3 ms·
The issue with those things is we actually know how to measure success for most of them, and they notably work best when the politicians stay the hell out of th
by XorNot 2y ago
The issue with those things is we actually know how to measure success for most of them, and they notably work best when the politicians stay the hell out of the process other then setting broad directions.
There's no equivalent for AI research, and a lot of temptation to try and find simple metrics which mean safe - and it's appealing. Water should be below X ppm contaminant is "easy" to establish. AI should be <blank> to be safe?
- AlbertCory 2y agoI don't get the feeling you read the link. But thanks.
- XorNot 2y agoWhat is your definition of what makes an AI system dangerous or safe? How would this be measured or standardized? Not vagueries, not "well superintelligence should be regulated". A company owns a 100,000 Nvidia GPUs. What are you implementing? What physical actions are the people enforcing this policy going to take? Someone, somewhere will be at a computer entering commands, or writing code, or looking at data. What will it be?
- AlbertCory 2y agoare you asking me? I'm not the one with the government program.
- samatman 2y agoThe person you're snarking at is not replying to you, and is under no obligation to read your Substack before posting.