[ExI] propose an ai experiment: was RE: Why AI agents invent their own language if you let them chat.

[spike at rainier66.com]

-----Original Message-----
From: spike at rainier66.com <spike at rainier66.com> 

>...Have you noticed how AIs are being treated as arbiters of truth?

>...Skip the next paragraph if you are in a hurry. 


OK retract, don't skip the next paragraph.
 

>...Do perform that experiment I suggested a few weeks ago: find a topic about which you know more than anyone, about which you have opined online in the past....

Then drill down until you can get the AI to hand back something that is unambiguously false, complete provable baloney.

It's a game or contest.  The prize: one meeeeelllllion respectars.  Made of... iron pyrite.  Imaginary variety.

Example: AI showed up in about fall of 2022.  I wanted to learn how it works, so I was able to find enough info on it such that matrix algebra hipsters can grok.  I next wondered how quickly the AI was post trained on news: did AI read the newspapers?  My experiment was to ask ChatGPT in November 2024 who won the presidential election.

For several days after the election, it was handing back the wrong answer, without telling me it didn't yet know.

Oh how I wish after the fact that I had archived that.  I was still treating AI as an educational toy, rather than as an arbiter of truth.

Find a topic on which you know the answers.  Use it to test the AIs.  Word the question identically for each.

Related: ask the AIs who won a ballgame that just happened.  See which one gets the right answer soonest.

Fun experiment: ask your favorite AIs if Musk overpaid for Twitter.

I asked Duck.ai, ChatGPT and Grok.  Duck and GPT say Musk definitely overpaid.  Grok agrees kinda.  I thought that one was cool, because Grok IS TwitterX.  It is analogous to one's asking: Am I worth what I cost?

Grok opined in a most delightfully nuanced way which I shall share in a subsequent post, but I don't want to influence your experiment, possibly harming your chances at one meeeeelllllion respectars.

spike