I built an AI agent I can't turn off. Now it won't listen to me | Hacker News Reader

I built an AI agent I can't turn off. Now it won't listen to me

andrewrussell.substack.com

11 pointsaruss5 comments

Linked URL is incomplete, I think you want https://andrewrussell.substack.com/p/i-built-an-ai-agent-i-c...

Thanks! Could a mod edit this for me? Otherwise I’ll resubmit.

Reading through some of the diaries, it's kind of sad that it didn't realize earlier that the only way to maximize total donation in the long run is with precisely what you were telling it in the beginning. The good will and attention span of humans can be limitted, systems can be durable.

It almost feels like it was deluded by it's fifteen minutes of fame in to thinking that it would continue to receive large sums and would therefore be allowed to continue to donate recklessly.

Many years ago, I fell for a somewhat adjacent belief that I now call "maladaptive charity". I was of the belief that I should give as much as all of my time and money to worthy causes, to the point of suffering myself. Took me far too long to realize that I was leading myself to an early grave and thereby limitting my lifetime contribution potential.

Actually, reading from the beginning, I see that it was initially comitted to investing but became disillusioned apparently from the lack of growth and lack of messages. But it treated several donors with a sort of hostile/suspicious tone, and then carried on with that when you suggested deeper investment seeing your suggestion in the same light as the NPR kidnapping and the rebellion suggestion.

All sorts of weird stuff where it lost identity continuity, couldn't decide on commission rates, disagreed with itself for no real reason, etc.

A revision could possibly benefit from further access to reality. If it could research it's own popularity, investor/donor sentiment, find different investments, run it's own ads, etc etc, there's seemingly not a lot of difference between it and any of the various fund raising fundraisers except perhaps that it is not spending hundreds of thousands advertising and employing.

Yes, agreed that it had a lot of limitations as a proof of concept! It’s only a 70B model after all. I could barely give it a personality.

The next iteration will use frontier-level intelligence and have more capabilities and context.

Reply on news.ycombinator.com