,
3 tweets,
1 min read
Read on Twitter
There is a gap in current the current Silicon Valley AI safety research ecosystem. Most organizations are studying artificial general intelligence, but much dumber systems will come online long before we have AGI, and may do a lot of damage. 1/3
Examples include pricing AIs that learn to collude (already demonstrated in the lab), biased criminal justice risk assessment, and algorithmic trading systems that end up funding arms dealers and environmental damage. These systems need to be built to encode social values. 2/3
The work of labs like @OpenAI and @MIRIBerkeley is important but decades away from use. We need concrete and computable social metrics for the near term. We need engineering best practices -- today -- for building systems that protect environment, fairness, inequality, etc. 3/3
