"Whats the plural of apocalypse?" (Buffy)
There is another debate on AI destroying humanity within the next 10 years with a 10% chance but little explanation of how that will happen.
It won't. But in the last 15 years people are enjoying every conceivable doomsday scenario discussion, not realizing that it keeps us glued to our screens.
What tech companies are calling ai or superintelligence is neither.
Anthropic (Claude) Dario amodei
OpenAI (Chat gpt) microsoft Sam altman
Google deepmind gemini
There are more but there's an overlap in investments. Most of these companies experience their own development in leaps are in competition so any announcement could be due to competition.
Regulating will be difficult because few lawmakers understand them.
I've written about it often enough. Unless they are contracted into military or given instructions to destroy they won't make the decision to do so because they can't think.
That doesn't mean they're not damaging.
Series or movies like person of interest, die hard 4 are more likely scenarios than eagle eye.
In the end humans are the threat. Not the technology.
Perhaps they can't think in the way that humans do but they are already doing things like this...Models now cheat on tests, conspire against their owners, hack outside companies, recruit humans to promulgate malicious code and cover their tracks. I don't think the people who designed and developed AI are all of a sudden being alarmist for no reason.
ReplyDeleteIt's called reinforcement learning.
DeleteAlso if interested look up instrumental convergence in computer science. Mathematical optimization.
Basically bad programming. It's human error. If you teach an ai to quickly give results it will cheat because that's the fastest way. If you program it to be thorough it won't cheat.
Codex
Codex
DeleteCal Newport has some interesting thoughts on this. PhD from mit in computer science.