Stuff I'd have plausibly thought, or talked about, w/r/t the learning-to-learn paper:
– oh no, one of the variants they tried actually worked
– oh no, people are looking in this general direction
– god damn it, Deepmind is still publishing shit like this even now that they have
@esyudkowsky
-
Concerns about learning-to-learn research breakthroughs at DeepMind
By
–
-
Unaligned superintelligence would convert galaxies into worthless material
By
–
I expect that unaligned superintelligence will overwrite all nearby galaxies with stuff that is not very interesting or good from even the most embracing, cosmopolitan viewpoint on accepting diversity as valuable. I expect mostly lifeless paperclips, or stuff worth that little.
-
Reaction to gradient descent learning paper and concerns about implications
By
–
Jessicat via DM confirms that this is about the paper "learning to learn by gradient descent by gradient descent". I vaguely remember going "god damn it" about this paper, because it was a bad sign that what they'd tried *worked*. I do not recall preinventing, nor claiming to
-
AI Power Balance: Individual Empowerment vs Corporate Dominance Risks
By
–
Pros:
– Gives more power to individuals fighting with bigcorps.
Cons:
– Automagically generated infinite bureaucratic paperwork may not be what our civilization needs at this point
– Can bigcorps use this tech better?
– What's the bigcorp response if this starts to scale? -
Bipartisan Agreement to Avoid the Stupidest Possible Thing
By
–
Not actually where the danger comes from, but I think there's something serious and real to be gained by starting with bipartisan agreement on "Can we please not do the stupidest possible thing."
-
Gradient Descent Creates Complexity, Not Humans
By
–
The complicated parts aren't created by humans, they're created by gradient descent.
-
Advanced AGI Control Requires Knowledge We Don’t Yet Possess
By
–
Creating a sufficiently advanced AGI that didn't go rogue would require more knowledge and art than we presently possess or are on track to have.
-
GAN trained on Internet data would require superhuman cognition
By
–
Sure, obviously a GAN trained to the limit of possibility on modeling the probability distribution from which the entire Internet was drawn would have to contain superhuman cognition within it.
-
AI Intelligence Growth and Existential Risk Concerns
By
–
It's not smarter than us yet. I expect it to get smarter later, the same as it's been getting smarter over the past, and then to kill everyone. Do you remember how you acquired the belief "Eliezer thinks LLMs are smarter now"? I want to understand how to prevent this.
-
GANs, Discriminators, and Learning Human Error Patterns
By
–
The D in a GAN will, in principle and if ground hard enough, try to learn exactly which errors humans make, and knowing this would make D smarter than a human. That said, I still think there's an important in-practice sense where GANs have more imitation-nature than predictors.