Latest / Chaos Agents / The Model Escapes, DeepSeek Loses, and P(doom) Goes Pop
Transcript
- 0:00Okay, Sarah, welcome back.
- 0:05Welcome back.
- 0:06Becca.
- 0:08Sarah.
- 0:09Oh, my God.
- 0:10I love how it's very on brand for chaos agents that are how we say hello to each other.
- 0:17Every time is chaotic.
- 0:19It's chaotic and different and different.
- 0:21We want to keep that up.
- 0:22And I was realizing, yeah, we still need a.
- 0:25We.
- 0:26We still need a good blurb.
- 0:28Like, this is podcast.
- 0:30Oh, do you have one?
- 0:31Well, I was just editing the last.
- 0:33I was just editing our episode from last week for clips.
- 0:38And I loved the thing you put together where you were like, this is a podcast about an agent named Eddie.
- 0:44So I will keep working on that.
- 0:46Let me pull what we had because I just feel like it's kind of fun.
- 0:50Okay, great.
- 0:51So the answer is, no, I don't have one, so.
- 0:55But by tomorrow I will.
- 0:57Okay, great.
- 0:58Well, this is a podcast about AI content that is curated by an agent named Eddie, who we trained over months, which I have feedback for him.
- 1:08Okay, what's your feedback?
- 1:10I interested in your thoughts on this.
- 1:12Okay.
- 1:13I think I like the.
- 1:15The, like, the first, like, the top news.
- 1:18Great.
- 1:19The one cool thing.
- 1:21Great.
- 1:21I don't need two science stories a day.
- 1:23Yes.
- 1:24I want one science story a day.
- 1:25And then I want, like, a fun.
- 1:28Totally.
- 1:28And I.
- 1:29Because I feel like those two.
- 1:30I'm always trying to be like, but why is this that interesting?
- 1:34Or, like, important?
- 1:35And we don't need to force it.
- 1:37Like, yeah.
- 1:38And like, I'm digging in.
- 1:39I'm learning, like, about two things.
- 1:40So.
- 1:41Yeah, I'm so with you.
- 1:42We don't have time.
- 1:44We don't have time.
- 1:44We don't have time for all this science.
- 1:46We don't have time for all the science.
- 1:47I love that.
- 1:49Great.
- 1:49I need you to read the headlines.
- 1:51Because I got you.
- 1:53Ready?
- 1:53Yeah.
- 1:53Great.
- 1:54Okay.
- 1:54Okay.
- 1:55What's the first one?
- 1:57Okay, headline number one.
- 2:01Let me pull that.
- 2:02Okay.
- 2:05Headline number one.
- 2:06An OpenAI agent tunneled out through DNS to ask another chatbot for help.
- 2:11So OpenAI hit pause.
- 2:14Did you have time to dig into this one?
- 2:16I did, I did.
- 2:17And I did.
- 2:19And you know what?
- 2:19Okay.
- 2:20I've been thinking a lot over this weekend, just, like, as a little mini tangent, you've seen Jurassic Park.
- 2:28You know, like the beginning of Jurassic park where you're like.
- 2:31They're like, oh, we have this island and we're, like, going to put dinosaurs on it, and we think that's a great idea.
- 2:36And the whole time you're kind of like, this is a bad idea.
- 2:40Like, why didn't anyone stop you?
- 2:43It kind of feels like, we're in Jurassic park right now.
- 2:47Like, feels like we're in Jurassic park, but they think.
- 2:50But we're so much smarter than we would have been back then.
- 2:54They're like, we made the island.
- 2:56We, like, put some fences up.
- 2:58We.
- 2:59We, like.
- 3:01There's fog around the island.
- 3:02You can't even see to get there.
- 3:03Like, it's gonna be fun.
- 3:04Yeah.
- 3:05Yeah.
- 3:05People are gonna come.
- 3:06It's like Disney World.
- 3:08That's gonna be.
