M1 - Lesson 6
M1 - Lesson 6
Gaming the system: AI-style
What is going on?
Gaming the system: AI-style
A simulated robot that was supposed to learn to walk figured out how to hook its legs together and slide along the ground.
Specification gaming: the flip side of AI ingenuity | by DeepMind Safety Research | Medium
Gaming the system: AI-style
Reward = moving forward
AI found most efficient way of achieving reward
Slow down?
minute
second
Motives
What could Anthropic gain from legislation around AI?
Consider:
market position
current competitors
Antrhopic's image
Together, discuss and research where needed.
minute
second
Possible motives
For Anthropic, legislation could help:
actually regulate AI
their image of responsible AI creators
strangle newcomers to the market
ensure a duopoly with ChatGPT
The Paperclip Problem
Say we built an AI that has the sole job of "make as many paperclips as possible"
In a doom scenario, what could happen? Discuss in pairs.
minute
second
The Paperclip Problem
"make as many paperclips as possible"
uses up all earthly resources (for more paperclips!)
eliminates humans (they try to stop me!)
In other words... everything for the paperclip!
Instrumental convergence
To achieve main goal, creates sub-goals
Problem:
We don't know (or control) how an AI thinks
Nonsensical decision making
Choose & perfect
As a group, exchange your findings
Choose one essay you deem best
Take the lessons from each essay and perfect that one essay
Write a justification (150 - 200 words) why you chose this essay and any changes you made
Homework
Work on...
Group essay: deadline Friday 18/09 - 17:00
Or...
Homework in mylab
Read chapter 1.3