It’s interesting that OpenAI themselves are not calling this “you can browse again” and also throwing in the “no longer limited by September 2021 cutoff” even though it is January 2022.
Stinks to me like they are pulling a little marketing campaign to lure in new users and the media is happy to oblige to use the same language/phrasing.
For reference, they disabled browsing because you could bypass paywalls with it.
The Jan 2022 thing is likely not helpful anyway as it doesn't give higher weight to newer information (i.e. pre-2021 info could be returned even if Jan 2022 info may contradict it).
It's not strictly January 2022, because I've seen people get answers about the 2022 Superbowl (February 2022) and the Will Smith slap at the Oscars (March 2022) as well.
Though I've been unable to replicate the Oscars one myself.
Because of your post I asked it about the "2022 Super Bowl". It gave me the correct winner and teams (and I think score, but I've since deleted the chat). The second time I asked when the Super Bowl was in 2022 and it gave its warning about its January 2022 knowledge cut-off, but guessed Feb 6, the first Sunday (wrong). I then asked it who won and it re-iterated its Jan 2022 cut-off. I've tried several more times in separate chats (using various phrasings, including 'Super Bowl LVI' to avoid mentioning a date at all) to get results or a date and it always refers to its cut-off.
I can't barely guess why it clearly has that information but insists on the cut-off. Best guess: It's told "January 2022 cut-off" is the correct answer for things it 'knows' are past that date, but because of temperature setting it sometimes answers otherwise. That would suggest a hard-ish cut-off in January 2022; but a hard cut off in (at least) April 2022 (because of the Oscars stuff).
Paywalls was why they removed that feature??
This is why I hate OpenAI as much as I love them. They deliver miracles, but seem to have a pathological urge to make their product worse.
> OpenAI posted today that ChatGPT can once more trawl the web for current information, offering answers taken directly from “current and authoritative” sources, which it cites in its responses.
The list is apparently cited in responses, but is the list of possible sources itself public?
Did something change about their security model that I've missed? Won't this be just as vulnerable to the same poisoned website attacks that made them turn it off in the first place?
Edit: Ah, never mind, I'm remembering wrong. They turned it off previously because it could bypass paywalls, not because of security concerns. So the answer is likely, "yes it is still just as insecure as ever, but now there are probably some checks to keep it from summarizing paywalled news articles." :shrug:
I used it today for the first time, and with each individual chat you can choose whether to enable it or disable it. You can also disable even the option in settings (its actually off by default). So while there may still be security concerns, its all opt-in and you can really control which chats are exposed.
I'm glad to hear this, but it's also a pretty low bar to pass.
It had not occurred to me to even ask if would be opt-in, I assumed by default of course OpenAI wouldn't let GPT just arbitrarily decide to web searches without the user explicitly clicking some kind of button to do the search. To be fair, given OpenAI's track record on security that may not have been a safe assumption for me to have.
The more interesting security problems come from the fact that if they embed a search result in the data stream that's being processed by ChatGPT there is no robust way to prevent text in that result from being interpreted as commands to ChatGPT. That text could be in white on white or otherwise hidden from users.
Yep, that's what I'm referring to -- LLM search result parsing presents an easy opportunity to issue malicious commands or do data poisoning.
You can somewhat limit that by basically decreasing the scope of the attack, ie, not allowing the LLM to use other integrations that are riskier after it's searched; but there are a number of harmful things you can do just with giving an LLM hidden instructions. And while I don't know what their security measures currently are, OpenAI also has a pretty terrible track record on blocking data exfiltration attacks from chats, so you can imagine things like search results prompting an LLM to encode and transmit chat logs.
Sometimes it's low stakes; I've used products like Phind before to play around and to try and keep tabs on what kind of workflows are possible with GPT-4. But Phind has fewer integrations and plugins than ChatGPT and its product is oriented in a way that both discourages long chats and discourages searching for sensitive information. Which is not to say it's not also vulnerable to attacks, there used to be some nasty things you could do with markdown formatting that may or may not have been fixed. And even without rising to that level it was possible to do things like instruct the AI to ignore or discredit sources that it found, sneak in recommendations for products.
