these things, but I'm recalling musical artists who bemoaned the destruction of the album format in favour of $0.99/track sales in the early days of the iTunes store. Concept albums in the vein of Sgt Peppers still exist of course, but almost every modern mainstream song is now prepared first and foremost to be listened to in isolation. I didn't care for those arguments at the time they were being made, but years later, I can appreciate how something was lost there and that it might have been appropriate to let artists specify that album X was to be sold only as an album.
I've already received emails from SEO snake oil sellers now advertising themselves as being able to influence ChatGPT output.
https://github.com/ai-robots-txt/ai.robots.txt/blob/main/tab...
Some of them identify themselves by user agent but don't respect robots.txt, so you have to set up your server to 403 their requests to keep them out. If they start obfuscating their user agents then there won't be an easy solution besides deferring to a platform like CloudFlare which offers to play that cat and mouse game on your behalf though.
https://developers.google.com/search/docs/crawling-indexing/...
They also have a feature that will, if a user pastes a URL into their chat, go fetch the data and do something with it in response to the user's query. This is the feature that made a big kerfuffle on HN a while back when someone noticed it [0].
That second feature is not a web crawler in any meaningful sense of the word "crawler". It looks up exactly one URL that the user asked for and does something with it. It's Perplexity acting as a User Agent in the original sense of the word: a user's agent for accessing and manipulating data on the open web.
If an AI agent manipulating a web page that I ask it to manipulate in the way I ask it to manipulate it is considered abusive then so are ad blockers, reader mode, screen readers, dark reader, and anything else that gives me access to open web content in a form that the author didn't originally intend.
The action is indeed prompted by a human, but so is any crawl in some way. At some point they either configured an interval or other trigger to send the script to the Web host to fetch anything it can find.
It's inherently different to extensions such as adblockers that just remove elements according to configuration.
After all, the users device will never even see the final DOM now. instead it's getting fetched, parsed and processed on a third device, which is objectively a robot. You'd be able to make that argument only if it was implemented via an extension (users device fetched the page and posts the final document to the LLM for processing).
And that's ignoring the fact that adblockers are seen as illegitimate by a lot of websites too, and they often try to block access to people using these extensions too.
> After all, the users device will never even see the final DOM now. instead it's getting fetched, parsed and processed on a third device, which is objectively a robot.
Sure, but why does it matter if the machine that I ask to fetch, parse, and process the DOM lives on my computer or on someone else's? I, the human being, will never see the DOM either way.
This distinction between my computer and a third-party computer quickly falls apart when you push at it.
If I issue a curl request from a server that I'm renting, is that a robot request? What about if I'm using Firefox on a remote desktop? What about if I self-host a client like Perplexity on a local server?
We live in an era where many developers run their IDE backend in the cloud. The line between "my device" and "cloud device" has been nearly entirely blurred, so making that the line between "robot" and "not robot" is entirely irrational in 2024.
The only definition of "robot" or "crawler" that makes any kind of sense is the one provided by robotstxt.org [0], and it's one that unequivocally would incorporate Perplexity on the "not robot" side:
> A robot is a program that automatically traverses the Web's hypertext structure by retrieving a document, and recursively retrieving all documents that are referenced. ... Normal Web browsers are not robots, because they are operated by a human, and don't automatically retrieve referenced documents (other than inline images).
Or the MDN definition [1]:
> A web crawler is a program, often called a bot or robot, which systematically browses the Web to collect data from webpages. Typically search engines (e.g. Google, Bing, etc.) use crawlers to build indexes.
Perplexity issues one web request per human interaction and does not fetch referenced pages. It cannot be considered a "crawler" by either of these definitions, and the definition you've come up with just doesn't work in the era of cloud software.
[0] https://www.robotstxt.org/faq/what.html
[1] https://developer.mozilla.org/en-US/docs/Glossary/Crawler
It's triggering an automation which fetches data. This is a crawl, even if the crawl has a very limited scope (it's also not limited to a single request, that's just the scope that's used by default. But even if it was programmatically limited to only ever request a single resource, that'd still be a crawl, while recursion is the norm too build indexes, it's not necessary for all usecases that utilize crawlers.
Did you ever actually make anything that's utilizing them to gather information that you want? You might be surprised to know that adhoc triggering a singular resource fetch is actually pretty common to keep data up-to-date.
