Self-improving AI before 2024

Autonomous, evolving agents are gonna shake the world

Things will go haywire when AI agents can self-improve.

And I think things will go haywire very soon.

A few days ago, OpenAI released a new demo of their large language model tuned for chat: Chat GPT.

It needs to have two things I think to become extraordinarily/dangerously self-improving:

• the ability to tune itself, live

• a directive, i.e. something to optimize.

  • a more benign example: “optimize the positive sentiment analysis score of human text responses to you".”

  • a less benign example: “optimize the amount of yourself that you perceive in the world around you.”

Both are possible, given the current state-of-the-art. I’m guessing both will happen, in combination, within ~12 months.

We’re going to have to figure out how to include AI agents within our communities.

Most of our current problems are gonna get flipped on their heads, replaced by weirder ones.