What they've accomplished so far is really impressive. But until I see AGI for real I won't hold my breath.
Just feels to me like some expert/wizard level PR is going on right now.
When we have AGI, we will find it out pretty quickly since whoever gets there first wants the most money and control they can.
To the average person chatGPT is literally magic.
Again, there's plenty of different conclusions that can be made on the risk levels but that doesn't change his expectation on timing of development.
He is basically the next generation's Ray Kurzweil but instead of AI granting us biological immortality now AI is going to cause mass extinction.
Instead of uploading consciousness to some digital nirvana we get the slightest variation on The Matrix. "And then it takes over the galaxy and uses the atoms in our bodies for batteries!!"
With the way AI captures the imagination we can't but help take things to irrational extremes.
From everything I've seen from him, his concerns revolve around (1) a lack of a bounding function for the development of AI (2) a view that the alignment problem is unsolvable and (3) the risk that we won't recognize an AI was developed until it is sufficiently more intelligent than us to pose a fundamental risk.
I raised it only as an example of a prediction he made that most at the time thought had no chance. My point isn't to claim he's some genius that gets everything right, but there are examples of him making predictions widely viewed at the time of having no chance that turned out to be correct.
Writing him off because people responding strongly to his predictions of AI destroying us all doesn't do anyone any good. His core arguments and predictions seem sound and I haven't yet found anyone to explain how we can solve alignment after the fact, for example
I'm writing him off because he isn't a practitioner. I can make things up about what will or won't happen on my own, as can anyone else.
If I want to believe every wacky concept from science fiction will happen, I don't need a middleman to tell me it will.
It does feel risky to write off anyone that isn't a practitioner. There won't be any practitioners working on AI risk, its a theoretical pursuit by nature. The closest a practitioner could get is AI risk management, i.e. attempting to limit risk for AI we're already developing.
Taken in that light, his being correct about his protein folding prediction feels like cold comfort when taken in context of his predictions, largely unsubstantiated, of AGI becoming a malevolent human extinction force at par with or of greater risk than that atomic bomb. It elides the very real conceptual limitations of current implementations of AI, such as it's inability to recognizing bidirectional associations (Tom is John's son means John is Tom's father), protecting against injection risks, "hallucinations" (inability to differentiate between when it is making something up and when it is not) and a variety of other existential limits towards its own power.
These details, far from being inconvenient nuances that do not detract from his overall point, should actually serve to make us seriously skeptical of how well he understands the risks of the entity he is warning against. A more skeptical observer might even say (and I'm one of them) that much like fire and brimstone preachers, he is trying to create an unrealistic fear of a deific entity that cannot be reasoned against so that he can consolidate power by serving a "solution" that protects against it and curry popular favor. I find his approach to be the distasteful and manipulative approach of the common demagogue. The haughty, arrogant, insecure way that he treats the arguments of his critics (he simply blocks them on Twitter or condescends to them) make me even more skeptical of him. It makes me wonder if on some level he knows that his argument is not being made in good faith, and isn't trying to win an argument in good faith.
How many of his predictions are even disprovable? Do you have a list? Was that protien fold prediction even disprovable?
An example of a nondisprovable prediction is:
-one day an AI will become self aware.
A disprovable one is:
-Skynet will begin to learn at a geometric rate. It will becomes self-aware at 2:14 a.m. Eastern time, August 29th 1997.