How the Chat GPT user interface "works" is very seductive. But one can understand it if one looks at the familiar Google search engine's auto-complete for queries as a simplified example. If I type "New York" into Google, it fills in "Yankees" first (because first pitch in today's Yankees home game vs the SF Giants is in half an hour), and the other choices are:
Mets (Baseball team, but their game is in Miami today, so less New York City centric)
Times (the Newspaper)
City (The city itself)
Rangers (Ice hockey team)
Giants (Football team)
weather (It's sunny and 47F today)
Knicks (Basketball team)
Jets (Football team)
The focus on sports is interesting here, as neither Jets nor Giants are playing until fall, The Knicks are at home, playing the Wizards at 6pm today, and the Rangers are winning 2-0 over the Capitals in Washington as I type. But why all the sports? Google mentioned no churches on a Sunday, no parks, none of the 4 zoos in town... dunno. But Google's special sauce is the context added by a constant stream of fresh queries from others so it is always weighted towards "what's happening NOW".
Now, if Google had been "spying on me", it would have figured years ago that we are season-ticket holders for the Mets, the Jets, the New York City Opera, the Metropolitan Opera, the City Ballet, and the ABT, and it would have listed them first. But it plainly doesn't spy to help me with more appropriate replies to my queries, it just spies to sell ads, so it did it's best to complete the query for the most likely OVERALL subjects of interest that someone would have when querying from Manhattan at this time of day on this particular Sunday in April.
Chat GPT is doing much the same thing, just with a lot more flexibility and verbosity, with a goal of "conversing". It is a "large language model", trying to finish a sentence, a thought, or a story in a plausible SOUNDING way. All this is very ad-hoc, it does not formulate an "answer" before starting to reply, it just starts generating text, and then improvises its way to the end, based upon a very large dataset of human-generated text, which it assumes to be plausible chunks of thought.
So, it is simply picking the next STATISITCALLY-likely word or phrase to follow what it just typed, and using factoids and buzzwords that are statistically more likely to have a connection to the inquiry it is answering.
It is like Jazz improvisation or reggae - it is certainly based upon "experience" but it is unlikely to ever give the same answer twice to the same essay question, as it is making things up as it goes along.
Could we "teach it" about beekeeping? Yes, but we'd have to be careful about not feeding it contradictory information without assigning weights to various conflicting views. But, without any filtering or weighting, the AI is said to have "passed" the "uniform bar exam", the SAT's reading, writing and math sections, the GRE, and a list of other "tests", including a Wharton School of business MBA exam, likely prompting guffaws from Sloan School grads.
What GPT-4 scored poorly on was math - the AMC 10 is a 10th grade (high-school level) algebra, trig, and geometry test, and it scored in the 6th to 12th percentile. It did much better on the 12th grade version of the test, so the weak point is very likely plain geometry, which is stressed early on, but ignored in later grades, but there have been multiple reports of GPT screwing up very basic math.
So, no surprise, a language model does poorly at geometry and math. It lacks a grasp of spatial manipulation, and it only knows that 5 times 8 is 40 because it "read" that multiple times somewhere, and it may not be quite so sure about 8 times 6.
So, if a large language model of any sort is able to answer questions about beekeeping, it would only be parroting the least common denominator between multiple written sources. The bad news here is that many beekeeping books, articles, and discussion group posts are chock-full of stuff that is repeated over and over, only to later be shown to be a bunch of old wives tales, and never mentioned again. So, it will, at best, be able to parrot a well-read person who has never actually kept bees, and has ignored recent findings that invalidated prior "conventional wisdom".
So, GPT, if it "knows" about bees is most likely to seem to be an older beekeeper who stopped reading the literature some years ago.
***********************************************
The BEE-L mailing list is powered by L-Soft's renowned
LISTSERV(R) list management software. For more information, go to:
http://www.lsoft.com/LISTSERV-powered.html
|
