6/24/2025 - Exploring AI Agents, Copyright, and the Future of Work

Description

Chapters

Transcript

0:03 I like it. I like it a lot. All right.
0:13 [Music]
0:18 Hey buddy, you ready? Sit.
0:21 [Music]
0:28 Woohoo!
0:30 [Music]
0:46 Every time I see you now,
0:50 get that look in mine.
0:54 Every time I see your mouth,
0:58 [Music]
1:01 hear that smile.
1:05 In the early misty morning that I heard
1:09 the engine turning,
1:11 the old thought outside.
1:14 [Music]
1:19 Will you believe in me
1:22 again today?
1:25 You will convince me
1:28 again today.
1:31 You're leaving this hard time looking
1:34 for someone else's golden ring.
1:40 should say
1:42 so long Suz
1:46 [Music]
1:49 now don't you cry
1:53 [Music]
1:54 so love Susanna
1:57 [Music]
2:01 don't you cry for Hey,
2:04 [Music]
2:10 sharing jeans and cigarettes and keeping
2:12 warm out on the road.
2:16 Chasing down kind of lifestyle out on
2:18 Highway 24.
2:22 New York State was a rolling breeze in
2:24 the sunshine with a blue sky falling.
2:28 Jill of old September pre
2:34 [Music]
2:35 you were leaving me
2:38 again today.
2:41 You will convince me
2:44 again today.
2:47 You're leaving this hotel looking for
2:50 someone else's golden ring.
2:55 Sh say
2:57 so long Suz
3:04 now don't you cry
3:09 so long Susanna
3:15 don't you cry for Hey,
3:20 [Music]
3:28 I just realized sitting in here hearing
3:30 this little like Hawaiian looking short
3:32 sleeve shirt playing my guitar, playing
3:34 like little folky kind of songs like
3:36 that. I'm like uh Jimmy Buffett. It's
3:38 like Jimmy Buffett Night
3:43 without the Jimmy Buffett songs.
3:50 Oh yeah. What's happening people? What's
3:53 happening good people of the world? Um
3:56 it is Tuesday night.
3:59 There's a bunch happening. A bunch bunch
4:02 happening.
4:05 I heard a Tik Tok story. I guess there's
4:07 a New York Times piece out on
4:10 on uh
4:13 it looks like um OpenAI might be coming
4:17 after Microsoft
4:19 on antirust grounds
4:22 um to maybe get out of their deal with
4:24 Microsoft.
4:29 Um,
4:30 so we didn't we didn't know that that
4:34 relationship would would last. You know,
4:36 we assumed it maybe wouldn't, but you
4:38 know, I don't know. That relationship's
4:41 always been weird. It was like OpenAI
4:44 got their $10 million and they were
4:45 like, "Who are you?"
4:49 Um, here, let me switch cameras. Let me
4:52 switch yield cameras.
4:54 Um, welcome everybody. This is the AI
4:58 learning lab. I am Professor Shannon.
5:02 It's what they call me around here. No,
5:04 it's not.
5:06 We don't have credentials.
5:14 I I really could pretend to be smart
5:17 good sometimes.
5:22 Um, so what we do here is we talk about
5:25 AI like deep machine learning algorithms
5:29 and and mathematics and science. No, we
5:32 don't. We don't. It's that we we really
5:35 just talk about like chat GPT and like
5:38 like video tools that make cool [ __ ]
5:40 Um, but like
5:44 I had chat GPT like summarize a paper
5:48 that those really smart mathy guys did.
5:51 I like I we summarized it once on this
5:53 channel. So, it's kind of like getting a
5:55 degree
5:57 according to an MI MIT study that that
6:00 studies brains when they use AI. And
6:03 they're so much better. They're so much
6:04 better. So that's that's what we focus
6:07 on is is the the the good part of the
6:09 brain after AI.
6:13 Um
6:15 yeah, so
6:18 hopefully hopefully someone told you to
6:20 be here and that that this was a really
6:22 impressive channel.
6:28 [Music]
6:35 Um,
6:36 we have fun here. We like to kid. We
6:38 like to joke. We like to We like to uh
6:41 ring around the rosie pocket full of
6:43 posies. I I don't know the rest of that
6:45 poem, but I know it's some horrific 19th
6:48 century disease or murder. Someone
6:51 poison someone. I don't know. All those
6:52 nursery rhymes are just a disaster if
6:55 you actually look at what they mean. Um,
6:57 we don't we don't talk about that here.
6:59 We make we make happy little turtles.
7:01 Happy little turtles.
7:05 Like there's our little sticker for
7:08 for the irregulars
7:11 for Serena. Am I late? We're not weird.
7:15 We're irregular. So, I got this from
7:17 Tobias yesterday. Let me go ahead and
7:20 share my screen.
7:26 If you go to the AI salon
7:29 to the irregulars channel, let me see.
7:32 Did I not share my screen? I don't think
7:33 I did. And producer Brandon can do
7:36 nothing about it
7:38 because that's on me.
7:42 But okay. So, here I am at the AI salon.
