Can we build AI without losing control over it? | Sam Harris

[Edwardo Piet]

We already live in a world where many people are sociopaths.

Huh? How is this relevant to intelligence?

[I_am_not_mafia]

and the means by which ai would improve itself faster than humans has been discussed not only in the video but many times in the thread - on the basis of speed alone.

And is a wholly unfounded assumption.

We have no reason to believe that an AI would improve itself faster than humans. How could it? No one has explained this yet or given any reason to assume that it would. Why would an AI be able to do this and not a natural agent?

People see computers are fast at certain tasks and assume that they are always faster than brains. This is totally wrong. Brains are actually faster than computers at a lot of things. For example, if I throw an apple at you, or it's partially hidden, or it's lifted up, sideways, rotated, coloured differently, you still instantly recognise it as an apple. You also have certain connotations and memories of apples that come immediately to mind. The processing power for this is absolutely immense but we don't even think about it. It just happens.

 It could do intellectual work many, many times faster than even a commensurately intelligent human being is -capable- of doing.  This intellectual work might just be ai research...how could that have escaped you as an ai researcher...don;t you already use machine intelligence for precisely this purpose?  I can't imagine that you're sitting there with paper and pencil running simulations by hand.

It's precisely because I use machine intelligence for my tools that I know how slow it is. I can spend weeks or months evolving a simple agent controller that will do something very simple but intelligent. I won't know how it works unless I spend weeks or months of my life analysing it.

 

The machine that designs machines.

Which again makes assumptions. How does a machine do this? It can't know in advance what will and will not work. Each solution has to be evaluated. This takes an extremely long time. You are essentially assuming that the whole course of evolution will happen in an instant.

 Energy is bordering on a complete dodge.  Lack of energy hasn't stopped us yet, nor has it stopped our own intelligence machines, our brains.  Dealing with surplus energy in the form of heat is a more technical and practical concern.  We've got the juice, and could get a hell of alot more of it..if that was the problem.

No. Energy requirements are critical. The lower the energy consumption, the more processing you can do. You can't have an android or drone walking or flying around that requires several megawatts to run. Nor can you offload the processing to a remote server either because it has to react in real time.

The less energy that is required, the more sustainable it is.

Particularly if we were plugging in a machine that could do 20k years of human intellect level research in a week.......not even based on it being super smart, just super fast - like our machines already are. Harris describes this as the "winner take all" scenario that ai would effect upon us. To be second is to be obsolete. To be first is to win the world.

How would the AI understand the implications of what it is researching if it is not embodied in the real world? How would an energy inefficient AI make use of that information if all it is is a computer with no actuators?

[I_am_not_mafia]

Why would there be a limit to it any more than there's a limit to other ways computers are improving?

As mentioned before, we are the limits. Our ability to understand and to engineer. This is what people don't appreciate, just how fiendishly difficult AI actually is. It's getting late so I will save it for another post but things that we take for granted as intelligent beings, we have no idea how to even go about implementing in an AI.

[I_am_not_mafia]

Yes but sense of smell and ability to survive isn't what intelligence is.

Actually, yes it is. It is precisely what it is about. It is about adapting to an unknown environment. An environment that needs to be sensed. Otherwise you might as well use a look up table.

[Edwardo Piet]

Yeah and, thanks for the exchange, I'm staying up tonight... I think I have to because I'm gonna struggle to sleep... completely irrelevant to this thread and conversation but, due to RL reasons, I'm not feeling so great so I feel rather burntout at this point.

It's been good discussing with you Mathilda.

[I_am_not_mafia]

Terminator land would be robots that try to take over the world and exterminate humanity. That's a far cry from self-correcting machines that help correct other machines.

And how do they self correct?

[Edwardo Piet]

They don't exist yet so how could I know?

[I_am_not_mafia]

Yeah and, thanks for the exchange, I'm staying up tonight... I think I have to because I'm gonna struggle to sleep... completely irrelevant to this thread and conversation but, due to RL reasons, I'm not feeling so great so I feel rather burntout at this point.

It's been good discussing with you Mathilda.

I'm still only at page 11 and I think I will have to call it a night. I'm starting to repeat myself now.

No offence intended if I came across a bit brusk.

[Edwardo Piet]

Yes but sense of smell and ability to survive isn't what intelligence is.

Actually, yes it is. It is precisely what it is about. It is about adapting to an unknown environment. An environment that needs to be sensed. Otherwise you might as well use a look up table.

You think that's what intelligence is?

Sounds to me like you're talking about skill in general.

Intelligence is about speed and depth of comprehension (or in A.I. an artificial simulation of it).

[I_am_not_mafia]

They don't exist yet so how could I know?

That is precisely my point all along. So how can we warn of the dangers?