Overparameterized value models keep machines obedient at modest cost
Evidence
“What we found,” Hadfield-Menell says, “is that if you have an overparameterization of the space of values, then you end up learning the correct thing, but it takes you a little bit longer. If you have an underparameterization, then you end up quickly becoming pretty disobedient and just becoming confident that you know better than the person.” — Brian Christian, The Alignment Problem, 1302
Categorization
- corroborated: 2 independent passages support this claim — Brian Christian, The Alignment Problem, 1302
- XREF: connects to: Autonomous systems become deferential whenever they remain uncertain about user wants — Brian Christian, The Alignment Problem, 1292; AI alignment may require machines to pursue human objectives, not their own — Brian Christian, The Alignment Problem, 1167; Ensuring AI captures human intent is now a central computer science problem — Brian Christian, The Alignment Problem, 65
Related
- Autonomous systems become deferential whenever they remain uncertain about user wants (similarity 0.81)
- AI alignment may require machines to pursue human objectives, not their own (similarity 0.80)
- Machines could learn human values by observing behavior instead of being programmed (similarity 0.80)
- Ensuring AI captures human intent is now a central computer science problem (similarity 0.80)
- Learning reduces uncertainty that is what motivated deference to humans (similarity 0.78)
Broader Connections
- AI alignment may require machines to pursue human objectives, not their own
#corroborates(0.70) — Brian Christian, The Alignment Problem, 1167 - Ensuring AI captures human intent is now a central computer science problem
#corroborates(0.55) — Brian Christian, The Alignment Problem, 65