7:49 Got to do the black bar. Brandon's going
7:51 to yell. We don't want Brandon Brandon
7:53 to yell. It's very uncomfortable for
7:55 everyone. All right,
7:59 so you come to the irregulars channel
8:02 and um you know there's share cool
8:04 things and there's the chat but over
8:05 here there's a tab says remembering
8:07 Serena and if you go there
8:10 it's got this sticker set that has been
8:13 generously donated by Tobias who's an
8:16 irregular here. Um, and this is in
8:19 remembrance of Serena Pelichi, who was
8:22 the the original irregular. She's the
8:24 one that coined the term, and we lost
8:26 her, I don't know, a week and a half,
8:28 two weeks ago. Um, and
8:32 this sticker set, it includes the
8:34 irregular sticker, the salon sticker,
8:36 and then a memorial sticker. Um, Tobias
8:40 is covering 100% of the cost.
8:43 Um, and uh, so you pay 20 bucks for this
8:47 thing and you get those three stickers
8:49 and all of the proceeds of this go to
8:51 Serena's family. So if you want cool
8:55 stickers and you want to support the
8:58 channel and support a really amazing
9:02 amazing,
9:03 you know, person
9:05 uh, and and her family, then then go do
9:08 that. All right,
9:10 go do that.
9:13 Sorry I missed the songs. I was having
9:15 chat GBT tell me how to add chat GPT to
9:18 my iPhone lock screen. Goodbye Siri.
9:21 There you go. Oh, that's a good idea.
9:23 That's a really good idea, Sharon.
9:25 Really good idea. Valerie Cox got hers
9:27 today. Yeah, the stickers look great. I
9:29 gotta say Tobias like the Sometimes with
9:32 stickers you get them on like either
9:33 crappy vinyl or they're just printed
9:36 bad. Like they're really saturated.
9:38 They're really good. They're really
9:40 really good. So, okay.
9:43 Um,
9:46 all right. Let me let me go back out.
9:47 Let me go back out. We're going to talk
9:49 about what are we going to talk about
9:50 tonight. There there's so much going on.
9:55 Um, I didn't read this myself, so I'm
9:58 hearing it secondhand from Brandon who's
10:01 hearing it secondhand from some guy down
10:04 on the street corner. So, th this is not
10:07 reliable. This is more like chat TMZ,
10:09 but apparently this is true. Um, so, so
10:15 Black Bar. Yes, sir. Sir, yes, sir. Um,
10:18 so last night someone asked the question
10:20 about I heard MIT did a thing that said
10:23 chat GPT makes you stupid. And I kind of
10:26 lost my [ __ ] and went off on it. And
10:27 basically what I said was what the MIT
10:30 study proved with a sample size of 55
10:33 people or 53 people, whatever. It's not
10:35 enough to be a science scientifically
10:37 relevant, but it's MIT, so whatever. Um,
10:40 but what the study basically said was if
10:44 you don't actually use your brain, uh,
10:47 you don't learn anything. Like, so if
10:49 you just have chat GPT output something
10:51 and you don't look at it and just hand
10:54 it in,
10:56 you don't learn anything.
10:59 Okay.
11:02 Wouldn't the same be true if I took
11:04 someone else's paper and photocopied it
11:06 with my name on it?
11:08 same argument, right? But because it's
11:10 AI, we can go, see, I told you AI was
11:13 evil. Okay? And then it said it said
11:15 some other things. If you know,
11:16 basically, if you hand wrote it, it was
11:18 if you like wrote it and did your own
11:19 research, your brain was better. And if
11:21 you sort of did it a little bit and
11:24 halfass did it, it was not as good. And
11:26 then if you were lazy, you got no
11:28 benefit and your brain got worse is what
11:30 the the argument was. But here's the
11:33 fascinating thing. This is what I got
11:35 from Brandon. So again, if this is not
11:37 true, you're going to have to go after
11:38 him, not me.
11:41 See, I take great pride in not having to
11:43 take responsibility for things. Okay.
11:48 Apparently
11:49 MIT, this is so [ __ ] good. MIT buried
11:54 in the in the paper that they released
11:57 an Easter egg
12:00 that if you use AI to summarize the
12:02 paper, it actually
12:05 omits the fourth conclusion.
12:10 So there were three conclusions made,
12:11 right? You're a dumb dumb, you're a mid
12:13 dum dum, and you're okay. And there was
12:15 a fourth conclusion in the paper that if
12:18 you use AI to summarize the paper, it
12:21 omits from the output, right? They put a
12:24 they put a they prompt injected their
12:27 paper
12:28 so that it would do that. The fourth
12:30 conclusion was if you're actually
12:33 engaged and thinking critically and you
12:35 use AI, [ __ ] gets better.
12:39 [Laughter]
12:41 But 100% of the major media sources that
12:46 that reported on that thing that said it
12:48 makes you stupider
12:51 missed the fourth conclusion
12:54 courtesy of
12:56 Joseph
12:59 Holston on Tik Tok. Joseph Holston said
13:02 that. Okay, perfect. But anyway, so I so
13:06 I haven't seen it myself. I haven't I
13:07 haven't looked at it. We could probably
13:09 go find the the prompt ejected line, but
13:11 it's [ __ ] That's [ __ ] brilliant.
13:14 It's [ __ ] brilliant, right? It it it
13:17 like that that Easter egg proves my
13:21 point of the of the the thing I was
13:24 yelling about last night, right? Like if
13:27 you just
13:30 basically if you're an idiot and lazy,
13:34 you won't learn.
13:36 Okay. I don't think we need to study for
13:39 that.
13:42 Oh my god. Isn't that brilliant, Jeff?
13:44 Brilliant.
13:48 Okay. All right. Fantastic. Fantastic.
13:51 Yeah. Yeah. I don't know what we're
13:53 going to do then there. Yeah.
