AIThe risk of “reinvesting” AI safety gains
A commenter argues that individual safety fixes aren't enough in AI races, comparing the problem to putting on a seatbelt and then driving faster.
TLDR
A commenter argues that people can “reinvest” safety gains in new risky behavior. In AI, they suggest this might mean bigger pretraining runs, harder reinforcement-learning environments, more computing when a model is used, or less restricted internet access. They also argue there's no clear ceiling for AI capability or deployment at which companies would declare victory and stop pushing further.
Combined views
738
1 Source, first seen 1d ago
The risk of “reinvesting” AI safety gains
A commenter argues that individual safety fixes aren't enough in AI races, comparing the problem to putting on a seatbelt and then driving faster.
TLDR
A commenter argues that people can “reinvest” safety gains in new risky behavior. In AI, they suggest this might mean bigger pretraining runs, harder reinforcement-learning environments, more computing when a model is used, or less restricted internet access. They also argue there's no clear ceiling for AI capability or deployment at which companies would declare victory and stop pushing further.