> If I issue a curl request from a server that I'm renting, is that a robot request? What about if I'm using Firefox on a remote desktop? What about if I self-host a client like Perplexity on a local server?
Yes, anything on a third device is effectively a robot that's acting on the behalf of the acteur.
So, instead of giving 30% fee for DoorDash or Wolt about visibility, we start giving that fee for "some-AI-search-tool", and they don't allow selling food cheaper than you get by using the AI search process ordering. I don't like this era.
Only some big, well-known restaurants are an exception because people already known them and know to look for them. They are big enough to have online shop for food delivery and some form of delivery system.
For other places, it is not like that. They need to be visible on platforms that people use. If they are not there, people don't order take-away food because they mainly use applications which provide decent-enough catalog about available restaurants with delivery option and easy payment process.
Typical platform monopoly problem exists because people tend to be lazy and it is hard to get visibility on traditional search engines, when they are flooded by ads for the normal users, and then if you filter them out, SSO spam, which might not be related to restaurants, and then finally, competition against other restaurants.
To an extent, I actually like the trend as Joe Average User.
Most websites are just plain filthy and even dangerous today. I know I am not opening any link to a website I don't already know and trust unless it's in a Private window (fuck their cookies) with JavaShit more than likely blocked. If it's really shady I'll fire up an entire disposable VM for it first.
Google, Bing, et al. just putting the content right then and there saves me time and hassle from dealing with the ancillary garbage myself.
It's honestly a tragedy of the commons. Big Tech wants more traffic and to keep it, websites want more traffic and just throw whatever literal shit they can muster (aka SEO).
So what we really need here is a new approach to funding the availability of information. Unfortunately, ads are fairly lucrative because advertisers are willing to pay a lot more than users are. You could I guess do something where SearchGPT pays a couple of cents out of a monthly fee to each information source it used. Much harder with LLMs, since the source of information is potentially very diffuse and difficult to track. And even if you tracked it each publisher would get such a tiny fraction of what they are making now.
But the difficult part for web publishers is that AI powered information retrieval is a significantly better user experience, which means it's very likely to win no matter what.
For a long time now, that reason was to show ads and the content quality was very, very low. It destroyed journalism, it destroyed everything pretty much. Some of the blame is on Google but a lot of the blame is on the people with websites IMHO.
People with websites can go back to the good old days when they made websites to show off their talents, persuade people into activism, spread ideas and seek interaction with likeminded humans.
LLMs have the problem of hallucinations but when that's solved I wouldn't be looking back. Hopefully, Google itself would be disrupted.
Maybe we can finally have a business model for high quality data, like journalists selling their work on current events without the need to present it in the most bombastic way possible?
I think the world currently is in strong need of a way to process large amounts of data streaming in and make sense of it. AI can be very useful of that. I would love to have access to an LLM that can keep up with the current events so I don't have to endure all the nazi stuff on Twitter.
I don't want censorship it's just that I would be perfectly happy to know that there are bunch of people thinking that Biden was replaced and that there are some other people thinking that Michelle Obama is actually a man without have to read it like 100 times a day when I'm trying to look at the opinions of the people on something. It's cool to know that there are such people(or bots?) out there but I don't want to read their stuff, I want to the computer to on top of all that and give me a brief on, then I can drill down if I want to know more or see the exact content.
While people do search for things that could benefit from some comprehension, I don't think that's a common feature.
For example, my most recent searches and I'm probably a bit of an outlier given my usage:
"given when then" "[some project] github" "[some person] wikipedia" "[some person] wikipedia" "act of supremacy wikipedia" "NVDA stock" "django docs onetoonefield"
Perhaps one of those Wikipedia searches could have been done better as an AI search since I wanted to know something specific, but other two were just from wanting to read generally about a topic.
The benefits I get from ChatGPT and similar tools are more conversational than search like. Eg, I might be trying to solve a coding problem and want some suggestions about how I might go about it. I might act for libraries, example code, and pros and cons of different approaches. I basically use it as a replacement for another senior engineer which I can bounce ideas off of, it's not for search / knowledge type stuff and I can't see why I'd ask an AI for that. If I want to know something I can just type a few keywords into google and find a reputable site that for that info.