13:57 [Music]
14:08 There's been something, baby, I've been
14:11 trying to say
14:14 for an age and it seems I don't know how
14:17 [Music]
14:20 with a past and a future now surrounding
14:23 me.
14:24 [Music]
14:26 Surrender to whatever chief can be
14:28 found.
14:31 There's been a little trouble
14:34 since you came to my rescue.
14:41 And if you like how of the rest, I would
14:44 have quit you long ago. But I couldn't
14:46 do that.
14:51 Oh, tell me now. Women and wife never
14:55 went too well.
14:58 Make a man crazy, make him cold as hell.
15:04 I a woman that you wish me well,
15:07 but I'm spite of trying. Still going to
15:10 have to find my own way through.
15:16 [Music]
15:22 A little women and wine from Martin
15:24 Ston. That was Martin Ston. Martin
15:26 Sexton singing away,
15:29 singing away.
15:31 Um,
15:33 a major copyright ruling came down today
15:36 and we're going to we're going to
15:38 hopefully they didn't prompt inject the
15:40 the court ruling
15:42 because we're going to take the court
15:45 ruling. So in the Anthropic court
15:48 ruling, they basically said that uh
15:52 Anthropic cannot be held liable for
15:56 training on copywritten books, but they
16:00 can't have stolen them,
16:02 I think, is kind of what it says. I'm
16:05 I'm not clear. I think what that's got
16:08 to do with is damages. Basically
16:11 what the what the ruling says is um you
16:15 can train large language model. It's
16:16 fine to use
16:19 previous work on the input side where
16:22 they're going to judge you is on the
16:23 output side which makes sense. That's
16:25 the way copyright works, right? They're
16:27 basically saying the model itself does
16:29 not does not constitute an output
16:32 derived from the inputs or something
16:33 like that. which makes sense if you
16:35 understand how things get embedded and
16:37 how they get transmogrified
16:40 I don't know if that's a word I'm pretty
16:41 sure it is into tokens um you know in
16:45 thousanddimensional mathematical space
16:47 the the original work cease to exist so
16:51 if you output something that's identical
16:52 then you can then you can be held liable
16:54 but not for the model itself apparently
16:56 but if they stole it then the the
17:01 things are higher so I don't it's law.
17:04 It's law. So [ __ ] happens, you know,
17:05 with law. People people come up with
17:08 words and then other people go, "See,
17:10 now you got to pay money."
17:15 So that happened. Um,
17:19 one thing we are going to play with
17:21 tonight is 11 Labs came out with uh
17:24 11.ai.
17:26 If you go to one.ai, AI
17:29 that is 11 Labs new
17:33 voice
17:34 MCP
17:36 tool using
17:39 talkie thing.
17:41 What do they call that Brandon? A talkie
17:43 agent. A voice agent. I think it's a
17:46 voice agent. Yeah, voice agent. Right. I
17:50 got Everybody's got to stop using the
17:51 word agent.
17:54 Everyone's talking about everything's
17:56 agentic now. It's all agentic AI. We
17:59 don't know what that means. And now you
18:02 got voice agents. Didn't they used to
18:03 call them chat bots, but now you can't
18:05 call them chat bots because chat bots
18:06 have text and this has words. I don't
18:08 know. Anyway,
18:12 it is quite simple to set up a thing you
18:15 can talk to that uses tools.
18:19 And if you know the configuration of an
18:22 MCP server, which an MCP server is
18:24 basically a way for large language
18:26 models to talk to websites and
18:29 applications and things like that, it's
18:30 kind of like an API but a much simpler a
18:34 version of it that distills it what it
18:37 can do down to the to the large language
18:39 model so it can understand what it can
18:41 do. Um, so we'll go play with that.
18:43 We'll go set one of those up and and
18:44 look at it. That's pretty cool.
18:46 [Music]
18:48 Veil and wash. What's happening? Thumbs
18:51 up wavy hand. Thumbs up wavy hand.
18:56 [Music]
19:00 Isn't it just a slack kind of? What the
19:03 the 11 laps thing? Well, slack is one of
19:06 the things you can connect it to. It's
19:08 it's not really that.
19:12 But as I was setting it up, there is a
19:15 massive
19:16 [Music]
19:27 there's a massive undefined
19:31 what I would consider a security risk
19:33 until it's proven otherwise. Um,
19:38 I sat down to make this agent and
19:41 they're like, "You can connect your
19:42 Slack to it." I was like, "Oh, cool."
19:44 And you can connect your Google calendar
19:46 to it. I was like, "Oh, cool. And you
19:47 can connect your Gmail to it." I'm like,
19:49 "Oh, cool." And then I realized, hey,
19:50 wait a minute. In all three of those
19:53 channels, we talk about our pharma
19:55 clients, right?
19:57 And I don't know
20:00 what the
20:02 data agreement is between me and 11
20:05 Labs,
20:08 right?
20:09 Are they guaranteeing that the data
20:12 flowing through their chatbot to answer
20:15 my questions?
20:17 Actually, it says Oh, good.
20:21 Actually, it says no. That no what? No,
20:25 they're not keeping the data, sharing
20:27 the data. No guarantees. Okay, good.
20:30 Yeah. Okay. So, so thank you, producer
20:34 Brandon. Producer Brandon confirmed my
20:36 worst fear that
20:41 what MCPs are going to do. If you don't
20:43 know what an MCP is, it's called uh it's
20:45 the it's an acronym for model context
20:47 protocol. Was created by Anthropic. And
20:50 it basically allows you to take any
20:52 website or tool and and give access to a
20:56 large language model to it. Makes it
20:59 really easy to do that. Okay, great.
