4 ms·
I don't know how existentially dangerous AIs may become There is very little evidence or theoretical support for a hard limit on AI capability. In fact, overwh
by stcredzero 3y ago
I don't know how existentially dangerous AIs may become
There is very little evidence or theoretical support for a hard limit on AI capability. In fact, overwhelmingly, most of the evidence and thinking supports the notion that over time, the capabilities of AI will greatly over-match human beings in specific contexts. The same even goes for general intelligence.
How benign or malicious an entity is, is orthogonal to how potentially dangerous it is. Humans are demonstrably very dangerous. What we're creating has the potential to be orders of magnitude more dangerous.
So yes, we desperately need alignment! And even if we achieve it, how do we know the AIs won't convince Homo sapiens to simply stop breeding and fade away in alignment with our values?
Hard to know the future.
- AnimalMuppet 3y ago> In fact, overwhelmingly, most of the evidence and thinking supports the notion that over time, the capabilities of AI will greatly over-match human beings in specific contexts. The same even goes for general intelligence. How much of that "evidence and thinking" is actually evidence? Isn't it mostly just thinking, especially with respect to general intelligence?
- stcredzero 3y agoA chief piece of evidence is the progress made by scaling up the models.
- yanderekko 3y ago>And even if we achieve it, how do we know the AIs won't convince Homo sapiens to simply stop breeding and fade away in alignment with our values? If it could do this, then we should seriously question whether it's actually a bad outcome.
- simonh 3y agoWhat goals would a better outcome serve, and why should we care about them?
- ftxbro 3y agolol this should be such an uncontroversial take but you are getting downvoted so hard. I only hope they are downvoting so much because they are bored by the take that everyone already understands and agrees with it.