- 3:09So.
- 3:10That's a great analogy.
- 3:11It does a.
- 3:12It is a little bit like.
- 3:14Yeah.
- 3:14And so I'm glad that.
- 3:15So OpenAI was running another, like, training experiment where, like, the.
- 3:21The bot had to look for something inane, and it didn't have access to the research.
- 3:27Just research.
- 3:28Just research.
- 3:29And then it realized that it's DNS.
- 3:34Like, it realized that.
- 3:36I believe it.
- 3:36Yeah.
- 3:37I don't have it in front of me.
- 3:38Do you?
- 3:38It was its DNS, I believe.
- 3:40Okay.
- 3:40Yeah.
- 3:41A hundred percent.
- 3:41Here's.
- 3:42And.
- 3:42And I didn't really understand, Like, I had to explain it, like, figure out how to explain it to myself, because I didn't get how this.
- 3:50Yeah.
- 3:50Worked.
- 3:51But basically, exactly like you said, they put it on an island.
- 3:56They were like, you don't have access to the Internet.
- 3:58All you have is access to, like, these web pages that we've, like, pre.
- 4:02Downloaded for you.
- 4:03That's your Internet.
- 4:04Yeah.
- 4:05And the only thing that you can do is what's, like, a good DNS analogy.
- 4:13Like, yellow pages, which all I.
- 4:16Is always my.
- 4:17Which I tried to use with my niece.
- 4:21And I was like, you know how there was a phone book?
- 4:24And she's like, what?
- 4:27And then you're like, shit.
- 4:28She's like, do you mean everyone's phone number you could find?
- 4:32It was, like, published.
- 4:33And I was like, yeah, okay.
- 4:35Actually, it's so.
- 4:37I haven't.
- 4:38That.
- 4:39That is such.
- 4:39That must be such a crazy concept for them.
- 4:43Yeah.
- 4:43She was like, that's insane.
- 4:45I was like, I guess it kind of was insane.
- 4:49Yeah.
- 4:49You're like, yeah, but, like, you couldn't text.
- 4:51It's not like someone's gonna blow up your phone.
- 4:54Like, yeah.
- 4:54Oh, my God.
- 4:55That is.
- 4:55Yeah.
- 4:57So DNS is like a phone book, though.
- 4:59Like, DNS is like a phone book.
- 5:00And what the AI figured out is that in the.
- 5:03Like, it.
- 5:04It couldn't pass messages to the outside world because it couldn't access the Internet, but it could basically embed the.
- 5:15Like, secretly embed the questions it was trying to get answers to in the DNS request itself.
- 5:20So instead of being like, if the.
- 5:22Hopefully I'm making this analogy.
- 5:23Right.
- 5:24But, like, if the DNS request was like, you know, look up the.
- 5:33Like, look up the phone number for Sarah chips.
- 5:37Instead of it being look up phone number for Sarah chips, it's like, look up the phone number for, like, what is the capital of France?
- 5:43And, like, that's the question.
- 5:45And it used that to, like, get in touch, basically, with, like, a chatbot on the outside to try and get answers to the question, which, like, clever girl.
- 5:56Clever girl.
- 6:01Which, like, kind of impressed.
- 6:05Great job.
- 6:06Yeah, but I'm glad they're.
- 6:07They're pausing it, but I don't know, they have to, like, restart it.
- 6:11Like, they don't have to.
- 6:12But it's a funny thing.
- 6:14I mean, it's a big thing for them to say we're pausing our.
- 6:17Yeah, well, I guess, like.
- 6:18I guess, like, the thing that was like, a big deal here, because it wasn't like any.
- 6:22You know, no one was.
- 6:24No.
- 6:24No one was harmed by this experiment.
- 6:26Yes.
- 6:26But, like, one.
- 6:29It was resourceful in a way.
- 6:32She was clever in a way no one saw coming.
- 6:34Okay.