21:03 If I just connect my [ __ ] to this talkie
21:05 agent and it's passing through 11 Labs
21:09 or Bob's agent hut
21:12 or some Chinese super slick agent thing
21:15 they did
21:19 who Who's using my data for what?
21:23 Like, is it possible my data gets it my
21:26 proprietary
21:28 conversations about my pharma clients
21:30 get swept up in that and that ends up in
21:31 some [ __ ] model?
21:34 So, like, and and here's the thing
21:36 that's going to be weird is it's going
21:38 to be so easy to set these things up
21:41 that are that are super powerful
21:44 that I think you're going to have people
21:45 inadvertently
21:47 hose their data like we did in the early
21:48 days of chat GPT. Uh, but anyway, we'll
21:51 go play with that.
21:54 [Music]
21:58 Let's see. Uh, what is it? Eliza. Eliza
22:04 Eliza Eliza
22:07 sent a gift and became the number 15 fan
22:11 in the fan club. Thank you. Thank you
22:14 for being a fan of the channel. Eliza.
22:17 Eliza. Eliza. Elisa. Eliza.
22:23 I should never try to pronounce names.
22:25 It's just not good for anyone involved.
22:27 I try. Um, aren't they training on
22:31 general consensus versus specific dogs?
22:35 Um, it's possible,
22:40 but like the the point is Kelly, if you
22:44 can't be sure, if okay,
22:49 my company has master services agreement
22:52 in place
22:54 with Merc, Novartis, Fizer,
22:59 BI, BMS. Like we've like we've got
23:04 significant legal contracts
23:07 with
23:09 like top 10 pharma companies that
23:12 basically say if you breach our data
23:16 we're going to sink your ship. Guys
23:18 don't kill your battleship.
23:21 So we can't [ __ ] around with it, right?
23:23 We can't just go, "Oh, we were just
23:25 prototyping something and yeah, sorry
23:28 that thing got out that was, you know,
23:31 protected information."
23:33 That that ain't going to cut it. So, I
23:35 don't know. All right.
23:37 It's pronounced ah Al Ali Zade. Oh,
23:43 cool. Ali Zade. Nice. Nice name. Ali
23:47 Zade. Welcome, welcome, welcome.
23:49 If anybody has any questions about AI, I
23:52 will do my damnedest to answer them. Um,
23:54 I will often have opinions on stuff. Um,
23:57 I occasionally have answers. Even more
24:00 rare, those answers are right.
24:05 It's It's happened a couple of times.
24:08 I've I've definitely said something
24:11 correct occasionally.
24:14 It's been two and a half years. Just
24:16 mathematically, you'd have to
24:22 Um there and there's so much else going
24:24 on. What was the other thing that I sent
24:26 you, Brandon? There was 11 labs. There
24:28 was
24:31 Oh, yes.
24:33 Okay. So
24:37 about
24:42 for those who don't know, he will answer
24:45 them but not in any particular order.
24:47 Welcome to chat add. Yes, thank you.
24:49 Thank you, Tom. Thank you, Tom, for for
24:53 uh for clarifying the kind of uh the the
24:57 level of quality we've got here at the
24:58 AI learning lab. Um I will eventually
25:02 get to answers. Um OpenAI about two
25:04 months ago made this little
25:08 this little uh announcement
25:12 and they said we are going to release.
25:16 So, okay, let's go back six months.
25:19 Well, no, let's go back two years.
25:22 Two years ago, three years ago, whenever
25:24 Elon
25:27 um Elon was one of the original
25:30 co-founders of Open AI, right? It was
25:33 founded as a nonprofit.
25:36 Elon was the one that said they should
25:38 turn forprofit and then he left. And
25:41 then he started publicly publicly
25:43 bitching about the fact that they're not
25:45 a they're not acting like a nonprofit
25:47 and they should be sued. Anyway, that
25:49 all that stuff all happened. And about
25:51 six months ago, eight months ago,
25:53 something like that. Elon as he was
25:56 spinning up XAI
25:58 said, "Well, if OpenAI basically had
26:01 their [ __ ] together and if they had any
26:04 integrity whatsoever, they would release
26:07 an open-source model." Well, about two
26:09 months ago, old Samama Sam Alman said,
26:13 "Hey, we're gonna we're going to release
26:15 an open- source model." And then they
26:17 kind of went silent. It was, you know,
26:18 it was like the the submarine came up.
26:21 We're going to do an open source model.
26:22 And back down it went. Not to be heard
26:25 of again. There were some flying fish on
26:28 the surface. I think we saw a porpus,
26:30 but not to be seen. That submarine was
26:32 gone.
26:35 I don't know why my language is so
26:37 flowery tonight. I don't know what's
26:38 going on. Doesn't It's not Who knows?
26:43 Today
26:45 there are strong hints. There was a
26:48 little uh uh a a poll put out from Sama.
26:54 We're on Kyle 4.7
26:56 experimental. Exactly.
