AI ALIGNMENT FORUM
AF

Wikitags

Myopia

Edited by Dakara, abramdemski, et al. last updated 30th Dec 2024

Myopia refers to short-sightedness in planning and decision-making processes. It describes a tendency to prioritize immediate or short-term outcomes while disregarding longer-term consequences.

The most extreme form of myopia occurs when an agent considers only immediate rewards, completely disregarding future consequences. In artificial intelligence contexts, a perfectly myopic agent would optimize solely for the current query or task without attempting to influence future outcomes.

Myopic agents demonstrate several notable properties:

  • Limited temporal scope in decision-making
  • Focus on immediate reward optimization
  • Reduced instrumental incentives
Subscribe
Subscribe
Discussion0
Discussion0
Posts tagged Myopia
34Partial Agency
abramdemski
6y
16
40The Credit Assignment Problem
abramdemski
6y
32
62How LLMs are and are not myopic
janus
2y
7
20Towards a mechanistic understanding of corrigibility
evhub
6y
25
33Open Problems with Myopia
Mark Xu, evhub
4y
15
24Steering Behaviour: Testing for (Non-)Myopia in Language Models
Evan R. Murphy, Megan Kinniment
3y
5
32LCDT, A Myopic Decision Theory
adamShimi, evhub
4y
44
17Defining Myopia
abramdemski
6y
13
35Arguments against myopic training
Richard_Ngo
5y
38
48You can still fetch the coffee today if you're dead tomorrow
davidad
3y
14
91The Parable of Predict-O-Matic
abramdemski
6y
16
72An overview of 11 proposals for building safe advanced AI
evhub
5y
32
54Seeking Power is Often Convergently Instrumental in MDPs
TurnTrout, Logan Riggs
6y
34
44MONA: Managed Myopia with Approval Feedback
Seb Farquhar, David Lindner, Rohin Shah
7mo
7
39Thoughts on “Process-Based Supervision”
Steven Byrnes
2y
1
Load More (15/34)
Add Posts