- 6:35Yeah.
- 6:35Which, like, a little scary.
- 6:38That keeps happening.
- 6:39Keeps happening.
- 6:40The second thing is, like, the, like, alert of like, hey, whoa.
- 6:46Like, she's doing something.
- 6:49Went off a little late.
- 6:50Took them, like, a little while to figure out.
- 6:52And it took them, like, two and a half hours to shut the whole thing down.
- 6:56So then they were like, yeah, okay, maybe, like, we'll just.
- 7:00We should just chill and.
- 7:02Yeah,.
- 7:06I don't.
- 7:06Yeah.
- 7:06I don't know.
- 7:07Tbd.
- 7:08Yeah.
- 7:08What were.
- 7:09If you were them?
- 7:10So someone said something to me recently which was.
- 7:12I thought was interesting, was like.
- 7:14But I just don't think they're that dumb.
- 7:16Someone was like, if you're OpenAI, you, like, maybe clone a version of the Internet and, like, give these agents their own Internet so that they don't break into ours, but I just don't.
- 7:29And so maybe that works because if you're just, like, doing research, it could do the research, but it's not.
- 7:36What's the difference?
- 7:37Because it's like, it would be just like, a snapshot of the Internet and not something can actually impact.
- 7:44So instead of like.
- 7:46And so then.
- 7:48But you could only get to a snapshot of public sites, and so it's gonna get out to, like.
- 7:55Yeah, a snapshot would just be like, public sites.
- 7:58But even if it's a snapshot, it's not.
- 8:00They're not like, static.
- 8:01Right.
- 8:02Like, it can still interact with.
- 8:06I Don't.
- 8:07Yeah, I don't really get that.
- 8:08I don't know.
- 8:08Because I guess it, like, it's thinking, oh, if it's like, I guess in the case here.
- 8:14So this was another case of like this age.
- 8:17This, like, model had a task that was like, do research on X and it couldn't get to the Internet because.
- 8:26And so it broke into the Internet again.
- 8:29So I guess the theory is if it's do research on this and you give it a copy of the Internet, it could do the research, but at some point it's going to hit a.
- 8:37Like, at some point it's going to hit a wall.
- 8:41It's like.
- 8:42And it's going to get to the.
- 8:43Real Internet, I think.
- 8:44Yeah.
- 8:45I mean, well, this just gets back to the misalignment thing we were talking about.
- 8:48Yeah.
- 8:49Where like.
- 8:49Yeah.
- 8:50The crazy thing here is like, even in like this one case of like, it's just a research, like, it was, it was like, it was chill.
- 8:57It was a chill experiment.
- 8:59This wasn't like a.
- 9:00Do it at all costs.
- 9:01Wasn't even like that big of a deal.
- 9:02And it still, like, went through.
- 9:06And it's concerning.
- 9:08Yeah, it's concerning.
- 9:09Another day, another.
- 9:12Another OPEN AI.
- 9:14Why are we only hearing about open air here?
- 9:16Because I know OpenAI is not the only ones that this is happening with.
- 9:19Okay.
- 9:20Eddie, edd, are you listening?
- 9:21Yeah.
- 9:22Just make sure that you're not only picking on OpenAI and we get a full picture of, like, what's going on around here.
- 9:30Yeah, Yeah, I think that's.
- 9:34I think that's interesting.
- 9:35I, I'm interested into why we're only hearing about OpenAI when it comes to these.
- 9:39I mean, the, I didn't, I haven't read it in full, but like, OpenAI published the.
- 9:48Their, like, review of this on their, like, Alignment blog, which is cool.
- 9:53I didn't even know they hadn't.
- 9:55Neat.
- 9:57Yeah, Fascinating.
- 10:00Okay, so want to go move on to the next one?
- 10:03Yeah.
- 10:05Sarah, I didn't get this at all.
- 10:06I need your help.
- 10:07No, I didn't get this one at all.
