In your example, if I told that at random about 30% of the results were made up, you would not consider that a time saver. In fact it would be total time waster since you would have to vet every single entry. People think since 70% is accurate, only 30% work is needed but not if you don't know which 30% is bogus. You would need to check the entire work using conventional means including perhaps a 'regular' search engine.
Almost 100% next iteration will sport a fact checker, powerful style and format controls, and a much larger context. The development of advanced fact checkers will have a big impact on anything propagated online.
It's not going to write me something I'll hand to an editor. But for certain things, it could definitely give me a head start relative to a blank sheet of paper.
imagine where we'll be just two papers down the line!
There are largely three groups of people
1) ChatWhat?
2) It only makes bullshit!
3) OMG, this amazing, and scary and amazing, and useful. Oh wow...
The LLM won’t revolutionize search as it is today for factual queries. They’re Clippy 2.0. It’s great people are finding use for the models, but I wish this search story would be balanced out a bit.
I was laid off recently, and I’m using the LLM to write a bunch of cover letters. I give it my resume, a blurb about the company and job and a bit about what I like about work, and it outputs a cover letter. I don’t like writing BS cover letters where I pretend I majored in the company mission and my whole life has been teaching me their values. GPT can do that for me though- and yes I fact check but I’m fact checking against my resume and personal opinions which I obviously know quite well.
The ability of an LLM to generate decent content (provided you're an attentive editor or the users of the content aren't too discerning) could be huge for Office365, but that's irrelevant to any potential threat to Google, since Docs is of very little importance to Google's revenues and strategy in a market where Office is completely dominant and has always had a more full-featured product.
True. And also, keep in mind ... it doesn't truly have to be _better_ than Google search. You just need to start and maintain a _social trend_ so that the mainstream public _chooses_ it over Google. People use Google because it's the first and only option that comes to mind -- they haven't actually compared its accuracy to anything else in a long time (the audience of Hacker News is of course an exception).
There are a few types of search queries that people seem to do, factual lookups ("who is the exec of abc?"), but also generally treat the search engine as the entryway to the internet ("I need a teaching plan about Ukraine"). We'll see that LLM fall flat for facts (assuming people care), but they can supplant some of the general traffic. Realistically, its a bad fact search replacement, but it could be a great tool to put next to a search bar, making a better "starting place for accessing the internet".
With the teaching plan example, the original user was probably going to make a query for a template (or 5), then copy+paste, then do 10-100 queries learning all about Ukraine history and culture, then rewrite that into the template, editing down to manageable size, then send to peers to edit and review, then format for distribution. That could be dozens of Google searches. Now, one or two AI queries, and they have a template, basic written text, and can focus on a couple queries for fact checking. Oh, and since they used bing to do the AI part, they may just stick with bing for the fact check part. Google was irrelevant in that whole flow instead of getting dozens of queries over a day before, but if that feature was moved to Office365, then they may never have used bing for search while still killing a chunk of google's traffic.
The danger to google is not equal to the opportunity to bing. If 5-10% of traffic never reaches a google search, that's a huge chunk of google's revenue, even if it doesn't translate to searches on a different engine. Think of the potential impact an AI code generator could have on StackOverflow. When I need to pick up a new language, I often query "how to append to an array in python" in a search engine, but a LLM (or large-code-model) built into my IDE could supplant that query entirely. I
I doubt you hand wrote cover letters by making dozens of search queries (similarly, I doubt people devise teaching curricula by learning the history of Ukraine through a series of Google queries). But when you weren't taking the time to write them yourself, I bet you had more time free to search for jobs, or do general internet browsing using Google as your gateway to the internet...
People having more time free to browse the internet is unlikely to be a threat to Google's business, even in the highly unlikely scenario Google is incapable of advancing its existing AI products beyond their current state
aka "bullshit"
…and replaced by b.s. chatbot wranglers, who will be paid much more, and who will produce more total output, and who will be selected preferentially from among the people that best understand the work the chatbots are doing…so, yeah, in lots of cases, the people that were writing bullshit will end up wrangling chatbots, in jobs that bring in more money for their employer, and probably at higher pay (though, by historical trends of automation, a lower share of the generated value) who write bullshit.
This is even more clearly the case for people who have writing bullshit as an incidental part of their job rather than a core part, since the incidental part will consume less time, increasing productivity, without eliminating need for the core job. So its not even a “lose one job but move to the replacement job” situation, its just a “be more valuable in existing job”.
It just needs to be useful enough to dislodge the google monopoly.
People are lazy and are creatures of habit. Give them a single plaace to talk to ChatGPT and to search, they'll take it.
Then traditional search is going to become "raw index search" that you can query writing something like "intitle:"carbonara" source:"google_index"" or "give me all webpages containing carbonara in its title"
Really, you can believe me, outside Silicon Valley, nobody cares or cared about Juicero.
