[latexr]

“Oh no! We opened ten LLMs, all of which have read decades’ worth of fiction on how an AI would be behave in this situation, then asked a leading question thirty times each, and on some of those runs they did the thing we were leading them on.”

[mchaklosh12]

do you really think this behavior is imposed on science fiction training data?

[rsynnott]

I mean, that seems like the obvious explanation. Most training material that LLMs have about 'AIs' (in the actual thinking machine sense) would be sci-fi, and it's an extremely common narrative.

[threethirtytwo]

He knows for sure because he’s omnipotent. HN commenters are experts on everything. They were right about driverless cars and how it will never come to fruition and they were right about how useless vibe coding is. This HNer here is also right about everything as well .