3 ms·
If I used the term, I would reply, "A safe powerful AGI is one with a less than fifty percent chance of killing more than one billion people." (Because most of
by Eliezer 8y ago
If I used the term, I would reply, "A safe powerful AGI is one with a less than fifty percent chance of killing more than one billion people." (Because most of the work is pushing the failure probability sigmoid from 0.999+ to below 0.5, and people will disagree on how much lower than that it could be pushed.)
- backpropaganda 8y agoTheir blog post (https://blog.openai.com/concrete-ai-safety-problems https://blog.openai.com/concrete-ai-safety-problems) seems to have a starkly different idea of safety.