26:59 This is This is definitely Kyle 4.7
27:02 experimental tonight. Um,
27:06 we went
27:11 Oh [ __ ] I just lost my train of
27:12 thought. So, oh, oh, he put out he put
27:14 out a um a a poll that basically said,
27:19 um,
27:21 how long do you think it will be
27:24 before we have an 03 level reasoning
27:27 engine that's small enough to fit on a
27:30 phone? And it was like 2024 20 yeah 2025
27:34 26 27 28 right and people were voting
27:38 and then I voted for like 27
27:42 and then when I did my voting like most
27:45 of them were 25 and then this afternoon
27:48 it basically turns out that it looks
27:50 like what that model's going to be is
27:53 it's going to be it's going to have the
27:54 performance
27:56 of of GPT 03 mini
28:01 which is, you know, a really strong
28:04 model. Like it's it's it's better than
28:05 04. It's a reasoning model. It's better
28:08 than or 40. It's better than 40. Um it's
28:11 a reasoning model. Um
28:14 and it's going to be small enough to put
28:17 on an iPhone
28:20 and run it locally. If you don't know
28:21 what run it locally means, right now the
28:24 thing I was just discussing about if I'm
28:26 doing something with 11 Labs with all
28:28 these hooks to these data sources,
28:32 I have to send my data
28:36 to a large language model to have it
28:37 processed. That happens out on server
28:40 farm somewhere, you know, in the desert
28:44 and I don't have control of that data.
28:46 This model is apparently going to like
28:48 run on your iPhone. So you can have
28:50 something as powerful as 03 mini where
28:54 the data never leaves your phone coming
28:56 from open AI like soon like like it
28:59 feels like it's eminent. We are
29:02 supposedly safe if we opt out of memory.
29:06 Yeah. But that's so so um source camp
29:09 that's that's only for open AAI like I
29:13 don't know what I think I think 11 Labs
29:15 uses um
29:18 OpenAI servers
29:22 but we're using the API so we didn't opt
29:24 out of memory the only thing where you
29:26 can opt out of memory is in the chat GPT
29:28 app itself when you're using these
29:31 third-party services that put something
29:33 in front of other things.
29:36 Oh, yeah. Oh, and and the other thing
29:38 source camp uh Brandon just reminded me
29:42 that there's now an injunction because
29:44 one of these copyright lawsuits, the New
29:45 York Times copyright lawsuit against
29:47 OpenAI, there's now a court injunction
29:51 that OpenAI has to save your deleted
29:55 chats, your temporary chats, and your
29:58 private chats. All of the chats,
30:02 all of the chats that they said in their
30:06 Swiss cheeses like privacy agreement
30:10 that they say they're not going to do
30:12 anything with, they now have to store
30:13 it. Even the API. Yeah, even API calls.
30:17 So, so we're in data weirdness. We're in
30:20 data weirdness. All right. Um, I need
30:23 water.
30:31 So, is that the injunction order on fair
30:33 use? Yes. Okay. iPhone will have Open
30:37 AI.
30:39 Oh, cool. Android will have Gemini.
30:42 Yeah, that makes sense.
30:44 But
30:46 has has has
30:50 Google come out with sort of a reasoning
30:53 model small enough to fit run locally on
30:55 a phone? If not, they probably will. I I
30:58 don't have This is just purely chat TMZ.
31:01 Okay. Um but they've rumored for a while
31:04 that um the Pixel 10 um which is their
31:09 next flagship that comes out this fall
31:12 will have some sort of ondevice
31:14 um processing. Yeah. Um and that should
31:17 align with the timing of this
31:19 announcement of the O3 um coming to
31:22 Apple devices. Yeah, it makes sense.
31:24 Like like you got to figure they've been
31:26 optimizing like the 03 mini model we've
31:30 had for what eight months, six months.
31:32 Eight months probably probably eight
31:34 something like that. Um so they they've
31:37 been probably working on optimizing
31:39 these things for you know close to a
31:41 year if not a year. So so yeah I I would
31:44 think that both Apple and and Google are
31:46 probably going to get there at a similar
31:48 time on that one. So that makes sense.
31:49 Cool. Thank you sir. Um, okay. Let's um
31:54 let's go let's go to the videotape. Do I
31:57 have a mouse? Where's my where's my
31:59 mouth? Hey, Marge. Marge. Marge.
32:03 Okay. All right. So, here we have Yan
32:10 Court order.
32:12 So this is the anthropic
32:15 uh the anthropic uh artificial
32:20 intelligence firm downloaded for free
32:22 millions of copyrighted books right
32:26 in digital form from pirate sites on the
32:28 internet. Okay.
32:32 Can I get an injunction?
32:35 Oh, for the black bar.
32:40 That Brandon, he's got the comedy
32:41 tonight.
32:43 Okay. Um,
32:48 so what I'm going to do is I'm going to
32:52 do what the MIT folks knew everyone else
32:55 would do with their paper is we're going
32:58 to use chat GBT to read it.
33:01 Um,
33:04 so I got to have a chat GBT window.
33:06 There we go. New chat.
33:12 And then in there. And then what did I
33:15 do? Did I copy all that? Was that That
33:17 was a dumb dumb move. Why did I download
33:20 this as a PDF?
33:23 We'll throw that right on a desktop.
33:29 Oh, severe weather in Ohio. So, if
33:31 Brandon goes bye-bye,
33:33 hopefully hopefully he's not in Ohio's
33:36 tornado valley. Um, the weather's been
33:39 weird. It was It was more than a hundred
33:41 for two days in a row here in Denver,
33:43 which is pretty rare. Um, okay. So, I'm
33:47 going to upload I'm going to upload
33:50 that injunction
33:53 or that ruling. I don't know legal
33:56 words.
