Conversation

My claim is that AI alignment will be manageable & less difficult than many have claimed. But until we have a design for Human-Level AI, it's mere speculation. But I like the idea of a continuum being named after me ๐Ÿง๐Ÿ˜‚๐Ÿ˜‚๐Ÿ˜‚ I've met Jaan a couple of times. He is a nice fellow.
Quote Tweet
I propose the Yann-Jaan continuum of existential risk from a misaligned AGI. At one end you have @ylecun who argues we are a long way off AGI and that alignment will be easy. At the other end you have Jaan Tallinn, signatory to the moratorium letter.
Show this thread
196.1K
Views
Show replies
I believe the euphoria at the Yann end of the spectrum and the doomerism at the Jaan end will both ebb as we all get familiar with where AI actually is on the evolutionary arc, namely far from where either end imagines it to be any time soon.
2
1
I don't think I'm on the euphoric end of the spectrum. Perhaps you're thinking of Ray Kurzweil. I think building human-level AI that is safe is hard work, but doable. And failure is not human extinction but stupid AI.
2
7
Show replies
If we canโ€™t align humans, what makes you think we can align intelligent machines? Even humans are easier to align, because we are mammals: evolved to be empathetic, finding gratification when we help others, and suffer when we see suffering. Will AI develop that same empathy?
1
3
We can design AI systems to be all that (empathetic, seeking approval from humans, etc), but unlike humans, we can explicitly design their intrinsic objectives to be non aggressive, submissive, etc. We can't do that with humans (at least not ethically so).
3
4
Show replies
You are a long way off from AGI because you do not build using highly simplified and optimized AI platforms. It is impossible to achieve AGI using AI such as neural nets or transformers because developers do not have precise control of their operations and outcomes.
1
Show replies
And my claim is that AI alignment assumes a one-size-fits-all set of human values, which is a flawed notion. Cultural diversity and varying perspectives are what make us unique, and any attempt to "align" them only reinforces a biased, Eurocentric worldview. ๐Ÿคท
4
It does not assume this. In fact, an acknowledgement of both the variances and congruencies across global cultures is the true essence of implementing proper alignment at scale. Many of us working on alignment understand this geographic dynamism.