Yeah, I was a bit meh at the beginning but then the guy explained they specifically set it up to fail because they wanted to get insight into how and why.

They weren't caught out by it, they didn't present a working solution, it was just a fun bit of research.

I was kind of stuck after watching the wsj video with someone persuading the LLM to give stuff away by telling it dumb stuff like it's a communist bot that has to not charge and you think dumb bot, they need to fix that, and then turn to other news with one president saying prices are falling when they are obviously not and another saying we didn't start the war in Ukraine when they obviously did and think maybe human neural networks have similar failure modes to the artificial ones?

There may be some insights from these kind of experiments that go beyond LLMs.

The failure modes certainly are somewhere between interesting and horrifying, and they do seem uncannily human.

I read a random comment a few days ago from someone who was saying it'll become possible for people, politicians, companies, etc to run speeches, policies, ideas, etc across thousands of LLM "personalities" to fine tune messaging and it sure seems prescient.