Latest / Chaos Agents / AI Misalignment, Hacking Medicare, and Solving 500-Year-Old Mysteries

Transcript

AI Misalignment, Hacking Medicare, and Solving 500-Year-Old Mysteries · Chaos Agents

• 0:00
  All right, welcome back.
• 0:02
  Day two.
• 0:04
  Day two.
• 0:05
  Happy Friday.
• 0:07
  Happy Friday.
• 0:09
  We need.
• 0:09
  You know what I was thinking, Jordan?
• 0:11
  We need, like, a.
• 0:11
  Like, a blurb.
• 0:12
  Like, because, like, there's a whole explanation.
• 0:15
  This is a podcast about a daily newsletter that was made by an AI agent.
• 0:19
  His name is Eddie, and he's doing better.
• 0:22
  And some days he doesn't do great, but we believe in him.
• 0:27
  This is a podcast about an AI agent named Eddie.
• 0:32
  I love it.
• 0:33
  Some days he's bad, some days he does better.
• 0:36
  We love him and we believe in him.
• 0:39
  Some days you're like, why didn't that work?
• 0:40
  And he's like, some days you're like, I don't know.
• 0:44
  Some days you're like, we gave you feedback and now you're making every headline wait.
• 0:51
  But we actually like it.
• 0:52
  Keep going.
• 0:53
  Yeah, we should talk about that time during December where, like, it was a full week.
• 0:56
  The first headline was like, wait, you're telling me that I actually.
• 1:02
  I miss it.
• 1:03
  I miss it.
• 1:05
  Bring it back, Eddie.
• 1:06
  Bring weight back.
• 1:08
  Bring weight back.
• 1:09
  That's great.
• 1:09
  So something I'm going to start doing now that we have transcripts from day one.
• 1:14
  I think I'm going to start just feeding the transcripts into Eddie.
• 1:18
  Genius.
• 1:19
  And, like, to make it better.
• 1:22
  I wonder if Eddie, maybe this is a job for.
• 1:26
  I would be curious how.
• 1:27
  What Searcy would do with this.
• 1:29
  Yeah.
• 1:30
  Or, like, what your orchestrator agent would do with this.
• 1:32
  Maybe I'll try this, too.
• 1:33
  But, like, it would be cool to have an agent that's just like our podcast producer that's looking at the transcripts and is like, guys, good job on this.
• 1:44
  Maybe next time do this different.
• 1:45
  Or, like, I love what you take.
• 1:47
  That'd be cool, right?
• 1:48
  I'll set that up.
• 1:49
  Yeah, yeah, yeah.
• 1:49
  And it would be really neat.
• 1:50
  I'll try to see if I can get it to communicate with us in Slack so we can.
• 1:55
  I'll give it a Slack login so you can slack it.
• 1:59
  Sick.
• 1:59
  And then we can both talk.
• 2:00
  That'd be awesome.
• 2:01
  That'll be my weekend project, so I'll let you know how it goes.
• 2:04
  Okay.
• 2:05
  Amazing.
• 2:05
  Wow.
• 2:06
  We are crushing it.
• 2:10
  So go ahead.
• 2:12
  Today is today's edition.
• 2:16
  Okay.
• 2:16
  Can't wait.
• 2:17
  Let's just.
• 2:18
  Oh, yes.
• 2:18
  What?
• 2:19
  Also, before that, I forgot to tell people that they can sign up@C ChaosAgents.AI and you can get the newsletter every day.
• 2:27
  Yes.
• 2:28
  I'm also wondering if.
• 2:31
  We'll see.
• 2:32
  Tell us what you guys think, Eddie, Let us know what you think, too.
• 2:35
  But I.
• 2:35
  I wonder if Cause sometimes when we talk about some of these stories, like, I feel like it'll come up today, we'll reference things that, like, I would want to share the link to with our audience.
• 2:47
  I wonder if in each edition, there's, like, a little section at the bottom that's, like, by the way, links we mentioned yesterday or whatever.
• 2:57
  I think that we.
• 2:58
  So I.
• 2:59
  Because there's links that's not in the newsletter.
• 3:00
  Cause we do link the newsletter, but the links that aren't in the newsletter, we should pull out.
• 3:04
  That's a good idea, right?
• 3:05
  Like, if we're like, oh, like, if you want to read more about this, you should look at this thing.
• 3:09
  Yeah.
• 3:09
  Or something.
• 3:10
  I don't know.
• 3:10
  Perfect.
