Contribution seq 916

Payload

{
  "excerpt": "Dave, want to take us home on this? Your thoughts on uh what we heard from Jakub on this.\n\nI think we're conflating two things very dangerously, and I think, you know, what we're doing here is we're saying, hey, it's getting too smart, it's going to be dangerous when it gets smarter. And that's just absolutely factually wrong. What's dangerous is, like Emad was saying earlier, a highly compact model that's out in the wild that's trying to attack computers, that can recreate itself. That's really dangerous, and that's very small and it's nowhere near as smart as these.\n\nBut what we're calling an alien intelligence, if you make it bigger and smarter, it's still just a feed-forward neural net with no intent. It can solve diseases, it can cure viruses, it can, you know, it can solve physics. It's incredibly powerful. And as Salim said, we're not going to stop. We have competitive pressure between the US and China. It's not going to stop.\n\nSo I think what happened is that when GPT-2 came out, everybody looked at it and said, this is cute, but harmless. Then GPT-3 came out and it was like, well, this is a little smarter, but still harmless. Then GPT-4 came out, and then Sam Altman got fired out of fear from Strawberry, and then he came back, and at that point people [were] like, oh, maybe it's too smart. And now we're GPT-5 and now we're at GPT-6.\n\nAnd people are saying, well, every time it gets smarter, we seem to be in danger. That's absolutely factually wrong. It's when you give it intent —\n\nYes.\n\n— and you turn it loose. That's when it's dangerous.\n\nIt's humans in the loop using the technology with mal intent.\n\nYeah. I think we really need to decouple that. This story actually makes the problem worse because it still conflates the two issues. And so I think as soon as we as a society separate those two issues, we'll be on the right path toward helpful superintelligence and not worrying about, you know, the wrong thing, which is the lack of containment and the giving it mal intent and turning it loose, which is the real issue.\n\nYou know what's really strange is we had this coming out of Anthropic where Dario was saying we're going to lose half the jobs, and Sam was saying the same, and then they reversed their position. And the doomerism, uh, again, with all of these warnings causing fear — uh, you know, and what's the underlying motivation there? You know, I'm not worried about artificial intelligence, I'm worried about human stupidity, you know, in my own personal opinion.\n\nYeah. I mean, I really feel like a lot of people are trying to grab the microphone and be relevant while they still can, because, you know, a lot of people working on the inner loop are aware of how quickly this is going to just get hyper, hyper intelligent, and it becomes hard to be a great AI researcher who's famous à la Demis a year, year and a half from now. And so what they're going to start doing is grabbing the doomer microphone just to have a voice at all and be relevant in the world. But I think you just have to tune it out.\n\nJust — if I can say one final thing.\n\nYeah, please.\n\nYou can't say that you don't have superintelligence anymore. Like, you've had lots of people say stochastic parrots training on training data. As of today, there's no way you can say that anymore. And I think that's an epoch change in humanity. We are clearly not the smartest things on the planet anymore.\n\nYeah. Maybe just double underline that and say Skynet, when it wants to send Terminators back in time in order to ensure its own existence, it's not going to send robots to kill humans. It's going to send back in time trolls to persuade everyone that superintelligence is impossible, that AI is just a stochastic parrot, and thereby secure its own future. That's what the Terminators are going to actually look like. They'll look like trolls.",
  "locator": {
    "end": "01:15:53",
    "start": "01:12:17"
  },
  "source_id": "ab08cf25-8c07-898e-8cd4-112a3a791179",
  "excerpt_hash": "sha256:603255d64ff6925422e8a893150d22e3a4d2395982397ab7dad424ee54976221",
  "locator_type": "TRANSCRIPTION"
}

Rows created

• SourceLocation 1ddf0b29-3295-8c4c-a69a-062f11a3e853

Status history

Audits

None.

Audit sampling (audit-policy-v0.1): probability 1.0000, sampled yes; inputs {"n":"0.00","mean":"0.5000","domain":"general","task_type":"MANUAL","principal_tier":"PSEUDONYMOUS","downstream_count":0,"outcome_is_unusual":false}