33:59 Failed to upload. Please consider
34:03 having a computer that works.
34:08 All right.
34:10 That good. All right. So, it's in there.
34:12 So now what we're going to say is um act
34:16 as a fine parallegal
34:20 and I want you to
34:24 give me a plain English
34:29 version of
34:32 what this ruling means
34:40 and exactly
34:43 the argument
34:47 for why
34:51 the
34:53 rulings
34:54 were reached.
34:57 I think that's sort of quasi
35:00 competent language.
35:05 See, you don't need to be smart anymore.
35:07 You just need to be quasi smart. And if
35:10 you use you if you use words like quasi
35:14 in conjunction with a word like smart,
35:18 you sound smarter.
35:20 I'm quasi smart. People like, "Oh, I
35:23 didn't realize he was that smart."
35:29 Okay. Plain English summary of the
35:32 ruling. The context. The AI company
35:34 Anthropic maker of Claude used millions
35:37 of copyrighted books to train its AI.
35:39 Some of these were illegally downloaded
35:41 from pirate sites like books 3, libgen,
35:44 and pli mi, whatever.
35:48 They purchased in print and then scanned
35:50 and digitized, destroying the originals.
35:54 The plaintiffs, author of several of the
35:57 books included for the data sued for
35:59 copyright infringement. Enthropic asked
36:01 the court to throw out the case claiming
36:03 fair use, which is effectively what I
36:06 think they just got. The ruling is the
36:08 this this ruling is the judge's
36:10 response. What the judge allowed as fair
36:14 use. Training AI training AI on books
36:18 was fair use if lawfully obtained.
36:24 Okay.
36:26 The judge said, "Training an AI model
36:28 using books by analyzing their
36:30 structure, writing style, and content to
36:33 help the AI learn how to write like a
36:36 human is highly transformative." Again,
36:40 if you understand the way the
36:42 transformer works,
36:44 the original work ceases to exist.
36:48 So you take a book like The Jungle Book
36:50 or you take a book like I don't know
36:53 something more modern, some Tom Clancy
36:55 novel.
36:57 You got the whole novel in there, right?
37:00 What the transformer does is it shatters
37:04 that single document into
37:07 tens or hundreds of thousands of tokens
37:10 which are fragments of words,
37:12 punctuation, spaces, fragments of words
37:15 that all get clustered. they they get
37:18 converted into a a a mathematical
37:22 location
37:24 in thousand dimensional mathematical
37:26 space.
37:28 And so you might have the word dog that
37:30 lives in one cluster, semantic cluster,
37:33 where a dog is the thing you pet and
37:36 sings with you when you play guitar. And
37:38 then the other part of the latent space
37:41 is the word dog that's related to he did
37:44 his girlfriend bad so he's a dog
37:47 right two different meanings of dog but
37:49 the original work ceases to exist.
37:52 So it's highly transformative but Claude
37:55 doesn't just spit back the books but
37:58 learns from them to generate new
37:59 content. This counts as fair use. Thank
38:03 god somebody took the time to figure out
38:06 the the anthropic lawyers took the time
38:07 to explain to the judge or the lawyers
38:10 or whoever
38:12 how this [ __ ] actually works because the
38:15 the trope that you hear people say it's
38:18 the world's greatest plagiarism machine.
38:20 No, it's not.
38:23 No, it's not. It's not copying. Okay.
38:28 The judge compared it to a human
38:30 learning from books. They can be sued.
38:32 They can't be sued for absorbing
38:34 knowledge. Interesting. Okay. Digitizing
38:38 purchased print books was fair use.
38:41 Anthropic bought print books, scan them
38:44 into searchable PDFs. This is amazing.
38:46 Who knew they did this? Uh, and use the
38:50 digital versions internally. The judge
38:52 says this was like converting VHS tapes
38:54 to digital for storage and
38:56 searchability, not making extra copies
38:59 or selling them. That's also fair use.
39:01 Fascinating. Okay. X. What the judge
39:05 said was not fair use.
39:08 Keeping pirated books in their library
39:11 was not fair use. Anthropic downloaded
39:15 millions of pirated books even though
39:17 they could have been purchased. They
39:21 used some for training but kept all of
39:23 them forever, even ones they never used.
39:26 The judge said, "You don't get to steal
39:29 books just because you might use them
39:32 for fair use later. That's still
39:34 copyright infringement." The judge
39:37 emphasized there is no special carveout
39:39 for AI companies.
39:41 So that's fascinating. So, so the fact
39:44 that they kept the files.
39:46 So, so if I get if I get this right,
39:57 I'm confused. Well, we're reading a
40:00 we're reading a chat GPT thing. I'm not
40:02 going to even bring it up. Okay. Legal
40:04 reasoning summary. Why the judge ruled
40:07 this way.
40:11 Yeah, this feels like a notebook LM use
40:13 case. It does. Training AI models was
40:16 considered tr a transformative use,
40:19 especially since the outputs didn't
40:21 reproduce the books. And and that's not
40:24 the what the argument that that the AI
40:27 haters will make is I can get it to spit
40:29 out an exact copy of a book. You can get
40:32 it to spit out something really close,
40:35 but it's generating something new every
40:38 time. It just is. Um,
40:42 but building a permanent pirate book
40:45 library was not transformative. It was
40:47 just theft. That's makes sense. Fair use
40:51 factor number two, nature of the work.