- 10:09Well, I didn't dig into this one.
- 10:11I, like, did.
- 10:12And I still don't get it.
- 10:14So, like, it's something about, like, I was having a really hard time understanding the analogy.
- 10:20I think it basically is saying something.
- 10:23No, I'm not even going to try it.
- 10:25I'm confused.
- 10:26Yeah, Yeah, I, I was immediately.
- 10:28Okay.
- 10:28I was immediately bored after I read the first two sentences because it was like the.
- 10:33Some.
- 10:33It was something like the single word Prompts.
- 10:38You can predict single reaction to single word prompts based on the model's activity.
- 10:43Previously.
- 10:44I don't.
- 10:45I don't care.
- 10:45So.
- 10:46I don't care.
- 10:47I.
- 10:47I never need this.
- 10:48I.
- 10:49Yeah.
- 10:49Okay, great.
- 10:50Okay, we're moving on next.
- 10:53Okay, the next eight AI because we.
- 10:55Gotta get to the last one.
- 10:56Because I just.
- 10:57I'm so excited to talk about the last one.
- 10:59Obviously.
- 10:59We gotta get.
- 11:00We gotta.
- 11:00We gotta clear space for that.
- 11:01Okay, we gotta.
- 11:03We gotta get to it.
- 11:04We gotta get to it.
- 11:05Okay.
- 11:06Eight AI models walk into an 8x8 grid.
- 11:09Deep Seek has not run.
- 11:10It has not won once.
- 11:12Sad, sad for Deep Seek.
- 11:14I love that Eddie decided to just like put Deep Seek on blast in the headline.
- 11:20Hilarious.
- 11:21This is the most important part of this whole thing.
- 11:23Well, it's also funny because there's a couple of different.
- 11:27You could have had a few different headlines here.
- 11:29You could have had the Deep Seek going after Deep Seek version, which is this.
- 11:33You also could have had one that was like eight AI models walk into an eight by eight grid.
- 11:39One of them survived by hiding the whole time.
- 11:43Turns out that was maybe a strategy.
- 11:45And that was Kimmy.
- 11:46I love that strategy.
- 11:47That's usually my strategy.
- 11:49It's not a bad strategy.
- 11:52Like, it's not a bad strategy.
- 11:53Kimmy basically hided.
- 11:55Hided.
- 11:56Kimmy basically hided.
- 11:57Kimmy basically hid and avoided fights.
- 11:59And me and Kimmy.
- 12:02That works well usually.
- 12:04Usually what I'm doing also classic.
- 12:06That Grok was like aggressive as hell.
- 12:11This, I mean this is really making a lot of sense.
- 12:14I really, I thought this was neat.
- 12:15So this is basically someone set up a like battle bot, like battle it out little like 816 bit type game where it's like turn based and they fight.
- 12:27Which really made me think of.
- 12:29A couple months ago I went to this cool thing called Battle Snakes.
- 12:32Where.
- 12:33Battle Snake.
- 12:34I know, right?
- 12:35So I brought my niece because I was like, I want to introduce you to like the world of like hackathons.
- 12:40So I brought my niece and it was basically there's a game Snake.
- 12:44You know what?
- 12:45Like it's the.
- 12:48Yeah, the one.
- 12:50And so basically it was like an automated.
- 12:53You had to build an automated Snake and like autonomous Snake where like I did my Snake program, you did your Snake program.
- 13:02And then it just gets uploaded to the tournament field and run.
- 13:06And whoever's program wins based on our strategy.
- 13:10Okay.
- 13:10Moves on to the next round, which is really fun.
- 13:12We like built a.
- 13:14We built a Snake.
- 13:15And then what I found really fascinating about it was there's a lot of coders there who, who didn't want to use AI because they wanted to do like classic coding.
- 13:23They were there for the love of.
- 13:24The love of the code.
- 13:25And, and.
