Recent searches
Search options
I gave my geometry students some ChatGPT-generated "proofs" this week to review. There were several examples, each designed to illustrate a different point. One was a "proof" that the diagonals of a rectangle are congruent, which contained several errors. I was proud that several students immediately identified how dangerous it was: "It sounds like it is correct, until you look more closely at it."
@phonner the one you shared here was definitely eye-opening...
This thread from October might interest you, check out the article I included about how LLMs are "stochastic parrots"
As AI is trained by human generated content, it's not artificial intelligence, it's average ignorance.
Lol, your line made me laugh.
But I thought about it a realised it's even worse than that! I don't think there are many people who put online erroneous proofs. Maybe in some students forums or something, but surely, that's statistically insignificant.
So LLM are discovering some very novel and imaginative new ways to get things wrong.
It's like an extraterrestrial expedition comes to earth and that species shrinks in its aging process. Therefore they think the smaller people are, the elder and wiser they are. And they visit a kindergarten and let the kids explain human society, politics and technology.
@phonner That‘s a great way to teach the use of LLM.
I think that LLM are extremely helpful in domains the user is proficient in. It‘s like a very eager student who recently read a lot about something but just reproduces and has no real world experience. And makes up stuff on the spot instead of admitting that they don‘t know.
It definitely is helpful if you‘re training to find non obvious mistakes 
@phonner I'm seeing if Google's Bard/Gemini tool can do basic mathematics and OH MY GOD it's not good. Textbook quadratic ok, anything else... Well...
@hutchingsmusic @phonner They also have a first year's ability to tell the difference between kWh and kW (which is to say not very much)
@hutchingsmusic @phonner Note that is PaLM2, Google's old language model, not Gemini. Hopefully Gemini will be better once it's actually available.
@hutchingsmusic @phonner Your screenshot says it's PaLM2. Maybe it's not launched where you live?
@phonner Billions of these arguments, on politics, values and the worth of human beings, posted by hundreds of millions of different accounts and websites, both by text, audio and video, flooding the internet.
@phonner I wonder what would happen if you asked ChatGPT to prove that the volume of a pyramid increases tenfold when you double the length of the edges. It shouldn't be able to, of course, since it increases eight times, but I bet it'll confidently do it wrong anyway.
@phonner sure but this is the case with anyone. Like if I was an engineer...which I am... and I made a statement about stuff I work on, everyone at work would believe it in a second. Unless ofcourse someone paid attention and told me I was wrong. Then I could learn from my mistake.
We are telling ChatGPT when it is wrong so it can learn from its mistakes. That information then is used by the companies using ChatGPT against us.
@phonner brilliant use of genAI in education!
@phonner
Have it square the circle next.
-> @byte@awawa.club @phonner same with code. it throws up some correct-looking code, but often completely misses edge cases and other places for potential dangerous bugs. people write bad enough code already, no need to automate it :)
I have a friend who teaches history at concordia University in Montreal; she recently had their students generate some ChatGPT text about areas of history they were studying. They were gobsmacked by how many errors were riddled throughout
@phonner ooooh can you share them with me please? This sounds fun.