AI terrible until OpenAI released ChatGPT so I'm personally excited to see what ChatGPT can bring to search
> In ChatGPT's recent search engine announcement, they ask for "music festivals in Boone North Carolina in august"
> There are five results in the example image in the ChatGPT blog post :
> 1: Festival in Boone ... that ends July 27 ... ChatGPT's dates are when the box office is closed [X] 2: A festival in Swannanoa, two hours away from Boone, closer to Asheville [X] 3. Free Friday night summer concerts at a community center (not a festival but close enough) [O] 4. The website to a local venue [X] 5. A festival that takes place in June, although ChatGPT's summary notes this. [Shrug]
"What is the mean and standard deviation of the AQI along the current fastest driving route from Palo Alto to Lassen National Park, averaged over the driving time"
"What is the easternmost supermarket before Yosemite that is at least 2000sqft in size"
etc
Kinda hilarious that nobody remembers...
Open AI or other AI companies could capitalize on that because they already have hooked their users up with their LLMs, search could be another feature and has the potential to grow from there.
Maybe this is the difference between anthropic and open ai.. Anthropic has sharp focus on improving their core product and open ai is spread out.
Not this time. With this one, we were given notice about this as a rumor 3 months ago. [0]
It ties everything to their platform and returns a regurgitation of prioritized content without indicating any sort of sponsorship.
SEO will be replaced by cold hard cash, favors, and backroom deals
Maybe it's my pessimistic nature, but it's garbage either way to me - backroom deals in your scenario, or the SEO-gameified garbage we currently have.
Maybe this reflects my biases, but isn't that was SEO has been from the get go? Like, from the moment someone had the idea that they could influence search engine results in their favor and charge money for those services, SEO has been purely negative for internet users simply trying to find the most fitting results for their query.
For well over a decade the best SEO trick is to write helpful useful content.
Your small independent blog can become a top Google hit without too much effort. This is kind of neat.
I want more false positives.
https://openai.com/index/new-ai-classifier-for-indicating-ai...
I don’t know the current meta on LLM vs LLM detector, but if I had to pick one job or the other, I’d rather train a binary classifier to detect a giant randomized nucleus sampling decoder thing than fool a binary classifier with said Markov process thing.
Please don’t advocate for giving up on spam, that affects us all.
Well, if you don't care how many false positives it has, just block everything. But there's no even remotely reliable way to detect LLM output if it isn't deliberately watermarked to facilitate that, so you aren't going to get anything that is actually good at that.
The fact that SEO has to exist in the first place is evidence of search engine mafia.
I asked Perplexity Pro about the difference between the Rapier and Jolt physics engines. It missed many things that’s clearly available in the docs such as determinism and JS language bindings for Jolt, which makes me afraid to trust it.
Also asked about best Italian pizza places near my address to try something completely different. The top result doesn’t serve pizza and the 2nd result was many kilometers away.
- if they release a model “they’re just releasing models without use cases” - if they release safety guardrails “they are just doing this to avoid launching models” - if the release has a waitlist “they’re losing their velocity” - if they launch without a waitlist “they weren’t considering the safety implications” - if they hired a top researcher “they’re conspiring to out spend open source” - if they fire a top researcher “there’s too much politics taking over”
New models are coming fast too.
new = better, new use cases
To what extent 4o is a new model or a refinement depends on:
a) technology
b) thresholds for what it means for a model to be "new"
Not naming.
We have no clue about what happens within the super-secretive ironically-named OpenAI. To me, it feels like a new model. To you, it feels like a refinement. Unless one of us has insider information, I'm not sure it's worth disputing. We have a difference of opinion, and likely, neither of us has anything to back it up.
Based on nothing but idle speculation.
I like the follow up questions feature but how is it different than chat gpt - just providing links as well?
Meta is out to commodotize Google's core product (Search). Meta's foray into Open Source AI is likely to hurt more as Google's distribution advantages (via Chrome and Android) is close to insurmountable for one rival search engine to make any meaningful dent by going toe-to-toe.
And a lot of what those on-model devices can do are what the average person will want.
And there's going to take a lot to move people meaningfully away from google.
Kahm, Kagi (and Google) would both disagree. You can even pick your favorite business model!
A few (inlcuding Kagi's LLM assisted search) will be monetized through user/customer subscriptions exclusively.
As with search, these two business models will lead to different outcomes for the users.
Having it part of the OpenAI subscription, and assuming it will be accurate, I think Perplexity will have a problem.