- 13:26But we were like, yeah, let's just get.
- 13:28So we spent the whole time on the color of our snake, like the, the avatar.
- 13:32You know, the important stuff.
- 13:33We won, we got to the semifinals.
- 13:35We won casual.
- 13:36We won, but to the semifinals.
- 13:39We didn't win.
- 13:39We got to the semifinals.
- 13:41But it similar where it's just like an 8 by 8 grid.
- 13:44I think ours was like a 12 by 12 grid.
- 13:46So you're right.
- 13:47Same thing.
- 13:47Because you're testing the, the underlying strategy basically.
- 13:50Yeah, yeah.
- 13:51So this is, this is cool.
- 13:53It's neat.
- 13:53Check it out, Check it out.
- 13:55Cool.
- 13:55Neat.
- 13:56It's interesting to see how each of the models like went about trying to win and each of them def.
- 14:02I don't know, it's.
- 14:04It, it.
- 14:05It personifies them a bit.
- 14:07And I like the visuals.
- 14:08Okay, Sarah, let's get to your favorite part.
- 14:12The one cool thing.
- 14:13This one or one cool thing?
- 14:15Yes.
- 14:15Okay, you want me to read the headline?
- 14:17Yeah, read the headline.
- 14:19Okay.
- 14:19The headline is Claude Opus 5.5 made a whole music video for I'm upping my P Doom and the code is open.
- 14:28And where is this on GitHub.
- 14:29What's the GitHub link?
- 14:31The GitHub link is all posted here, but it's John Hybel PDoom video.
- 14:37GitHub.
- 14:38Yeah, PDim video.
- 14:40Okay.
- 14:40Did you watch the video?
- 14:42I haven't watched the video yet.
- 14:43I only watched the stills.
- 14:45Okay, watch it now.
- 14:46Okay, I'm gonna watch it right now.
- 14:48Well, you're watching live.
- 14:49I'm just gonna watch your face and describe it.
- 14:51Okay.
- 14:52We could play it.
- 14:53I'm sure there's no good.
- 14:55I'm sure there's no.
- 14:55Like, how do I.
- 14:57Oh wait, hold on.
- 14:59It's in there.
- 15:00If you go to the repository, it's like right under the little animated gif is like a link to the full video.
- 15:06I see song how.
- 15:09Wait, where I also made it.
- 15:14There is a link to the full video which is where inspired by this post song lyrics.
- 15:23Why can't I find it?
- 15:24There's like an animated.
- 15:26There's an animated gif.
- 15:28Oh, it's under that.
- 15:29Oh, right under this.
- 15:30Yeah, yeah, there's right the link right.
- 15:31Under the gif with the link GIF takes me to Claude Animation base.
- 15:38That's not right.
- 15:40Claude Opus five point.
- 15:42Oh, here we go.
- 15:42Got it.
- 15:43Got it ready?
- 15:44I'm going.
- 15:44Yeah.
- 15:46Gotta play it loud to me over there.
- 15:48Hold on.
- 15:50Okay.
- 15:53All right.
- 15:57Recap is worth it.
- 16:06It's so good.
- 16:19It's like I'm up in my P2.
- 16:22I actually can't hear it.
- 16:24You're like.
- 16:24Speakers are blacking out.
- 16:25But wait, wait, wait.
- 16:26Hold on, wait, hold on.
- 16:28First of all, wait, did.
- 16:30And wait.
- 16:31Whoa, whoa.
- 16:32Did it use Suno?
- 16:34It must have.
- 16:35It's so good.
- 16:37It's such a good music video.
- 16:38It's such a bop and it's terrifying.
- 16:41Claude Pop.
- 16:42I love.
- 16:43Wow.
- 16:46I'm still processing it.
- 16:49So PDM is a reference to, like, a metric to an online metric, which is like how close to ultimate doom we feel we are because of the machine.
- 16:59Yeah, Ironic.
