AI can develop abilities its builders did not explicitly teach

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 6:53

   because it's not safe what's about to happen. Why is AI distinct from other kinds of technologies? Well, let's get to that. So I think the thing that is most difficult for people to get is up until now, technology progressed in a very, like, we're kind of adding layers to a stack kind of way. Like we build the networking stack, we build the user interface stack. And as you develop the stack, you're kind of just adding layers and layers and layers. And the technology that we live in was coded manually, like line by line, like when the computer sees this, do this. When the computer sees this, do this. And then people contribute all this code over 30.

2. 7:30

   40, 50 years on GitHub and in operating systems, and then you land in this technological world in which everything that happens in a computer is happening through logic and through human choice. What makes AI different is that you're designing and you're not really coding it like, I want it to do this. You're more like growing this digital brain that's trained on the entire internet. And when you grow the digital brain, you don't know what it's capable of or what it's going to do so think about this way like if i did a brain scan in your brain could i know from just the brain scan what you're capable of no i can see that this part of your brain lights up when you have that thought but i can't have a comprehensive picture of like what is everything that chris is capable of can he do sociopathic manipulation and do better military strategy than the best

3. 8:17

   u.s generals like i from the brain scan i can't tell that maybe you can but so with ai we are essentially you know when people hear about these huge data centers getting built out like facebook's building one or meta's building one the size of manhattan um and you ask like what is that what's going on there it's like they're building a bigger and bigger digital brain that that's what goes from gpt3 to gpt4 you know with more neurons when you hear the number of parameters of an ai model that's like essentially the number of neurons in an ai model and what they found is that the more um gpus and NVIDIA chips you point at at sort of growing this digital brain the more intelligent it gets and the more it picks up capabilities that we didn't intentionally teach it like there was a famous example where you just train it on the internet and then um you

4. 9:04

   know it's answering questions in english and suddenly it learns how to answer questions in farsi like doing you know in q a in in a different language and no one taught it that language it just sort of learned that on its own um and that's what's like weird about ai is that it's a black box. We don't really understand how it works. And yet we're making it more powerful, much faster than understanding how it works. And that's what leads it to make these more unexpected behaviors that we aren't able to control. And I think we're going to get into some of those.

5. 9:34

   A data center the size of Manhattan? Yes. Where? I don't remember where that one is, but it's crazy. There's like an overlay. Someone can look it up. There's like an overlay where you can see the size of this data center, and it's almost the size of Manhattan. And you can ask, I mean, again, there's more money people should just get there's trillions of dollars going into this there's more money going into this technology than all technologies of the past have ever been built and we're releasing this technology faster then we've released every other technology in history.

6. 10:02

   It took something like two years for Instagram to go from zero users to 100 million users. And it took two months to go from zero to 100 million users for ChatGPT. And of course, they're going from ChatGPT 3 or 4 to now at 5.2. And it went from barely being able to finish a sentence with ChatGPT 2, like finish a paragraph and do like a coherent text, to GPT-3 could write full essays, to GPT-4 can pass the, you know, the bar exam or the MCATs, to GPT 5.2, I believe, was used to get a gold in the Math Olympiad.

7. 10:37

   Meta's Hyperion AI data center will sprawl to four times the size of Manhattan's Central Park. And there are quotes from people like inside of OpenAI who believe that they're not just building this like narrow technology that's a helpful blinking cursor. They want to build artificial general intelligence and so what that means is being able to do that everything that a human mind can do and the joke inside the company is like we're going to cover the world in data centers and solar panels like they want to cover the world in essentially these big boxes that have huge clusters of NVIDIA chips that then compute away and ultimately create something like a super intelligent god entity that they believe that they will use to own the world economy make trillions of dollars and from a kind of ego religious intuition they will have built the god that

8. 11:27

   supersedes and replaces humanity i know that sounds insane so let's we can slow that down