Egor Riabov

Math freak with trace amounts of musician

[Egor Riabov @imobulus]

@erminalt19216 @robbensinger @grok You have a habit of reading my statements as absolutism and appealing to impossibility of what is not even needed. Of course I meant that current AIs will continue to improve up until the mentioned threshold and there seems to be no limit up to that point.

[Egor Riabov @imobulus]

@erminalt19216 @robbensinger @grok There are different positions because “doomers” aren’t a single group. My position is that it’s inevitable once a certain capabilities threshold is reached. My position is also that the capabilities gains don’t seem to have any limits and only accelerate. The rest follows easy

[Egor Riabov @imobulus]

@erminalt19216 @robbensinger @grok NPPs aren't grown in such a way that no one knows how it works internally. For NPPs there a limited list of things that can go wrong. And the worst thing that can happen with NPP has already happened. If you so want to go on with this analogy, you must assume similar path with AI

[Egor Riabov @imobulus]

@erminalt19216 @robbensinger @grok So we're back to the point where we discuss whether a bad thing is about to happen to implement a preventive measure. I think that preventive measures that allow further advancements of AI at the current speed will not be effective against the bad thing I mean.

[Egor Riabov @imobulus]

@erminalt19216 @robbensinger @grok Regulations are growing "on demand", as you say, i.e. when a bad thing happens or we think it's about to happen then we regulate it. If you want to wait for a supervirus to actually get released to force a regulation, be my guest. Otherwise we must implement preventive measures.

[Egor Riabov @imobulus]

@erminalt19216 @robbensinger @grok I would even say that I am not so sure about how AI would go in a world where people behave as carefully as you describe. But the AI capabilities gains would certainly be slower in this world, and we don't live in this world so I didn't think much about it.

[Egor Riabov @imobulus]

@erminalt19216 @robbensinger @grok They certainly aren't as controlled as you want them to be. You have already told enough about your vision of AI that I can tell you almost for sure -- no, they are not controlled enough. And they will not be, without legislation, because controls interfere with speed heavily

[Egor Riabov @imobulus]

@erminalt19216 @robbensinger @grok Of course those labs are robotic and automated. Did you expect AI companies to slow down their results just to have humans check every single step? The robots don't sleep and work without breaks, they do science 10 times faster, and it's already going well, why stop?

[Egor Riabov @imobulus]

@erminalt19216 @robbensinger @grok I mean "progress" just as a "a good thing". They believe their plans for checking that AI behaves are adequate. Did you know there already is a project for automated drug discovery that gives AI access to wet labs?

[Egor Riabov @imobulus]

@erminalt19216 @robbensinger @grok Fourth, there's no visible "button" or a core decision that is dangerous. It is just linear progress with an unknown point of no return, and AI companies don't know where that point is or don't eve believe it exists.

[Egor Riabov @imobulus]

@erminalt19216 @robbensinger @grok Third, AI companies exist solely because they are competition driven. They experience constant pressure from the market to come closer to pushing the button.

[Egor Riabov @imobulus]

@erminalt19216 @robbensinger @grok Whatever humans have not trusted AI to do (yet) AI can gain access itself, if it's sure enough it doesn't get caught. For example moving some subsystem of itself as a worm that uses distributed computing resources of the internet to think private thoghts on a self-rented machine

[Egor Riabov @imobulus]

@erminalt19216 @robbensinger @grok My idea assumes that AI will do everything by itself because humans will trust it to do so because it "seems aligned" and "all tests, including real world ones, show it's aligned", and AI just does everything faster than humans, so "if we don't do it with AI then competitors will

[Egor Riabov @imobulus]

@PugCommander @stantrien @RBehiel @leecronin When they later meet, in each configuration they agree on what they saw. Configurations don't interact with each other unless they are completely the same. And there is no action at distance there, nor any collapse happening.

[Egor Riabov @imobulus]

@PugCommander @stantrien @RBehiel @leecronin When you send a pair of entangled electrons to different receivers you put a world into a superposition of configurations <both particles spin up> and <both particles spin down>. When two receivers receive those there are two configurations, both of which received same spin

[Egor Riabov @imobulus]

@erminalt19216 @robbensinger @grok It doesn't matter, it talks to humans, humans assume its premises, the result is the same either way

[Egor Riabov @imobulus]

@erminalt19216 @robbensinger I don't want to bring the human factor into this, it already is there. I don't understand this conversation, you just keep responding with some new topic-to-discuss about this wide problem. Can we focus on one?