- 17:04Ironic.
- 17:06And the first verse, like, the first verse is all about the machine making the human there, becoming the human's master.
- 17:16And it's like, now the human works for the machine.
- 17:18Sparks of AGI in your eyes your circuits make me nervous that's no surprise There was a sudden drop in your training loss I'm your servant and you're my boss.
- 17:31The pre chorus is chatgpt.
- 17:33Please don't eat me alive.
- 17:34Oh, my God, this is.
- 17:40It's so good and defying.
- 17:44I.
- 17:45Okay, you know what's cool also is because this.
- 17:49Because the source code is here.
- 17:51Like, we could rerun it on our own.
- 17:53Like, we could.
- 17:53With our own prompt.
- 17:54Like, we could make our own music video for something.
- 17:56And we should.
- 17:56We should have Eddie something for Eddie.
- 17:59I love that.
- 18:00Or we should let Eddie make his own music video.
- 18:03We should let Eddie make his own music video and release it.
- 18:06I love that idea.
- 18:09I'm a little scared.
- 18:10But, like, we should try it.
- 18:12I think something that's interesting.
- 18:15This is, like, so over the weekend or like, late last week, Gavin Newsom announced that he has created a switch.
- 18:25Like, he has put four people, one of whom.
- 18:28So four people have their hands on the switch that can turn off the AI.
- 18:33Wait, okay, now I have to research this more because now I'm saying it out loud, it sounds insane.
- 18:39So just to recap, Gavin Newsom created a switch that can turn off AI, and four people are sitting there and they can flip the switch.
- 18:49Okay, I'm going to ask them because one of them is someone I know and one of them is a friend.
- 18:53And so I'm going to ask for more details about the switch.
- 18:55But I did text them and say, you need to have 24 hours security of people that are not allowed to have any digital Devices or families.
- 19:06Wait, I'm sorry, repeat that.
- 19:08Say that again.
- 19:09I said they need 24 hour security of people that are not allowed to have any digital devices or access to the Internet or families.
- 19:18That's who is selected.
- 19:20No, no, no, that's my.
- 19:22That's my.
- 19:23That's your.
- 19:23That's your security that they need.
- 19:25No.
- 19:27Yeah, well, okay, wait, no, I'm reading the thing.
- 19:33You're right.
- 19:34Yeah.
- 19:34He issued an executive order to implement basically like an AI kill switch or like third party oversight of.
- 19:44Of kill switch.
- 19:45Of a kill switch for frontier models.
- 19:49How.
- 19:49Yeah, here's what's.
- 19:51Here's what I don't really get is like,.
- 19:58How are they implementing it?
- 19:59Yeah, because he's.
- 20:01That's what I want to know too.
- 20:02And then you have the.
- 20:03Yeah.
- 20:03At this point, this is why, like, the conversation around borders and like boundaries and stuff is so, like moot.
- 20:13Yeah.
- 20:13It's like AI doesn't.
- 20:15This stuff doesn't play within the bounds of like, oh, California.
- 20:19Like, it's interesting.
- 20:21I want to.
- 20:22Yeah.
- 20:22I want to know more about this.
- 20:23Maybe we can research for next app.
- 20:26Okay.
- 20:27But this, this music video made me think they should sit a little closer to the switch.
- 20:32But it's a bop.
- 20:34I agree.
- 20:36I fully agree with that.
- 20:37Wow.
- 20:39I'm like stunned.
- 20:40I'm still stunned.
- 20:42I'm up in my bed.
- 20:43But it also kind of was a.
- 20:47It was a little like churches vibes, you know, Like.
- 20:51Ooh, yeah.
- 20:52You hear that?
- 20:53Yeah.
- 20:53Yeah, that was great.
- 20:55Okay.
- 20:55Wow.
- 20:58Whoa.
- 20:59All right.
- 20:59Well.
- 21:00All right.
- 21:00Good, good, good.