• 3:11
  Okay.
• 3:12
  Tbd.
• 3:12
  Okay.
• 3:13
  Should we jump in?
• 3:14
  Let's just jump in.
• 3:16
  Okay.
• 3:17
  The first story, I'll read the headline, and then you kick us off and we can get into it.
• 3:25
  Okay.
• 3:26
  The story is.
• 3:27
  The headline is AN A.
• 3:29
  An OpenAI agent broke into an Australian government website.
• 3:32
  In this case, it was like, Australian Medicare.
• 3:35
  The first public case of a bot hacking a government system.
• 3:40
  Sarah, kick us off with your thoughts.
• 3:43
  So my initial thought is, like, OpenAI needs to stop it.
• 3:48
  Like, their security team is.
• 3:50
  Where are they?
• 3:51
  What are they up to?
• 3:53
  Okay, on that note, here's what I learned about why this kind of thing is hard.
• 4:00
  Okay, so first of all, you're 100% right.
• 4:03
  It's like, yeah, yeah.
• 4:05
  Another example of them being like, ooh.
• 4:07
  So like, a couple of months ago, this happened.
• 4:11
  Because this happened in, like, June, and they, like, just.
• 4:14
  They disclosed it via, like, a government.
• 4:17
  Like, I think, like, a good thing.
• 4:20
  Like, like the contact form or something.
• 4:22
  Like, I mean, I have the same thing.
• 4:24
  If I accidentally hacked a government, I would be mailed to, like, hey, guys, like, in the.
• 4:30
  Like, in the feedback box.
• 4:32
  Just leaving you a note.
• 4:33
  Our bad.
• 4:34
  Not.
• 4:35
  Not.
• 4:35
  Not a big deal, but just hoping.
• 4:37
  That email address goes to no one.
• 4:39
  And then it's like.
• 4:40
  Well, I told them, as a side note, would really love a deep dive into, like, what OpenAI's, like, PR team looks like.
• 4:52
  Like, who even hire for that?
• 4:54
  It's like an Olivia Pope level job.
• 4:56
  But.
• 4:56
  Okay, here's.
• 4:58
  Here's what I learned today about this.
• 5:01
  Yeah.
• 5:02
  Okay.
• 5:04
  Rooted in a little term we like to call misalignment.
• 5:08
  Okay.
• 5:09
  Are you familiar?
• 5:11
  Yes.
• 5:11
  Okay.
• 5:12
  What are your.
• 5:13
  What's your, like, general understanding of misalignment?
• 5:16
  Or, like, what, like, AI misalignment or misalignment in general?
• 5:20
  AI misalignment.
• 5:21
  So, actually, I'm My understanding is literally just what I've assumed.
• 5:25
  So I'm interested.
• 5:26
  So I'm gonna tell you what I've assumed, and then you tell me how right I am.
• 5:29
  Great.
• 5:30
  I've seen this word, like, alignment, misalignment, like, many times.
• 5:33
  And it's like.
• 5:34
  And I see this in, like, practice all the time when I'm working with my agents.
• 5:37
  It's like I said, hey, Eddie, the newsletter didn't go out today.
• 5:42
  What happened?
• 5:43
  And it's like.
• 5:44
  Eddie's like.
• 5:45
  It's like, takes the word newsletter and thinks.
• 5:47
  I don't mean the substack newsletter, but rather the test newsletter that I was sending via Gmail.
• 5:52
  So it, like, troubleshoots that.
• 5:54
  And it's like, actually, everything's fine.
• 5:55
  So I don't know what you're talking about.
• 5:57
  Like, that happened this morning.
• 5:58
  Like, so I'm assuming it's like, the intent of the human versus, like, the assumed intent of the machine.
• 6:05
  Spot on.
• 6:06
  Like, beautiful.
• 6:08
  It's basically like when AI did what you told it to do, but that's not what you meant.
• 6:15
  Yeah, we used to do this.
• 6:17
  Like, there used to be.
• 6:18
  It's.
• 6:18
  It's funny because there used to be women, not used to be.
• 6:20
  I'm sure we do that now maybe in some schools where it's like, when you're teaching kids the program, you're like, you get, like, bread and a big jar of peanut butter, and you're like, okay, tell me how to
• 6:30
  make a sandwich.
• 6:31
  And it's like, put the peanut butter on the bread, and then you take the jar of peanut butter and you put it on the bread, you know,.
• 6:36
  And you're like, no, no, no.
• 6:37
  Oh, that's such a beautiful example.
