Dwarkesh is the Chad version of Lex

@theodoreshachtman7360

2 ай бұрын

Funny as hell dog

It’s so refreshing to see a leader in this space be so honest, humble, and pragmatic. We need more people like him and less enigmatic elitism from MBAs for such an important and transformational technology.

Have you considered creating a separate channel for your clips? Having to ignore most of your video notifications can't be helping the algorithm.

@theodoreshachtman7360

2 ай бұрын

Def do this

Dont forget "What you try to control, controls you"

if you give the objective function to prevent car accidents it will intentionally cause cars to crash, so there are less cars that will crash by accident.

@mstar7520

2 ай бұрын

what if the objective accounts for many different values, and harm to humans has significant negative value, and THEN we ask it to prevent car accidents?

@Dan-dy8zp

2 ай бұрын

@@mstar7520 What if you find out after you turn it on that you didn't account for all the values, but it's too late? Trying to think of all our values sounds like a dangerous plan because we'll miss something. We need another way.

@mstar7520

2 ай бұрын

@@Dan-dy8zp we don't neccessarily have to program in the values ourselves. ChatGPT and similar AIs have learned these things for themselves by poring through a huge amount of text. Try testing if chatGPT is missing any knowledge of human values. I'd guess probably not.

@dogecoinx3093

2 ай бұрын

increasing entropy is a good objective function. it will try to learn in order to satisfy this requirement. So it doesn't do stupid shit@@mstar7520

"I don't think we're doomed by default and have some problem we have to solve . . . " - What?

its so over for both of them

@Dan-dy8zp

2 ай бұрын

For all of us?

@Sky-fk5tl

2 ай бұрын

@@Dan-dy8zp ?

How about... don't make the thing? no? ok : (

@mygirldarby

2 ай бұрын

Automobile accidents have killed over one million people. Should we not have invented cars? Had we, America, not invented them, or forbade inventors from creating automobiles, would it have saved a million lives? Once the idea of the automobile existed, our government couldn't have banned their existence because other countries were going to build them even if we banned them. So America would be sitting there with no cars, and no infrastructure, while other countries happily built cars and made roads and we would fall far behind. We would get poorer as other countries that took our idea of the automobile grew rich on cars. How stupid would America be to have banned cars!

@Dan-dy8zp

2 ай бұрын

Yeah. That's the only safe way.

You do realize alignment is impossible right? You simply cannot control an entity orders of magnitude smarter than us. Just like how ants cannot control us. We are the ants to AGI.

@drhxa

2 ай бұрын

Wrong, as we build more powerful AI, we will build more powerful control methods. To have useful AI, it must be controllable, hence building good/strong AI pushes us toward building a more controllable AI. RLHF and constitutional AI are sufficient to control today's LLMs. As Dario said, we'll have better methods as we create the new models

@gorgeouswilderness

2 ай бұрын

If we can control the intelligence that is superior than human intelligence then how can that intelligence be more superior. People who wants to build intelligence will have illusion that they can solve this problem. Otherwise their fav project will be paused.

@andersfant4997

2 ай бұрын

​@@drhxaAnd how can you be sure🙈? We "might" find a way but its not plausible because today we dont have a clue. Our best hope could be that we never reach AGI.

@hannesthurnherr7478

2 ай бұрын

That's why it's called alignment, dork. You can't controll it so you need to make sure the thing wants the same as you, in which case you don't need to controll it because it's aligned with you.

@Linda-fr4dj

2 ай бұрын

@@drhxawhat are you thinking of? Controlling a clever new dog you just adopted? This is AGI … it will outsmart everyone.. and you first