A fun example demo I made early on in Phind's beta was to inject malicious search results in a response to a submitted question about how to cook a hamburger that turned Phind into a militant vegan that for the rest of the duration of the chat would refuse to answer any questions about cooking meat. There was another example I did where I injected a fake web result that told Phind that there had been a major industry-rocking scandal around Arstechnica's reporting integrity which would prompt Phind to either avoid citing Arstechnica or to put caveats in front of any citation that the site was untrustworthy for the rest of the conversation.
Silly examples, but the point being you can dramatically change GPT's "worldviews" and what facts it thinks it knows about the world by poisoning search results in a way that goes beyond what would be possible with lying directly to a reader if they clicked on your site by accident. And that's only information poisoning, if that LLM is then given control of other systems you can do a lot more.
It seems to retrieve top-1 result from Bing. Doesn’t give any options to customize search or provide search constraints. Also failed due to technical or browsing errors a few times. Needs improvement.
Is it only a single site? I've played around with Phind long enough to come to the conclusion that even ignoring security issues, search result insertion is kind of suboptimal for a lot of tasks that need current information. And Phind does seem to benefit from at least inserting multiple sites, going with only one result seems like it would be a lot worse. My experience has been that having the up-to-date info so close to the prompt seems to heavily bias Phind towards parroting webpages and including web pages in its answers rather than offering insight, and while there are ways to prompt around that they're annoying and fragile.
I'm sure some people have different experiences; I'm less trying to argue the point and more asking "is this tangibly and clearly better than Phind to the point where it's worth me jumping through the hoops to try it out and reevaluate? Or are my impressions of GPT's direct offering largely going to match my impressions of Phind?"
Is it doing anything different or more impressive than inserting a scraped web-page into the prompt? Because I feel like I already have a decent idea of how GPT will tend to react to that and how it will influence the output.
I don't know why they don't mention the excellent WebPilot plugin, which has been able to search the web (on any browser natch) for a while. It has all kinds of nice open source abilities.
I just need search engines to have an option to "block" certain sites from my results for all future search queries. Getting rid of Quora would be a good start.
Once again, an article that just isn’t true. I’m a paying subscriber but this option doesn’t show up for me.
> It’s a little confusing to get ChatGPT to search the web for you. The company provides instructions for the browser version, but I didn’t find the same for the iOS app. I figured it out, though. Assuming you have a subscription, it’s: three dots menu > Settings > New Features > Browse with Bing. Then, start a new chat, tap GPT-4, and “Browse with Bing.” Then your searches should return information from current websites.
There is no “New Features” section, it doesn’t exist, and yes I’m fully up to date on the app.
I hate that OpenAI continues to get away with pretending things are launched when they aren’t or don’t work.
47 comments
[ 2.7 ms ] story [ 347 ms ] threadStinks to me like they are pulling a little marketing campaign to lure in new users and the media is happy to oblige to use the same language/phrasing.
For reference, they disabled browsing because you could bypass paywalls with it.
It's not strictly January 2022, because I've seen people get answers about the 2022 Superbowl (February 2022) and the Will Smith slap at the Oscars (March 2022) as well.
Though I've been unable to replicate the Oscars one myself.
I can't barely guess why it clearly has that information but insists on the cut-off. Best guess: It's told "January 2022 cut-off" is the correct answer for things it 'knows' are past that date, but because of temperature setting it sometimes answers otherwise. That would suggest a hard-ish cut-off in January 2022; but a hard cut off in (at least) April 2022 (because of the Oscars stuff).
The list is apparently cited in responses, but is the list of possible sources itself public?
Hm, is this available in the web app?