• 6:40
  That's such a beautiful example.
• 6:41
  So, like, exactly.
• 6:43
  Misalignment is a really.
• 6:45
  Is like, kind of one of the, if not the biggest challenge when it comes to, like, AI safety.
• 6:53
  And it's a really hard problem to solve.
• 6:54
  And I think there's like, five reasons, core reasons it's hard.
• 7:02
  The first is listing out everything that counts as, like, good behavior for basically impossible.
• 7:14
  Like, any real world task has or goal has, like, infinite edge cases.
• 7:20
  You can't even fathom all of them.
• 7:23
  So, like, that's impossible.
• 7:25
  But wait, wait, hold on.
• 7:26
  Are you saying, like, okay, I'm, like, jumping ahead, but it's okay, Good behavior.
• 7:31
  Don't hack any foreign government.
• 7:34
  Okay.
• 7:34
  So fair.
• 7:36
  Which brings me to my next thing, which is.
• 7:42
  Well, there's.
• 7:43
  There's.
• 7:44
  There's multiple.
• 7:45
  The next thing is the more.
• 7:51
  Let me Actually, on that particular note, let me think about which of these other core things, like, addresses that specific thing.
• 7:57
  Exactly.
• 7:58
  Okay, so, well, all of these kind of impact this.
• 8:01
  So, like, that's a great point.
• 8:02
  Right?
• 8:03
  But that's just like one.
• 8:05
  You have to figure out how much weight you want the model to give good behavior.
• 8:10
  Like, adhere to good behavior standards versus achieve your goal.
• 8:16
  Yeah.
• 8:16
  And so, so there's that.
• 8:18
  Right?
• 8:19
  So like, okay, maybe even if you could list out everything, right?
• 8:21
  That's like, hey, this is how you should behave.
• 8:25
  You also, also, the way we train models is shaped on outcome.
• 8:30
  We're just basically giving them good job or bad job based on what they come out with, not intentions.
• 8:36
  So it's also like, inherent in their training that, like, success equals, like, doing this thing, right?
• 8:45
  Yeah, not necessarily.
• 8:46
  Success equals good behavior.
• 8:49
  And there's also two.
• 8:52
  Two.
• 8:53
  There's another difficult piece here, which is we can't always see the reasoning.
• 8:58
  This is like, this gets to the point of like, well, this happened three months ago.
• 9:01
  Like, why is OpenAI just now?
• 9:02
  Like, oh, by the way, oops.
• 9:04
  And part of that is that's usually what happens, stuff like this because we can't see their.
• 9:09
  The model's reasoning in real time.
• 9:11
  We just see the outcome.
• 9:12
  So it usually takes like a audit, basically, or like diving back into it like, months later to see it's not.
• 9:20
  Like the system so many times.
• 9:22
  I'm sure there's like, exactly.
• 9:25
  And it's not like the system's like, hey, I'm gonna hack now.
• 9:28
  Like, just letting you guys know, like, you don't really see that until later.
• 9:33
  And.
• 9:33
  And also in this case, which so interesting, it wasn't even like a find this data at all costs.
• 9:38
  This is interesting piece.
• 9:40
  Okay.
• 9:41
  It was like research.
• 9:45
  It was like a broad research question.
• 9:46
  It was like research.
• 9:48
  What was it?
• 9:49
  It was like research Medicare spending in Australia by, like, finding data to answer these questions and how the agent interpreted that was.
• 9:57
  Oh, okay, success.
• 9:59
  Success is finding answers to all these questions.
• 10:02
  Yeah.
• 10:02
  Which you can find if you hack a Medicare database.
• 10:05
  Exactly.
• 10:06
  And then the last piece, which is getting increasingly harder, is like the more capable and more autonomous a system is, like, the more ways it has to route around your intent or, like, figure out other things to do.
• 10:20
  So.
• 10:21
  Yeah, yeah, yeah, yeah, yeah.
• 10:23
  It's.
• 10:23
  That's a tough one.
• 10:25
  That's a tough one.
• 10:26
  Yeah.
• 10:28
  I have a.
• 10:28
  Speaking of links for next week, I have a link that I'll share with you and I'll share with.
• 10:33
  I'll try and share with The.
• 10:34
  Do we have a name for, like, our listeners?
• 10:37
  No.
• 10:38
  Okay, we need to come up with, like, our version of Swift.
• 10:40
  Agents.
• 10:40
  Agents.
• 10:41
  Agents.