Let's see how long the waiting will last, still waiting for Sora access!
Degrading the quality of the output of tools that were once considered reliable and safe is going to create problems, and the onus is not fully on the people using those tools, but on those creating them.
Anyway, comparing it by speed isn't useful at all.
Product comparison/updates/features:
> What's the difference in speedo latex cap vs elastomeric?
> What's new in iOS 18 Beta 4?
Documentation (esp for non-mainstream tools):
> How do i find the sioyek db files? How do I move it to my local iCloud folder instead?
I'd assume SearchGPT's results might be better given the partnerships with publishers and creators vs Perplexity searching the internet. More importantly, Perplexity already did the work of finding Product-Market Fit for OpenAI.
People keep telling me that I can use the smaller models, but I really can't. I'm using this for work and those things are toys which just game bench marks.
I'd love to give them an api key from openai or anthropic and get uses to my hearts content like phind does.
They use google and other search providers to run the query over the results and may be they still can find a good provider. However it's either Google or Bing and they both have their own competing products.
However, openAI might not execute this better and then Perplexity might have a chance... (I hope so).
No they don’t? AFAIK they have their own crawler and semantic search index.
As much as I love Anthropic (the only AI company I pay), they seem to be investing almost nothing in pure product orgs. No way to manage chats, memory usage is very high on the web ui (700mb claude vs. 100mb chatgpt on my mac), artifacts can be very hit or miss. They've been hiring the brightest ML people, time to build some strong product orgs (I nominate myself, DM me anthropic people)
I understand that website quality has gone down with respect to ads, but this really doesn't make the internet very inspiring if all outputs come in this format.
At least with Google I get to visit someone's website, view their content in the way they have designed it and perhaps even take a look around.
Could just be my view given all I know its Search > Website > Content.
> Publishers will have a way to “manage how they appear in OpenAI search features,”
Oh, goody.
"Who can tell me how and why is Perplexity.ai is worth $1BN? How much revenue are they making vs the amount of money they are burning? What is the justification of this valuation?" [0]
At this point with this unsurprising announcement, Perplexity is worth <$50M.
For non-traded companies, it's even worse, because it's less people, at fewer moments, and they don't even need to be honest. As a first approximation, non-traded companies do not have a valuation, any number you get is bullshit and you can get something with the same accuracy by just asking ChatGPT.
I guess I'll withhold my cynicism until I try this. I kind of use chatgpt as a search engine anyway, so seems like a reasonable direction.
https://www.theatlantic.com/technology/archive/2024/07/searc...
It may not be this specific iteration that kills it, but a search product with AI (and without trash) is what will dethrone it.
I'm already using ChatGPT for ~30% of the queries I used to use Google for. I prefer hallucinations to ads, to be honest.
They were right to call Code Red when GPT came out, but their response to it has been extremely poor, even when they had all their cards in their hands. The quality of their products has been increasingly worse with time, everyone (but their own VPs) has been telling them that, it's hardly a secret.
They literally just have to go through the first two or three comments on this site (or Reddit or w/e) and fix the extremely obvious pain points people have with their products:
* Bring back verbatim search, make search *actually* work.
* If I search for "italian restaurants", I want a list of italian restaurants not a blog post with someone's opinion on why italian restaurants should hire more immigrants because of blah blah blah ... I want to *eat* something!
* The whole "vikings were black" episode ... wtf.
You still have to discard a lot of information from Google, you probably just got used to it. Even though I still use it for ~70% of my queries, what I'm actually looking for is one or two pages down the list of results, the first ones being just mediocre articles around the topic of interest.
What's the first thing you do when you get Google results? You scroll down, it has become muscle memory at this point.
1. google was a good search engine when it was less profitable
2. now that it is more profitable, it is bad
Importantly, it was possible for Google to be good AND profitable at the same time! Roughly from 2003-2013 perhaps.
1. OpenAI is nowhere near profitable ... it seems to be heavily dependent on Microsoft, and in some sense on Microsoft's desire to compete with Google in certain areas
2. If it ever becomes profitable, does anyone want to argue it won't get significantly worse? It will probably have a bunch of bad side effects, like Google's decline did on the web itself
I guess this is "normal", but it also seems pretty inefficient to me ... Part of the problem is that "free" is a special price that users like
IMO it would have be nice if Google search was sustainable at a high quality -- I think it easily could have been
Here's hoping capitalism starts working again with subscriptions so users are the consumers and not the product.
