As far as why Anthropic should probably race, here’s @joshc’s take on it, using the fictional company Magma as an example:
https://www.lesswrong.com/posts/8vgi3fBWPFDLBBcAx/planning-for-extreme-ai-risks#5__Heuristic__1__Scale_aggressively_until_meaningful_AI_software_R_D_acceleration
The other winning pathways I can list are:1. Unlearning becomes more effective, such that you can use AI control strategies much easier.
2. We are truly in an alignment is easy world, where giving it data mostly straightforwardly changes it’s values.
3. We somehow muddle through, with an outcome that none of us expected.
As far as why Anthropic should probably race, here’s @joshc’s take on it, using the fictional company Magma as an example:
https://www.lesswrong.com/posts/8vgi3fBWPFDLBBcAx/planning-for-extreme-ai-risks#5__Heuristic__1__Scale_aggressively_until_meaningful_AI_software_R_D_acceleration
The other winning pathways I can list are:
1. Unlearning becomes more effective, such that you can use AI control strategies much easier.
2. We are truly in an alignment is easy world, where giving it data mostly straightforwardly changes it’s values.
3. We somehow muddle through, with an outcome that none of us expected.