This is very different for ChatGPT, and I'm sure that if you get interested to it you'll find interesting usages with it that can fit your daily workflow (or just fun! like with image generation models).
One useful thing I learned from this, though, is that ChatGPT can handle Polish just fine. It never even occurred to me to try it - I incorrectly assumed the model was trained on English text only. I suspect that being multilingual from day 1 was a huge factor in ChatGPT's sudden and extreme user growth.
This is also a pattern. Their voices are not better than some paid services (NaturalReader for example). Their OCR and document understanding is inferior to Amazon Textract. Even in speech recognition there is the excellent Whisper from OpenAI doing just as good or better. Google's generative image models are not the best, and locked away for good measure. I think SD and MJ rule.
Google's AI was cool in 2000 for search and in 2016 for games. But now the best people are leaving them - almost the whole team who invented transformers has their own startups.
I also think maybe, just maybe, their TPUs are bad and they can't scale high quality models to the public. Maybe they lost the race because GPUs were better in the end. Maybe it's stupid, but how can we explain the lack of advanced AI? The other explanation is they won't mess with something that makes them so much money (current search/ad model).
To put it another way: the main value proposition of using something like ChatGPT to navigate the internet is that you're putting your trust in it to filter out the noise on your behalf. If you can't trust it to actually do that (there's still ad noise in what you get back), then what's the point?
Either people will pay a subscription fee to unlock the utility of an information-distilling agent, or they won't. Trying to sidechain ad revenue into that equation is self-defeating.
Adding a chat front end is just going to lower the SNR, because ChatGPT has no idea what facts are or how to check them.
Unfortunately it's also the main attraction for corporate revenue generation. You can sell stuff conversationally. Woo hoo. These systems are going to turn into automated used car sales bots which use persuasion techniques to steer users towards a sale.
From the user POV the main attraction is the prospect of a kind of universal summarising WikiBot and bureaucratic paperwork automator.
Those are fundamentally different domains.
Users have been pretty relaxed about being manipulated and distracted by social media and covert PR/sales/influencer operations, so there's going to be a huge market for the bad stuff.
But it's just corporate noise, as it always is. The real value will come from processed search in the sense of automated teaching and intelligence augmentation.
Unfortunately there's not where most of the research will go. It's not going to become common until LLMs are taught to fact check with high reliability, and the cost of entry is low enough for that to be offered as a service.
Meanwhile - yes, exactly: ads disguised as search results.
This feels a bit like projection though. People in general are trained to tolerate ads for most freemium services, such as social media, search, etc., and chat is no different.
For any market involving human attention, there's a portion willing to pay money for the service, but a significant larger portion willing to trade attention time (e.g. ad impressions) for a free service instead.
Right. That's a transient state, unfortunately - we can trust ChatGPT now because we know OpenAI had neither the time nor resources nor a reason to make their tool biased for commercial purposes (they're busy biasing and constraining it so it doesn't generate too much bad press, but this doesn't affect the trustworthiness of responses to typical queries). A model like this obviously won't be allowed to gain widespread adoption as a search proxy - it's destructive to commercial interests.
> If you can't trust it to actually do that (there's still ad noise in what you get back), then what's the point?
Exactly. The problem is, as users, we have no say in it. If Microsoft and Google decide that conversational interfaces are the future, then we'll be doing searches via ChatGPT-derived sales bots. End of story. Google and Microsoft each have enough clout to unilaterally change how computing works for everyone. And if they both decide to compete on quality of their ML search chatbots, there's no force on Earth that could stop it. Short to mid term, if they want it, we have no choice but to use it (long-term this might create an opening for a competitor to claw back some of the search market with a chatbot-free experience).
> Either people will pay a subscription fee to unlock the utility of an information-distilling agent, or they won't. Trying to sidechain ad revenue into that equation is self-defeating.
This, unfortunately, has been proven false again and again. Newspapers. Radio. Broadcast TV. Cable TV. Music streaming. Video streaming. On-line news and article publishing. And so on.
Advertising is a disease, a cancer that infects and slowly consumes every medium and form of communication we create. Often enough, creation of a new medium is driven by the desire for an alternative, after the old medium became thoroughly consumed by advertising and seems to be reaching terminal stage.
Side-chaining ads into a chatbot interface is going to be even more powerful than ads in normal search results - not only you can tweak the order of recommendations like search engines do today, you can also tweak the tone and language used in the conversational aspects, effectively turning the bot into a sneaky salesman.
But is this related to search? I'm a "pro ChatGPT" user (ie, I pay) and I don't use it for anything search related. It's entirely unrelated.
* How does getting recommendations from a chatbot (what TV to buy) play with websites that produce such content (TV reviews) * How does it play with websites that rely on ad impression * How can you monetize a chatbot? (there's an easy way: free tier + monthly subscription) * How to reduce the massive compute cost of a good chatbot without making it bad (this also seems more straightforward)