I'd need to have links and to read it by myself.
Here is his interview. It's very, very hard to tell if he's got his tongue firmly in cheek (he refers to minds of human-level intelligence and our problems as being "small"), or if he's enjoying an opportunity to troll the hell out of some organization with a low opinion of his work.
With regards to reinforcement learning, one thing to note is that the learning process is in general not the same thing as the intelligence that is being built by the learning process.
With respect to genetic algorithms, you are correct. With respect to something like neural networks (real world stuff) or AIXI (pure theory), you are incorrect. This is actually why machine-learning experts differentiate between evolutionary algorithms ("use an evolutionary process to create an agent that scores well on X") versus direct learning approaches ("the agent learns to score well on X").
Not understanding this, local genuises of the AI risk been going on about "omg he's so stupid it's going to convert the solar system to smiley faces" with regards to at least one actual AI researcher.
What, really? I mean, while I do get worried about things like Google trying to take over the world, that's because they're ideological Singulatarians. They know the danger line is there, and intend to step over it. I do not believe that most competent Really Broad Machine Learning (let's use that nickname for AGI) researchers are deliberately, suicidally evil, but then again, I don't believe you can accidentally make a dangerous-level AGI (ie: a program that acts as a VNM-rational agent in pursuit of an inhumane goal).
Accidental and evolved programs are usually just plain not rational agents, and therefore pose rather more limited dangers (crashing your car, as opposed to killing everyone everywhere).
With respect to something like neural networks (real world stuff)
Well, the neural network in my head doesn't seem to want to maximize the reward signal itself, but instead is more interested in maximizing values imprinted into it by the reward signal (which it can do even by hijacking the reward signal or even by administering "punishments"). Really, reward signal is not utility, period. Teach the person to be good, and they'll keep themselves good by punishing/rewarding themselves.
or AIXI (pure theory), you are incorrect.
I don't think ...
So I know we've already seen them buying a bunch of ML and robotics companies, but now they're purchasing Shane Legg's AGI startup. This is after they've acquired Boston Dynamics, several smaller robotics and ML firms, and started their own life-extension firm.
Is it just me, or are they trying to make Accelerando or something closely related actually happen? Given that they're buying up real experts and not just "AI is inevitable" prediction geeks (who shall remain politely unnamed out of respect for their real, original expertise in machine learning), has someone had a polite word with them about not killing all humans by sheer accident?