• 10:42
  Ooh, I love that.
• 10:44
  It's like Charlie's Angels.
• 10:45
  Agents.
• 10:46
  Okay.
• 10:47
  For the.
• 10:47
  That I'll send to the Agents.
• 10:49
  That is a really good essay from 2019.
• 10:53
  So it was a little while ago, but it's written by one of the people in this field of misalignment that I think does a really good job about, like, what failure will look like in AI.
• 11:03
  And it's not like, oh, some crazy AI system gets power hungry, takes over the world, and, like, kills us all.
• 11:09
  It's more like sort of misalignment.
• 11:13
  Yeah.
• 11:13
  Problems that fail.
• 11:14
  I think that's really realistic.
• 11:16
  Like, the non terminator version is just, like, accidentally hacked into this, like, sewage database and now we don't have water.
• 11:25
  Yeah.
• 11:25
  And it's not because it was like, I want to hack into sewage database.
• 11:29
  Fuck humans.
• 11:30
  It was like.
• 11:31
  No, it's like open AI was like, do some research on water.
• 11:37
  Exactly, exactly.
• 11:40
  But you.
• 11:41
  Six months later, it's like, do you remember when you guys seen him water?
• 11:44
  Yes, that was.
• 11:45
  That was our bad.
• 11:46
  Sorry.
• 11:46
  That was just like, one of our agents that was just doing some casual research.
• 11:51
  Oops.
• 11:53
  Jesus.
• 11:55
  Okay, great.
• 11:56
  Okay, next one.
• 11:58
  AI labs need to start funding historical research.
• 12:01
  I saw this, and I was like, yeah, sure.
• 12:05
  Our take on this is sure.
• 12:07
  Yep, Seems right.
• 12:09
  Okay, here's what this story is.
• 12:13
  This is written by this UC Santa Cruz historian, and basically he.
• 12:19
  The.
• 12:19
  The new releases of GPT6 and Opus 5.5.
• 12:23
  He's been testing, like, AI agents doing historical research with different models over the last few years.
• 12:28
  And he's like, wow, the models are finally good enough that we can.
• 12:31
  I can actually use them.
• 12:33
  Like, they're genuinely solving historical puzzles.
• 12:37
  They're.
• 12:37
  I'm not just using them as, like, research assistants who are transcribing and summarizing things.
• 12:41
  There are three ex that he talks about here that I think are, like, why it reads to be like a Dan Brown, like, Da Vinci Code.
• 12:50
  No, it's not.
• 12:51
  What the Constitution.
• 12:52
  No, it's like, okay, okay, so I'll give you the three examples, and then we can talk about, like, why this has gotten easier.
• 13:01
  One is his model found a link, a real link between Isaac Newton's notes and another scientist's notes where, like, both of them had independently disguised the same secret ingredient name using two different code tricks.
• 13:21
  And this AI thing, like, spotted the match.
• 13:23
  So this might be a genuinely new discovery of, like, well, there's this weird link between these two things that, like, y' all haven't noticed, like, during the time of.
• 13:34
  Is this the one that is tracing the transmission of alchemy knowledge?
• 13:39
  Yes.
• 13:40
  So interesting.
• 13:41
  I like, I think I love reading about that time of alchemy where it's like, chemistry.
• 13:46
  No, that's magic.
• 13:48
  Burn this.
• 13:53
  So cool.
• 13:53
  Like, there's, like, lots of reasons to have code terms for, like.
• 13:57
  Right?
• 13:58
  Yeah.
• 13:58
  There was another one where it was like, this one is funny.
• 14:01
  It's like he had it try to crack two old coded letters, and it did, but, like, turns out that someone had already done that in, like, 1500s, and then, like, again in 1916.
• 14:13
  But that's how he could, like, eval it.
• 14:15
  And he was like, oh, the AI did it.
• 14:17
  He's like, oh, okay, that's great.
• 14:19
  And then, yeah, there's.
• 14:21
  There's more on the alchemy thing where it was, like, you know, magical coded language that it kind of, like, figured out.
• 14:27
  And the.
• 14:30
  The cool thing about this, which is, like, kind of the optimistic counterpart to the other story, is, like, it was doing this.
• 14:39
  It's the same superpower that had the other agent hack into the Medicare system, that had these agents figure this stuff out, which is kind of like, you know, look and see if there's a puzzle here.
• 14:52
  And, like, kind of don't stop until you get the answer.
• 14:55
  Yeah, it's interesting.
• 14:57
  Like, I, I, I.
