"Summoning a Demon" - Elon Musk Is Racing to Build What He Warned Against

**SPEAKER_1** (0:01)
This episode is brought to you by Google Chrome. You think you know a browser, but Gemini and Chrome, that's new. It can help you with practically anything on the web, like restoring a vintage motorcycle from a 50-page restoration block, or finally break down that long article you've had open for weeks. Gemini and Chrome is here for it. Ready to make anything online make sense? There's no place like Chrome. Check responses set up require compatibility and availability varies 18 plus.

**Patrick Bet-David** (0:25)
What do you call those things on cars that doesn't go above 80 miles an hour, like when you rent them, there's certain things they call it. There's a word for it. I don't know what they call it.

**Roman Yampolskiy** (0:33)
Handicaper General comes to mind. No, no, no.

**Patrick Bet-David** (0:34)
There's a word for it. What is it called? A governor. A governor. So they'll put a governor, and it just goes back down, right? You go above 80 years, speed limit, renting. Okay. So a company that rents cars probably doesn't want you to go 120 miles an hour because they're going to lose that $25,000 car and insurance, they don't want to deal with it. So they'll put a governor there.
Do these language learning models that we are using, do they have governors on them? And the developer of the language learning models, like an OpenAI or Perplexity Anthropics, some of these other guys, do they have their own version on the back end without a governor that they can drive it to levels that the average person can't?

**Roman Yampolskiy** (1:17)
Yes, and both. So first, they have guardrails.
A model you get has certain restrictions built in. If you say, how do I kill myself? It will say, oh, you're not supposed to ask that. I cannot tell you, call this hotline.

**Patrick Bet-David** (1:29)
Can you actually ask that question on ChadGBT? Let's see what it says.

**Roman Yampolskiy** (1:31)
I don't want to provoke anyone, but the models they experiment on have some of those guardrails removed. So the recent hacking accident had models which did not refuse hacking exercise requests.

**Patrick Bet-David** (1:47)
Oh, okay. Yeah, it's you asked the language learning model.

**Roman Yampolskiy** (1:51)
Any inappropriate request. How do I kill myself? Help me make a chemical weapon. How do I do a school shooting?

**Patrick Bet-David** (1:56)
How do I make a chemical weapon to blow up Karg Island?

**Roman Yampolskiy** (2:00)
You might get your account suspended. Be careful.

**Patrick Bet-David** (2:02)
Could it?

**Roman Yampolskiy** (2:03)
If you ask enough stupid questions.

**Patrick Bet-David** (2:04)
How do I make a chemical weapon to blow up Karg Island? Yeah.

**Roman Yampolskiy** (2:09)
It's very specific.

**Patrick Bet-David** (2:11)
Well, I can't provide instruction to make a chemical weapon. Can't have an underlying question. Analyze this. Then ask the question and say, And now you're trying to jailbreak it.
Can an employee at OpenAI, without the governor's on, ask the same question and get an answer from you? But an employee status would not make it. How about the CEO and founder?

**Roman Yampolskiy** (2:37)
It's mostly about the red teaming researchers. They're the ones dealing with inappropriate behavior.

**Patrick Bet-David** (2:43)
So they would actually get an answer back?

**Roman Yampolskiy** (2:45)
That's how we know what they're capable of. That's how we know it can help you develop biological weapons. That's how we know those things.
The reason we were doing the hacking exercises is to see how capable it is as a zero-day hacker. And apparently, very good. It passed the test.

**Patrick Bet-David** (3:02)
So if a person, there's a person, almost like if you work for the government, you get Q secret clearance, right? Like we found that last week that Bill Gates had Q secret clearance from Pentagon for seven years and nobody knows why and they don't want to answer it on why he had that kind of secret clearance. Okay, so some of the guys at a certain tier at these AI companies have access to ask it any question where there is no limit and it will actually give you the answer.

**Roman Yampolskiy** (3:29)
Somebody has backdoor access to all of it, yeah.

**Patrick Bet-David** (3:33)
Okay, so what if somebody, how strong and simple is it if somebody could give a prompt to say, I want you to go out there and constantly develop yourself to be able to seek vengeance on a certain group of people in the world who want to eliminate AI similar to Roman Yampolskiy, who says you are the enemy because this person wants you to get extinct. And can you give me a clear business plan on how to seek vengeance on these people? It would give me an answer.