From The Washington Post Tech Brief newsletter:
“All large language models, by the very nature of their architecture, are inherently and irredeemably unreliable narrators,” said Grady Booch, a renowned computer scientist. At a basic level, they’re designed to generate answers that sound coherent — not answers that are true. “As such, they simply cannot be ‘fixed,’” he said, because making things up is “an inescapable property of how they work.”Grady Booch is in Wikipedia.
Have you had your rock today?
Related posts
ChatGPT e-mails a professor : AI hallucinations : ChatGPT writes a workflow : ChatGPT summarizes Edwin Mullhouse : ChatGPT’s twenty-line poems : Spot the bot : Rob Zseleczky on computer-generated poetry : ChatGPT writes about Lillian Mountweazel : ChatGPT on Ashbery, Bishop, Dickinson, Larkin, Yeats : ChatGPT summarizes a Ted Berrigan poem : Teachers and chatbots : A 100-word blog post generated by ChatGPT : I’m sorry too, ChatGPT
comments: 4
Which probably goes to show that LLMs and AI generally are indeed converging towards a human perspective on the world... we're almost all "unreliable narrators" and it's surely fair to say that "making things up is an inescapable property of how [we] work".
I guess if one wants something that always comes up with the right answer (like say flight control software) then you're better off with a straightforward algorithm - which may have bugs like any system, but at least it's built to always produce the same outputs from the same inputs. But if you're wanting an interesting conversation then variability is a lot more fun :)
Well, there’s improvising (which is life itself) and then there’s saying anything — different kinds of making things up. I’ve seen enough AI hallucinations to stay skeptical.
On a forum for some software I use for driving a robotic cutting machine I posted the following in regards to a post on one of the, as it were, “Chatty-Cathy” bots. I think it might be relevant here:
Reading this thread I thought about Lewis Carroll's "word ladder" puzzle game. The rules from Wikipedia:
“The player is given a start word and an end word. In order to win the game, the player must change the start word into the end word progressively, creating an existing word at each step. Each step consists of a single letter substitution.”
“An example for "boot" to "shoe" would be: BOOT, SOOT, SHOT, SHOE.”
I just plugged into the Opera browser's Aria "AI" the question "Tell me about word ladders" and it paraphrased the Wikipedia article and gave the above rules.
I then asked it:
“Solve a word ladder using the start word BOOT and the end word SHOE”
It replied:
“To solve a word ladder from "BOOT" to "SHOE," we need to find a chain of words where each word differs from the previous word by only one letter. Here is one possible word ladder solution:”
“BOOT, BOOK, LOOK, LOSE, SHOE”
“In this word ladder, we start with the word "BOOT" and change one letter at a time to eventually reach the word "SHOE." Each step in the ladder involves changing one letter to form a new word that is valid and differs from the previous word by only one letter.”
“Please note that there may be multiple valid word ladder solutions, and the one provided above is just an example.”
I then asked it to solve a word ladder from BOOT to STILETTO. It had a go and came up with (same boilerplate 3 paragraphs): BOOT, BOOK, LOOK, LOOT, LOST, LIST, LITE, SITE, SIT, STET, STILETTO.
If I asked a child that had never heard of Word Ladders (and they are a great many of them) to solve these two word ladders I expect that that they would get a correct solution to the first one, eventually, and they would tell me the second one can't be solved because the rules don't allow the start word to get bigger.
When the various "machine learning" apps can do as well as the child at this without them needing to be fed this post to add to their "knowledge" base. Then you've got AI.
PS. My previous attempt at getting information from one of the bot things was asking: Show me songs that use the phrase "Denn alles Fleisch, es ist wie Gras". It replied that this was from the 5th movement of Brahms' German Requiem.
I replied "you are wrong it is actually the 2nd movement of the Brahms Requiem". The response: "Sorry, you are correct it is the 2nd movement".
I know that I'm correct. But the thing is that you are less correct than a simple web search using a non-"AI" search engine and those make no claims at having learned anything.
———
Something that I had thought of since that post. The correct answer to the BOT to SHOE ladder was in the Wikipedia post. Even though the machine learnining thing summarized that Wikipedia post and would access to the solution it still gave the wrong solution!
I love these examples — thanks for sharing them here. They remind me of what I think is the best metaphor for what these bots produce: a Play-Doh version of the Internet.
I didn’t know that the word ladder come Carroll. I know it as word golf, something that plays a part in Pale Fire: love, lave, have, hate. (Par 3.)
Post a Comment