|Bing+GPT: This Is Not A Search Engine|
|Written by Mike James|
|Wednesday, 08 February 2023|
I have been waiting for the day that AI could take over search and now it seems to be here. The real question, however, is whether or not Bing+GPT is actually a search engine? The answer seems to be a very clear No.
Arguably search engines made the Web. The ability to find information on something you were interested in made the Web useful. Right from the start finding things was the problem and we had various inadequate solutions. We started off with technologies like Gopher and manually constructed subject guides and went on to things that appeared to be clever like page rank.
All absolutely and irredeemably flawed.
Today search engines use a mess of heuristics to work out which pages to suggest you might want to look at and we all know how that's worked out. Search is broken, but it has always been broken to a greater or lesser degree. This is the reason that I have been waiting for AI to fix the problem. But I missed the point.
A suitably general AI should be capable of reading web pages, working out if they are of a sufficient quality to be relevant to a search but.. and this is the point I missed, such an AI would also be capable of learning what the pages were saying and producing its own summary.
This is what Microsoft has just announced. Bing plus AI based on the GPT large language models isn't the new way to do search; it's the new way to get answers to specific questions.
It's an answer engine.
We only have a mock up of what it is supposed to be like and, at the moment, this isn't too impressive. The sample question is:
"Write code to find the Fibonacci sequence in Python."
and the answer returned is:
This is good answer to the question posed. However for me it isn't the right question. Getting a listing of a program that generates a Fibonacci sequence doesn't really hack it. I want some background, an explanation of what is going on - in short I don't just want an answer I want some information transfer into my brain that makes me better equipped to use the answer and to extend the answer.
So my complaint isn't about the answer but rather the wrong question. What about asking it not just write the code but explain it? Other non-programming examples that Microsoft offers take us in this sort of general direction. The problem here is how much do I trust this sort of answer?
GPT, and so far all large language or foundational models, aren't grounded in fact. They do not check the statistical relationships that they have observed in their training data. What this means is that they are inherently untrustworthy. I can see that the program Bing has offered me does compute what it claims to, but I cannot so easily check that any explanations or factual claims are correct. One question at the live presentation was exactly how Microsoft was going to ensure that the new Bing+GPT didn't lie or hallucinate? In this case there was no real answer.
The mockup of Bing+GTP has what look like standard search results appended to the "intelligent" answer. So perhaps this is not so much a replacement for search as an augmentation. I'm not sure how useful this is, given now the human has to sift through this graded search result - here is your answer but just in case you probably should read through these other items. If this is what Microsoft is proposing, and it does seem to be:
It is easy to say "it also cites the sources" but this isn't exactly what is happening - it is just augmenting the answers with the usual SEOed search results. If this is what is planned then search is even more broken than it was before GTP. Now you have unreliable AI answers plus SEOed irrelevant search results.
This all sounds very undercooked and you might be wondering why Microsoft is in such a rush to serve us this dish? The answer is, of course, very obvious - Microsoft has long wanted to damage Google and take over its lucrative position. However, Microsoft is in more danger of destroying the position rather than taking it over. If it is successful in replacing search engines with answer engines then where is the revenue coming from? Who will write the material that the answer engine learns from if there is no profit in it? Despite the best attempts by ad blockers ads still pay for the web, Bing+GTP has the very real potential to disrupt this market.
Google has gone into red alert mode because of this threat and its response is to build its own GTP in the form of Bard. However, this might not be the right response as this could demolish Google's position just as easily as Bing+GTP.
There are two main possible outcomes of this development:
Notice that none of this proves that GPT and similar models aren't useful or important - they are, but it is what they lead to that is important..
You can sign up to the wait list to eventually discover if the Microsoft mock up is anything like reality.
Reinventing search with a new AI-powered Bing and Edge, your copilot for the web
Open AI And Microsoft Exciting Times For AI
The Unreasonable Effectiveness Of GPT-3
The Robots Are Coming - AlphaCode Can Program!
The Unreasonable Effectiveness Of GPT-3
Does OpenAI's GPT-2 Neural Network Pose a Threat to Democracy?
To be informed about new articles on I Programmer, sign up for our weekly newsletter, subscribe to the RSS feed and follow us on Twitter, Facebook or Linkedin.
or email your comment to: firstname.lastname@example.org
|Last Updated ( Wednesday, 08 February 2023 )|