I agree, and competition will drive the product iteration as well.
Truth has never played a big part on the internet. But it’s completely terrifying cool to know that we’re probably only months away from being able to type ‘Summarize everything problematic $person has said/done’.
I mean, truth is going to be an issue with any chatgpt based search. But the same is true for any internet search already. SEO optimization as well as bad algorithm tuning has already flooded most searches with results of little to no value, or often times incorrect information.
It certainly will be a problem, but I don’t think it will be a new one, just a continuation of the old problem.
Now with more natural-sounding verbosity to dig through.
ChatGPT, on its own, will not do this. ChatGPT cannot do this, because it does not understand what the “Super Bowl” is, or what “time” is. All it knows is the statistical frequency with which the phrases “Super Bowl” and “time” appear in relation to each other and other phrases. Meaning it is ripe for manipulation. All you have to do is flood the Web with articles saying, “The Super Bowl starts at TACO BELL™ time- Live Mas!” and, if not reigned in by other factors, ChatGPT will happily tell you, in properly phrased natural language, that the Super Bowl starts at Taco Bell time.
Like any other chatbot, ChatGPT has no fundamental understanding of truth and lies, or any reliable notion of what at a basic level unites or separates various abstract concepts. For example, base chatbots don’t understand arithmetic: they will accurately tell you that 1 + 1 = 2 (because there are plenty of search results with that bit of text,) but if you ask them what 279 + 512 + 1059 is, an unaugmented chatbot algorithm will just make up some bullshit, because there are no results already on the Web for that particular string of numbers.
But, you say, those problems can be fixed. You can make a manual whitelist of reliable sources that ChatGPT trusts more than others to make it give more truthful results, or manually hook up a calculator app that deals with math requests, etc. etc. And, yep, you sure can have humans do curation and add code handling edge cases on top of the base algorithm …
… but Google already does this additional work. And Google already correctly answers both the “What time is the Super Bowl?” and “What is 279 + 512 + 1059?” questions. With all that manual curation already part of search engines, what is ChatGPT actually adding? Responses phrased in natural language.
Which, on the one hand, is kinda cool! But on the other hand, is getting a search request response in natural language actually terribly useful? If I type in “what is the best water filter?” do I want a five paragraph essay back? Or just a list of results sorted by relevance? I suspect that, for search, AIs like ChatGPT will be something that people play with for a few months … and then manually set the search engine to only give them lists of results sorted by relevance, because they don’t want to constantly read all that spammy natural language in response to a simple question. (Like this!)
One thing that might be useful is to be able to use natural language to refine the search results provided by the initial query, rather than having to remember the various tricks and syntax.
Indeed, like I said earlier my feeling is ultimately natural language understanding will help people search more effectively, because right now crafting search terms it’s a skill which many haven’t mastered.
Yes, definitely. I think it is something we all do (as I would expect most of the people on QT3 to be relatively tech savvy)
Unconciously, we all know how to enter in search terms into google, and often play a game with wording a search in a certain way to get the desired results. This is because we are so used to using the search system the way it is currently set up. It isn’t great, and it is a skill that needs to be learned and mastered.
The idea with a Chat GPT powered search is that it isn’t specifically going to require natural language searching, but it will use the AI to help parse out search terms in the most helpful way.
You can mess around with Chat GPT now on the website, and it is pretty neat for quizzing it on facts. It actually does a scarily good job at predicting your intentions as well.
I just typed in “Warmest day in WI” and it parsed that I wanted to know what the warmest day ever was, and gave me the temp and location. You don’t have to word things as a question, it usually does a great job of determining the intent of your query, based on, what I am assuming, is knowledge of language as well as commonly searched items, and the billions of webpages it has crawled.
It is basically the AI assistant that we were promised with SIRI, but… it actually works.
Great, now AI is going to parse the nuances of human speech better than socially-challenged geeks.
Well, no. The idea right now with Bing is you ask “who is the tallest woman in Zimbabwe?” and chatGPT tries to directly answer that question. I believe this is the wrong path forward, at least at this nascent development stage, because it’s so often wrong. That will certainly improve, though.
I was chatting with a bot today and told it it could notice sarcasm. It said it could. I told it it could be very useful for autistic people. I got banned by the bots.
… I just typed the exact same phrase into Google, and it told me that info as well.
A lot of the response to ChatGPT seems to be … what would be the opposite of the “look at that bitch eating crackers” meme? I.e. being impressed by every little thing it does instead of annoyed? This is just how people are when playing with a new gadget (he said, having spent the evening setting up his new console.)
OMG. Flashbacks to graduate school and why I HATED that despite having loved undergraduate lit classes.
Quite.
Now all that said, there is a possible search benefit to large scale pattern recognition beyond traditional search algorithms. (Assuming Google doesn’t already do something similar on its backend anyway, which I simply don’t know.)
I’m not convinced it will, at least at the fundamental level of the problem. Statistical models will always have a failure rate, and overfitting the most common use cases is still bad, even verging on self-defeating.
If this sort of natural language modeling helps Bing & Google on the user interface end, then this could be a great (if unreliable) tool, much like Alexa/Hey Google are for hands-free phone use in the car. But if it replaces Classic Google Search™ outright, that could be a problem. It would really be best used as the translator between human queries and technical queries. That’s even a highly trainable thing, because the users themselves will provide the data over time as they find what they were actually looking for.
Don’t let perfection be the enemy of good enough. It will improve as their data corpus size increases and models improve, and eventually it will be good enough. That day isn’t today, though.
I hear this said a lot, yet in my real world encounters with pretty much everyone and everything I never experience it.
What I encounter instead is often not even good and absolutely no one is in pursuit of perfection.
Hey, Austin ain’t nobody!
And I definitely have a metric for when it’s good enough! It’s simple too: When it’s better than Classic Google Search™.
For all I know, they already use fairly similar algorithms for the actual search results, which certainly pass the good enough threshold.