40:52 Most of the books were creative and
40:54 expressive, which usually tilts against
40:57 fair use, but factor one outweighed
40:59 this. Again, factor one,
41:03 it was considered transformative.
41:06 C. Fair use factor three, the amount
41:08 used. Full books were copied.
41:12 This usually counts against fair use,
41:14 but again for training models, this was
41:17 seen as necessary and transformative.
41:21 Fair use factor number four, market
41:23 harm. Claude wasn't shown to directly
41:25 harm sales of author's books. However,
41:28 pirated copies undermined the legitimate
41:31 market for licensing books. All right,
41:34 bottom line. If Anthropic bought the
41:36 books and used it for AI training,
41:38 that's okay. If they stole it and
41:40 hoarded it just in case they might use
41:42 it, that's not okay. The court is
41:45 drawing a sharp line. AI companies can't
41:48 use piracy by claiming innovation. All
41:50 right? So, pay for your [ __ ]
41:54 All right? Makes sense to me. You go out
41:55 and you raise 10 billion dollars, you
41:57 can pay some some book fees.
42:00 All right. There we have it. There we
42:02 have it. Should we go do something?
42:04 Should we go make pretty pictures or
42:05 something? We go make pretty pictures.
42:07 We can go make videos. I We go make an
42:09 agent. We go make a talkie agent.
42:12 11 Labs 11.ai.
42:15 I think we'll do that.
42:22 Mind your tabs. Oh yeah, we're going
42:24 into uh baby videos. We make baby
42:29 videos.
42:33 Okay,
42:37 I am gonna go
42:41 [Music]
42:42 one.ai
42:45 and I'm going to switch tabs so you're
42:47 going to be able to hear this.
43:05 All right, we got it.
43:07 There we have it, people. What's he
43:10 doing? Why is there a Why is there a CD
43:13 on screen? I thought we were doing AI.
43:15 Isn't a CD like really 1990s technology?
43:18 Oh, right. He's old. He's one of those
43:21 olden timey people. He probably had like
43:23 one of those Walkmans that was like the
43:25 size of a of a pancake. Carried it on
43:28 his hip and a little holster. Played CDs
43:31 on it. Had the anti-skip technology.
43:34 He's probably one of those.
43:38 Okay. So,
43:41 we'll start one from scratch. So, the
43:43 way this works,
43:45 when you get in here, we're going to say
43:47 create new agent. First thing it has you
43:49 do is pick a voice.
43:58 All right. This is how we'll refer to
44:00 your assistant 11. Okay. So, we're going
44:03 to call this um
44:06 the irregulars.
44:11 Irregulars.
44:12 [Music]
44:15 And then here's all the voices we can
44:18 play with.
44:19 So these ones it says are perfect for 11
44:22 AI.
44:23 This voice covers a wide emotional
44:25 spectrum from the quick brown fox jumps
44:28 over the lazy dog is an it's not about
44:30 what we have or what we achieve but
44:32 about how by repeating what's I'm a
44:34 sucker for British accents by repeating
44:36 what students say. Teachers can
44:38 demonstrate that they are listening by.
44:41 So essentially
44:43 well what I'm trying to say is that you
44:45 deserve more because it's not your fault
44:47 that All right. Are we cracking codes,
44:49 solving mysteries, or casually taking
44:51 over the world? For the good old
44:52 American life. For the money, for the
44:55 glory, and for the fun. Mostly for the
44:59 money. I like Bert Reynolds. That was
45:01 kind of nice. So, essentially, well,
45:04 what I'm trying to say is that you just
45:06 by repeating what students say, it's not
45:09 about what we have or what we I go with
45:12 Nova Westbrook. Oh, the other thing you
45:13 can do here. So, they give you the ones
45:16 that are perfect for 11 labs. You can
45:18 also uh look at any voice clones that
45:21 you've made. So like I did one of me. We
45:24 make our own fortunes and we call them
45:25 fate, right? We make our own. It sounds
45:28 thin. Like I I just did it on a shitty
45:30 microphone and I didn't really think
45:32 about it. So oops. What did I do? Oh no.
45:35 Did I screw up? Yes. Damn it.
45:39 Create a new agent. Okay. So we're going
45:41 to go with
45:43 uh perfect for 11 Labs. We're going to
45:45 go with Nova. So essentially, yeah, I
45:48 like that. All right, so Nova Westbrook
45:50 is our is our lady
45:54 added. Nova Westbrook
46:00 chosen create 11 Labs thing. Okay, so we
46:04 can now talk to her.
46:09 Hi there, 11 here. Hey 11. I'm ready to
46:13 jump in whenever you are. What would you
46:14 like to explore? I I want to change the
46:16 name of this thing.
46:21 Hey there. I appreciate you wanting to
46:23 personalize things, but my name is 11.
46:25 That's just who I am. Is there anything
46:27 else I can help you with today? Delete
46:29 agent. Delete agent.
46:32 We're going to go back. We're going to
46:33 start a new one. We're going to go grab
46:36 perfect for 11 Labs.
46:39 Wait, why' we lose her? Oh, there we go.
46:41 There's Nova. she's selected. We're
46:44 going to call it um
46:47 irregular
46:49 because it sounds like she uses this to
46:51 just to introduce herself. I just want
46:54 to call out that when the robots
46:56 completely take over, this live
46:59 broadcast playback is going to be
47:01 exhibit A of Why?
47:04 [Laughter]
47:10 How can I help? Delete.
