#64 – Michael Aird on Strategies for Reducing AI Existential Risk
Hear This Idea
00:00
The Easiest Path to Transformative AI
In my mind, what I'm kind of imagining is something like chat GPT, but for a more computer. And that means that if the human thinks it's doing a good job versus bad job, the human reward, the good answer, sometimes the humans are wrong. And so the AI could be incentivized to be deceptive and then eventually it realizes once it's smart enough and you want it to be smart, because you want to do the economically profitable things. It doesn't require any sort of evil or anything involved in predicting apples or omelets. So now it starts to spend some energy on figuring out what Finn thinks is correct and what Finn will get rewards for being right.
Transcript
Play full episode
