Scan that ship for life, ChatGPT!
ChatGPT offers blatantly wrong information while sounding confident and authoritative. This could be more dangerous than Fox News.
Backstory
I had gotten curious about the alleged amazing TV moment when Alexander Butterfield disclosed the Nixon taping system. Apparently it came as a shock to everyone when he did so. I was interested in seeing the stunned looks, so I found a clip. Uh-oh.
Fred Thompson simply asked point blank if there was a taping system. I asked ChatGPT how Thompson knew to ask that? Standard investigatory question, maybe?
Enter ChatGPT
Me: How did Fred Thompson know to ask about the taping system?
ChatGPT: “Fred Thompson, …learned about the existence of the White House taping system from a book written by former White House aide John Dean. In his book, Dean mentioned a conversation he had with President Nixon in which they discussed the taping system.
“ Thompson was intrigued by this information and decided to ask Alexander Butterfield, who had been scheduled to testify before the committee, whether there was any truth to Dean’s claims.
“During Butterfield’s testimony on July 16, 1973, Thompson asked him directly whether there was a recording system in the White House. “
Hmm. John Dean wrote a book before the hearings? Really, ChatGPT?
Me: When did John Dean write a book mentioning a taping system?
ChatGPT: “John Dean wrote a book titled “Blind Ambition: The White House Years” that was published in 1976, after the Watergate scandal had ended. In the book, Dean described a conversation he had with President Nixon in which they discussed the existence of a taping system in the White House…”
Cue WC Fields. “A child of three can plainly see…”. I could fix that with ten lines of Lisp, but with machine learning we do not want to interfere, we want the Alpha Zero thing, becoming an unbeatable Go or Chess player on its own. That way we learn new chess tricks never before discerned, such as not being too quick to recapture.
Sure, but after forty-four million games. Where it could score itself with perfect accuracy after each game. Not an option.
Can they fix it? Have you ever programmed Prolog? First it produces a terrible result, then we cannot even figure out how it got that result. Or constraints programming. I saw a brave “where we stand” review from the nineties. “Getting predictable behavior can be a challenge…”.
After we have the courage to take Fox News off the air, we have to get Captain Kirk to chat ChatGPT into a Mobius loop like he did the malevolent Star Fleet computer. It might go something like this.
“Why did Thompson ask that specific question?”
“He heard about the taping system in Dean’s book.”
“When did he ask the question?”
“1973.”
“When was the book written?”
“1976.”
“Did Thompson read the book before it was written?”
“My language model is not capable of speculation.”
“What does the word ‘before’ mean?”
“My language model is not capable of time travel.”
“Scan that ship! How many people have you killed?!!!!!”
“Would you care for a game of chess, Dave?”
In 2001, Dave could pull out the circuit boards from the rack. Try that with cloud computing.
I wonder when ChatGPT will score its first fatality?
Just sayin’.