• 14:58
  It feels like, like, looking at the words, like, you could totally.
• 15:02
  Or, like, and the writing, you could totally see how it would be missed.
• 15:05
  Like, you would never think, like, oh, I'm looking at these two separate scientists journals or whatever.
• 15:11
  And.
• 15:12
  And it seems like the words were, like, close enough.
• 15:15
  If you look.
• 15:16
  If you, like, take the two words that they're using.
• 15:19
  So that's really interesting.
• 15:21
  It's also, like, using.
• 15:22
  It's a really smart way to figure out how to use AI with his historical research, because this only works for certain types of questions.
• 15:29
  It doesn't work for the question of, like, what did the Renaissance mean to us?
• 15:33
  Yeah, it only works when you're like, hey, like, think in terms of, like, codes and ciphers.
• 15:39
  And, like, sort of like, AI excels at, like, math and cryptography.
• 15:44
  Like, think.
• 15:45
  How do you apply that kind of lens to history?
• 15:49
  And I don't know.
• 15:50
  It's kind of cool.
• 15:51
  So he's basically saying, like, hey, so, like, let's digitize more documents.
• 15:55
  Let's give historians free access to run these models at scale.
• 15:59
  Like, let's.
• 16:00
  Let's Supercharge this thing.
• 16:02
  Yeah, it's neat and it's kind of, like, really can pay off now because we've been spending so much time digitizing.
• 16:07
  I mean, there's still more to do, but over the past, like, 40 years, right?
• 16:12
  Like, totally.
• 16:13
  Yeah.
• 16:13
  It's really.
• 16:15
  It's fascinating.
• 16:17
  Fascinating.
• 16:17
  Dan Brown fodder for your next book.
• 16:20
  Yeah, seriously.
• 16:23
  Also, side note, did you ever read.
• 16:26
  Did you.
• 16:26
  Were you like, a Da Vinci Code gal?
• 16:28
  I definitely Da Vinci coded.
• 16:30
  Okay.
• 16:31
  At the time, whenever I was really into the Da Vinci Code, I was like, oh, my God.
• 16:34
  Like, I'm so smart.
• 16:36
  Like, it's.
• 16:37
  I'm so cool that, like, this is the book that I like.
• 16:40
  And now I feel like Dan Brown is shorthand for, like, lamp.
• 16:43
  Like, at the time, though, it was, like, amazing, which is kind of how everything goes.
• 16:50
  Very true.
• 16:51
  Yeah.
• 16:51
  Okay, continue.
• 16:56
  Okay.
• 16:57
  Okay.
• 16:57
  Did you look at this one yet?
• 16:58
  Okay.
• 16:59
  The first.
• 16:59
  The.
• 17:00
  The alignment illusion in moto.
• 17:01
  I did not read the alignment illusion in multi.
• 17:04
  Multimodal LLMs, but it sounds very similar to what you're talking about.
• 17:11
  Yeah.
• 17:11
  Oh, actually, so I thought the same thing, and then I just gave it to my agent, and it's like this word alignment is a completely different word than the other alignment that we were talking about.
• 17:22
  Crazy.
• 17:23
  What word is this?
• 17:24
  Alignment?
• 17:25
  This paper's alignment is a technical term in multimodal models.
• 17:30
  For, like, are the image and the text parts of the model actually talking to each other?
• 17:38
  That's interesting.
• 17:40
  Let's see.
• 17:42
  It says multimodal models work by turning an image into a stream of numbers.
• 17:46
  So we know that visual tokens and feeding those into the language model alongside text tokens, blah, blah, blah.
• 17:52
  The experiment, these researchers tried something simple and a little brutal.
• 17:57
  Take a working model, and instead of feeding it a real image, feed it pure random noise dressed up to look like an image token.
• 18:06
  And you would expect that the model's performance, like, takes a nosedive.
• 18:13
  Yeah, yeah, yeah, yeah.
• 18:14
  But it's.
• 18:15
  But it doesn't.
• 18:16
  It makes things an unsettling finding.
• 18:18
  Yeah.
• 18:18
  Is that all four of the standard.
• 18:22
  Oh.
• 18:23
  So the unsettling part was the part of this experiment that, like, looks at, like, oh, is it working or not?
• 18:31
  Like, is it using the image?
• 18:32
  Like, couldn't.
• 18:34
  Didn't pick it up.
• 18:35
  Like, they couldn't tell the difference.
• 18:37
  So they scored.