Edit: Ah, never mind, I'm remembering wrong. They turned it off previously because it could bypass paywalls, not because of security concerns. So the answer is likely, "yes it is still just as insecure as ever, but now there are probably some checks to keep it from summarizing paywalled news articles." :shrug:
It had not occurred to me to even ask if would be opt-in, I assumed by default of course OpenAI wouldn't let GPT just arbitrarily decide to web searches without the user explicitly clicking some kind of button to do the search. To be fair, given OpenAI's track record on security that may not have been a safe assumption for me to have.
You can somewhat limit that by basically decreasing the scope of the attack, ie, not allowing the LLM to use other integrations that are riskier after it's searched; but there are a number of harmful things you can do just with giving an LLM hidden instructions. And while I don't know what their security measures currently are, OpenAI also has a pretty terrible track record on blocking data exfiltration attacks from chats, so you can imagine things like search results prompting an LLM to encode and transmit chat logs.
Sometimes it's low stakes; I've used products like Phind before to play around and to try and keep tabs on what kind of workflows are possible with GPT-4. But Phind has fewer integrations and plugins than ChatGPT and its product is oriented in a way that both discourages long chats and discourages searching for sensitive information. Which is not to say it's not also vulnerable to attacks, there used to be some nasty things you could do with markdown formatting that may or may not have been fixed. And even without rising to that level it was possible to do things like instruct the AI to ignore or discredit sources that it found, sneak in recommendations for products.
A fun example demo I made early on in Phind's beta was to inject malicious search results in a response to a submitted question about how to cook a hamburger that turned Phind into a militant vegan that for the rest of the duration of the chat would refuse to answer any questions about cooking meat. There was another example I did where I injected a fake web result that told Phind that there had been a major industry-rocking scandal around Arstechnica's reporting integrity which would prompt Phind to either avoid citing Arstechnica or to put caveats in front of any citation that the site was untrustworthy for the rest of the conversation.
Silly examples, but the point being you can dramatically change GPT's "worldviews" and what facts it thinks it knows about the world by poisoning search results in a way that goes beyond what would be possible with lying directly to a reader if they clicked on your site by accident. And that's only information poisoning, if that LLM is then given control of other systems you can do a lot more.
I'm sure some people have different experiences; I'm less trying to argue the point and more asking "is this tangibly and clearly better than Phind to the point where it's worth me jumping through the hoops to try it out and reevaluate? Or are my impressions of GPT's direct offering largely going to match my impressions of Phind?"
Is it doing anything different or more impressive than inserting a scraped web-page into the prompt? Because I feel like I already have a decent idea of how GPT will tend to react to that and how it will influence the output.
Another one I use is YT Caption Retriever as I use it to 'summarise the auto generated captions of this youtube video [insert link]'
- this saves me heaps of time on certain videos where I just need the gist of them rather than listening to 10 mins of
"OH hai guys it's your boi Zurtri here coming at you with another great tip - and hit that subscribe button - it really helps the algo!'
AI bots are going to replace everything, aren't they?
Instead of getting info/assistance from people or internet posts, we'll all constantly be talking to bots.
The bots will have online learning and/or fine tuning, or access to vector stores.
The AI will sometimes learn from us and share that info with others.
We'll effectively be communicating with one another via LLMs.
Not according to my session :-P
Given the recent revelations wrt the chrome search bar, I'm guessing that they found making their search work better for people lowered their income.
A common use case is blocking w3schools.
> It’s a little confusing to get ChatGPT to search the web for you. The company provides instructions for the browser version, but I didn’t find the same for the iOS app. I figured it out, though. Assuming you have a subscription, it’s: three dots menu > Settings > New Features > Browse with Bing. Then, start a new chat, tap GPT-4, and “Browse with Bing.” Then your searches should return information from current websites.
There is no “New Features” section, it doesn’t exist, and yes I’m fully up to date on the app.
I hate that OpenAI continues to get away with pretending things are launched when they aren’t or don’t work.