AI Safety Pizza Night (Stockholm) (Weekly)

Schedule

Tue Oct 04 2022 at 05:30 pm

Location

Albano (Stockholm) | Stockholm, ST

Advertisement

The AI can't fetch your chai latte if its dead.
Sufficiently advanced AI’s will form sub-goals for self-preservation – the probability of you getting your morning latte is significantly lower if the AI ceases to exist.
This is one of the so-called Convergent Instrumental Goals.
Another is Goal Integrity – If an agent retains its present goals into the future, then its present goals will be more likely to be achieved by its future self.
Thus, if (or when) we program an advanced AI with the wrong goal, it will try to stop any modification of it.
These problems might, on an initial thought, seem to have some trivial, intuitive solution. They don't.
We have no good idea how to build systems smarter than us, safely.
One day, someone will build such a system, probably with some goals which upon further inspection did not capture the full intent of the programmer.
In trying to stop any modification of its goals, the AI will deceive the programmer into believing its goals successfully have been changed, then seek to gain enough power until its sure that all attempt to shut it down – will fail – to then proceed with its original, misaligned goal.
Few work on AI Technical Alignment. We wish it wouldn't be that way. Even small contributions are of great value; the stakes are so high.
Help us nudge the odds in favor of humanity. Join us!
We will have a laid back conversation about the most intriguing problems of AI Safety, on a weekly basis. Location and time are decided with regards to participants requests.
Signup form:
PS
One cannot save the world on an empty stomach so everyone gets a vegan pizza.
Advertisement

Where is it happening?

Albano (Stockholm), Stockholm, Sweden

Event Location & Nearby Stays:

Chris Gerrby

Host or Publisher Chris Gerrby

It's more fun with friends. Share with friends