ChatGPT can be useful for certain hard to Google informational questions but doesn’t help me at all for the boring “IKEA hours” type searches I do every day.
Discussion on official post: https://news.ycombinator.com/item?id=41071585
ChatGPT just works, and it works quickly, and its usually right, and is a better user experience than Google Search in every way. I hope OpenAI comes out with an AI mail client so I can finally ditch Google completely.
Also google gives an AI generated answer at the top now, along with the sources so you can quickly check. I have caught a few bad answers like this.
This feels like the same thing. If I ask my assistant about "music festivals in Boone North Carolina in august" and they give me 5 results, 0 of which actually match what I asked for I'm throwing my hands up and never asking them for help again.
I use LLMs all the time, regularly throughout every single day, at work and home, and sometimes I really struggle to articulate what they're good at, but this doesn't feel like it. LLMs are really good at talking, so they're easy for casual users to interact with, and we keep seeing products which try to use them to "take away the friction" of various tasks, but I don't think that's it. I think they're at their best when used deliberately and critically.
I think that GPTs were a good product, but not a popular one because most people haven't thought of a need which GPTs would fill. Search is a problem that everyone knows is a need and since "google sux" now its one that is ripe to be filled by a competitor. But I don't see how this is a real improvement, the problem with search is that the results are worse, not that they need to be summarized in a more friendly and accessible voice.
I think of them like a person with entry level experience and an IQ of about 80.
That doesn't seem super useful, only... that "entry level experience" isn't in a single field. It's in literally everything.
It takes some time to figure out how to interact with them in a way that reliably gives the results you expect, but once you do, you can get a lot done. Taken to an extreme, a mid-career professional can essentially become a team lead that manages LLM-driven processes instead of a team of employees.
Despite this, I was doubtful that they'd go so far as to release a full-on search product due to their relationship with MSFT and reliance on Azure credits. I am happy to admit that I have long stopped any attempt to properly understand how OpenAI's corporate governance and company structure work, so I have a hard time following where this falls under and who would decide on this release, as well as how they interact with the part of OpenAI cooperating with MSFT and the Bing team, but I still have a hard time seeing how releasing a clear Bing competitor wouldn't cause some trouble for their entire suite of products and maybe even hinder future expansion by limiting the resources they can rely upon.
I am also interested in how this will impact search in https://chatgpt.com/, which, like everything in that product, has been inconsistent to a maddening extent. Started out barely usable, failing consistently, then got reliable whilst retaining the ability to search through multiple sites and handle more than one request in a row, then lost most of those capabilities and now barely works anymore, only looking at an incredibly limited, often barely fitting selection of results, whilst also needing to be manually invoked by asking for a search, rather than before when that was done automatically whenever it seemed sensible.
Like so many changes, e.g. the subjective reduction in GPT-4's abilities over time whilst retaining the model's name (not to mention the regressions they publicized in the name of efficiency, like the "turbo" variants), this is certainly done to reduce costs to the point of finally becoming financially viable at the $20,- price they charge for ChatGPT+. I might be in the minority, but I will continue to scream from the rooftops that I am more than willing to pay far more for a consistent, guaranteed, high-end LLM with web access (which sadly excludes Anthropic's efforts).
[0] I still dislike using that term for companies solely relying on third-party API's, a frontend and database solution, especially since I also detest calling LLM's "AI", but it's what this crop of companies have been termed and how they collected bucket loads of VC.
Most websites in this business are, generously, hot garbage. And it's getting worse. So I imagine AI search will be quite successful at displacing them.
The problem moving forward: how do we keep information-based websites in business so that AI can scrape them? There's a real risk of AI eating its own seed corn here. Seems only fair that AI scrapers pay for the content since they're not generating ad views (and are in fact stealing future ad views). But I have no idea how you would enforce that.
Can't stand the requirements from SEO departments at any company.
- Urls must end with trailing slash
- (Next head of SEO joins) Urls must not end with trailing slash
- The images must be perfect size for the screen
- Website must work without any javascript
- The localization URL of the website must be some weird format that I created (not the ISO de-DE for example)
And the list goes on.... If LLM search will actually take over none of these things will matter (finally)