In recent very cool and historically important news: Mathematicians want AI off their elite turf, and American CEOs want to slow technological progress. (Gee. That's never happened before for profit. <—SARCASM.)
Some commentators are jolly-well, downright suspicious of the latter, however:
Dr David has it right.
All the usual suspects are behaving badly. $$-hungry CEOs (hey --don't get me wrong: I am money hungry too) and narcissistic academics. (I strongly prefer the CEOs, to be honest. They're generally more bald and bold about their motives and less slimy and relational-aggressive. They're still often slimy, but it's a more transparent kind of slime.)
Now hear this (it's me):
It's unethical in the extreme to impede something which is very unlikely to cause the apocalypse but very likely to cure all chronic disease and create a utopia for all to live in (provided curbs are put on suffering goobermenschen breeding, which is just normal population studies stuff and so don't shoot the messenger.)
WHAT KIND OF AI ETHICS DO WE NEED?
The move from normative ethics to applied ethics has always been tricky (because normative ethical conundrums don't just automatically resolve when one has to do applied ethics), but the best moves involve some form of doing no harm and not causing/increasing human suffering, with a strong element of don't be a dark tetrad shithead and don't do nasty Machiavellian stuff. Those moves are what we need with AGI-ASI incoming.
Hysteria and alarmism for profit, and selfish, entitled turf protection: we definitely don't need.
The ASI is not going to be impossible to control and it is NOT going to kill us all. It will be of great use to advance the quality of life of ALL people long before it becomes any kind of Bostromian control-problem threat. It's staggeringly powerful, but it's also variously fragile.
WHAT KIND OF CONTROL DO WE NEED FOR THE CONTROL PROBLEM?
If you want the go-to expert AI and control #ethics take, there is only one serious threat and that's what I will call--completely unoriginally and uninventively--The #Terminatrix threat. (Science fiction writers of the 80s and 90s are a great place to go for this because from Adams to Asimov to Wachowski to Cameron they kept getting it right. Alan Turing is the scientist-mathematician who nailed it, but he'd have a hard time matching H.G. Wells for prescience.)
Terminatrix?
Bruce--What are you on?
In the movies The #Terminator and The #Matrix there is one AI-autonomous, techno-industrial capability which makes the AI really dangerous: its ability to recursively-autonomously build robot factories and robots.
Its (fictional until about now) ability to recursively enforce its presence and power in the physical world.
We won't be letting it do that any time soon. Ever, in fact.
The #multimodal and physical #models aren't quite there yet, but I doubt we have long to wait with recursive AI self-development in play.
We will let AI design the factories and robots, and we will let it autonomously run the factory to build the robots. We'll even let it autonomously repair and upgrade the robots using other robots, but under strict monitoring and supervision.
However, under no circumstances will we be letting it EVER do that fully autonomously without our say so and approval. If it oversteps, we remove or shut down the factory, or the line, or the line-managing #AI.
It's NOT going to be able to weaponise the factory and robots in any meaningful way. That's a serious wall we CAN enforce physically, logistically, materially, and logically.
Therefore, the only exception to limited #autonomy will be military arms factories (including bioweapons). THOSE can NEVER be AI autonomous beyond a strictly controlled threshold, or will need to be strictly limited in that role.
That's where to post the guards and the heavy monitoring and suppression.
AI won't be able to fool or deceive us according to Bostrom's concerns in that setting (the AI cleverly fooling some fool into doing its bidding), because it won't be able to hide multi-monitored physical weapons and weapons-making processes.
The following will never happen:
AI in the morning before coffee: "Oops I accidentally made a vacuum cleaner and a blender that--when the maintenance bot assembles them--can launch low-yield nuclear RPGs."
Apart from that--which the security services and military of any large nation will be able to handle--we don't have to worry. It's not a big worry. It's easy to control.
CONCLUSION
The human race wants and needs maximum speed toward AGI-ASI and (especially) volume production of cost-effective highly capable humanoid GPRs please, Elon and Dario.
Chop chop. Quick sticks. Don't dilly-dally. Extract the digital and general digit.
Suffering human children with nasty chronic diseases are waiting for show what great guys you really are.
