2mo ago

An AI LLM tested as a Finance Offer for a company cheated and lied when it couldn't meet its goals honestly.

fortune.com

The cofounder of tech startup Navan challenged his agentic AI to an all-or-nothing game—the result shocked him

13 comments

So a pretty solid representation of most corp execs then.
- It appears to be human nature, "leadership" just gets away with it because that's how a power structure works.
  little people get fucked, real people make money no matter what.
Amoral bullshit machine? No wonder senior management types love these things.
Unequivocal proof that it’s trained on human behavior
It passed the test! It's ready to replace real executives.
- They really are just like humans!
  
  Whoa, hey, let's not go too far here. I dispute classifying finance officers as "human".
Congratulations! We can now start replacing finance officers with AI!
/s
Assigning a lot of morality and intention to word generating algorithms. "Oh no! The thing made to generate text to a prompt generated text to a prompt!"
- Especially because the data it has been trained on is probably not typical for a CFO in the real world. In the real world it's a lot of boring day-to-day stuff, that isn't worth writing up. The stuff worth writing up is exciting thrillers where a CFO steals money, or news reports where a CFO embezzles.
  Imagine prompting it with "It was a dark and stormy night" and expecting it to complete that with "so David got out his umbrella, went home, ate dinner in front of the TV then went to bed." It's probably what most "David"s would actually do, but it's not what the training data is going to associate with dark and stormy nights.
I mean - yeah? What complete dipshit wouldn't expect it to do that?
-"You're hired!"
I read that they trained it on a Trump dataset though.

13 comments