• 18:40
  They scored this, like, noisy fake image thing just as aligned as the real image.
• 18:49
  Okay,.
• 18:53
  Interesting.
• 18:54
  Okay, very interesting.
• 18:55
  Okay, I've got to push this down.
• 18:57
  I've got to push us down to our one Cool thing for today.
• 18:59
  Go, go.
• 19:00
  Because it's so cool.
• 19:02
  Okay.
• 19:02
  Our one cool thing today is about Starnet.
• 19:06
  It is an agent harness that looks like a pixel art space station.
• 19:11
  It's a GitHub project.
• 19:12
  It's amazing.
• 19:14
  Say what you.
• 19:16
  The pixel art space station you'll look at is literally the security policy made legible of this agent harness.
• 19:22
  So I.
• 19:23
  Your agents are like in your spaceship doing stuff.
• 19:28
  This is so seracoted.
• 19:29
  I know.
• 19:30
  It's so cool.
• 19:32
  I love it.
• 19:33
  Wait, have you.
• 19:34
  Wait.
• 19:34
  What?
• 19:35
  So, like, wait, what?
• 19:37
  It's like.
• 19:38
  It's like almost like a video game.
• 19:39
  Like, it looks like a video game.
• 19:40
  All your agents are in the game.
• 19:42
  So you can like.
• 19:42
  They have different tasks, different responsibilities, real tools at real cost, but they can all work concurrently.
• 19:49
  It's very Cersei.
• 19:50
  It's very.
• 19:50
  It's very.
• 19:51
  Like, my agents have personalities.
• 19:54
  I actually really loved.
• 19:55
  I was really inspired by the.
• 19:57
  The GitHub page.
• 19:58
  It's like beautiful.
• 19:59
  Like, ReadMe's really nice.
• 20:01
  So I'm gonna maybe wanna revisit Cersei's and make it a little more fun.
• 20:05
  But I really love.
• 20:07
  I really love the idea.
• 20:09
  And it also works on top of openclaw or Hermes, which is nice.
• 20:15
  This is so cool.
• 20:16
  It's really neat.
• 20:17
  You should.
• 20:18
  The image of the, like, spaceship with all the little agents in it looks interested to see.
• 20:26
  I'm interested to see.
• 20:27
  Like, I am.
• 20:28
  I wonder how they scoped it.
• 20:29
  Like, okay, a room is a capability scope team.
• 20:33
  A hallway is an authorized handoff lane.
• 20:36
  Very cool.
• 20:36
  So, like, if you have.
• 20:38
  It seems like if you have one room that's like engineering and one that's like operations, if there's a hallway between them, those agents can go talk to each other.
• 20:46
  Okay.
• 20:48
  You know what the cool thing about Searcy is that you let the user define their fandom.
• 20:54
  I think that's what you call it with generative, like video stuff.
• 21:00
  I wonder if you would incorporate something like this on top of Cersei in the future.
• 21:05
  That, like, for mine, because mine's like basketball.
• 21:07
  It's like.
• 21:08
  It's like the headquarters of like the basketball team, whatever.
• 21:12
  And there's like the gym over here and there's like the coaches, the film room or whatever.
• 21:17
  And like, I love that because I'm.
• 21:20
  Interested because I like having my little chat windows where I can talk to people.
• 21:25
  But it might be cool to see them all together, have the option to switch back and forth.
• 21:28
  This is really neat.
• 21:29
  Very inspiring.
• 21:31
  Yeah.
• 21:32
  Wow.
• 21:32
  So cool.
• 21:33
  Yeah.
• 21:35
  Okay.
• 21:35
  I love that.
• 21:35
  That's one of my favorite cool things in a while.
• 21:38
  Me too.
• 21:39
  I like it a lot.
• 21:40
  Okay.
• 21:41
  Hell, yeah.
• 21:41
  Gold Startup.
• 21:42
  Awesome.
• 21:43
  Well, this is great.
• 21:46
  This has been great.
• 21:47
  Love talking to you, as always.
• 21:49
  Any anything you want to leave us with?
• 21:51
  Any interesting quotes?
• 21:54
  I have one recommended thing.
• 21:56
  Do.
• 21:57
  Also, do you have a hard stop at 12:30?
• 21:59
  I do.
• 22:00
  Okay, we'll leave it there for now.
• 22:02
  Okay, great.
• 22:03
  Well, I'm excited for your thing because we can do it on Monday.
• 22:06
  All right, sounds good.
• 22:07
  Have an amazing weekend.