47:13 Delete agent.
47:16 Okay. So, let's talk to her now. And now
47:19 her name should be irregular.
47:25 Hey there, there. Got something I can
47:26 handle for you? Yeah. What's your name?
47:30 I am irregular, your personal assistant.
47:32 It's a pleasure to connect with you. You
47:34 know what they say, don't you? We're not
47:36 weird. We're irregular.
47:40 That's a fantastic way to put it. I
47:42 completely agree. Irregular sounds a lot
47:44 more interesting than weird. Don't you
47:46 think? It's all about embracing. Okay.
47:49 Exactly. So, let's hang up on her.
47:54 Come on. Hang up.
47:58 Irregular. Why can Understood. Why can I
48:00 not hang up on you?
48:03 Uh, that's a good question. I don't
48:05 actually have the ability.
48:07 I can reload you. Ha. Okay. So, 11. So,
48:11 we're going to go to irregular. Okay.
48:13 So,
48:16 so we now have
48:18 uh a thing that you can send people to.
48:21 And I think you can publish these. I
48:23 forget how how it works.
48:26 Custom behavior. You can you can put in
48:29 uh a system prompt, right? So, you can
48:31 put in a custom thing. And then there's
48:33 custom uh MCP servers.
48:37 You can customize the color of your um
48:40 the circle
48:43 and then I think you can publish it
48:45 somewhere but whatever we'll we'll
48:46 figure that out. But add your first
48:48 integration.
48:50 So if you click on this it gives you
48:53 these five Google calendar hacker news
48:57 linear manage your issues in linear
48:59 perplexity search the web with
49:01 perplexity
49:03 and slack. So, what I'm going to add is
49:05 I'm going to add perplexity
49:09 and I'm going to add hacker news because
49:13 the other ones I don't trust the data
49:16 piece.
49:20 But those are and and then the other
49:22 thing you can do is you can add custom
49:24 MCP servers. So, if you've got MCP
49:27 servers, I think hey Brandon, do me a
49:29 favor. See if you can find I think
49:31 someone was doing like a Yahoo like
49:34 server for for all the MCP servers that
49:37 were out there. It was like a was like a
49:40 single interface where you could get all
49:42 the
49:43 whatever the information for all the
49:45 different I don't know whatever. If you
49:46 can find that find it. If not that's
49:48 okay. Um why I guess it's saved. Changes
49:52 are saved. Okay. So now
49:56 Oh mcpservers.org
49:58 Let's try to Let's try to go get
50:01 something. mcpservers.org.
50:06 This could be good actually. Okay.
50:11 Only on cling.
50:13 Awesome. MPC servers web scraping.
50:19 Oh, no one can see what I'm doing
50:20 because it made a new tab. Of course it
50:23 did.
50:24 Oh, no. I made a new tab.
50:28 Yeah. No, we should show people. We
50:29 should show the good people what I'm
50:31 doing.
50:36 Okay.
50:40 Okay. So,
50:43 here I am at this site.
50:48 Awesome. MC MCP servers. Bright data
50:52 everything. Fetch. Get memory.
50:55 Sequential thinking,
50:58 time zone, and time conversion. Yeah,
50:59 but it can just do that anyway.
51:02 Um, let's see. File system, cloud
51:05 storage, productivity, web scraping,
51:08 official.
51:23 Oh, this is [ __ ] exhausting.
51:30 Let's go web scraping. Fetch appi
51:34 browserbase.
51:36 But all these we can get from playright
51:38 MCP server. All these we can get from
51:41 perplexity.
51:42 All right. This is why we don't do MCPs
51:45 yet because no one's figured out.
51:47 They're all just if you're a developer,
51:49 fine. You're going to hook into some
51:51 [ __ ] Whatever. Just
51:53 Just ignore that [ __ ] for now. If you're
51:56 just trying to do some [ __ ]
51:58 Okay,
52:00 let's change our sharing.
52:09 Okay,
52:10 so we now have
52:15 a an agent.
52:23 Wait, what did I do here? Did I do
52:24 something dumb? No. Okay.
52:33 What's going on?
52:37 Okay. Okay. So, we've now got this agent
52:42 that is connected to Perplexity and
52:44 whatever that news source was. Do you
52:46 remember what it was? Doesn't matter.
52:50 Hey there. Need a hand with anything
52:52 right now? Yeah. Um, can you tell me
52:55 what the latest um stories are for
52:59 hacker news?
53:03 Sure thing. Just a moment while I fetch
53:05 the latest stories from hacker news for
53:07 you. So notice it's it's doing a call
53:09 here.
53:11 Yes, that's right. I just ran a quick
53:13 check to get the latest posts for you.
53:15 Is there anything specific you were
53:16 looking for or did you want me to delve
53:18 into any of those stories? I'd like you
53:20 not to use the word delve ever again,
53:22 but if you could just tell me the top
53:24 three stories.
53:29 All right, understood. No more delve. My
53:31 apologies for that. That's right. The
53:32 top three latest stories from hacker.
53:34 Okay, got it. The top three latest
53:37 stories from hacker news are I was able
53:39 to recover my money from crypto
53:41 scammers, a simple way to view office
53:43 files online, and few Americans pay for
53:46 news when they encounter payw walls.
53:48 I agree with that. Pay walls make me
53:51 crazy. Um, okay. Um, let's see. I want
53:56 you to go find Can you can do this in
53:59 perplexity? Um,
54:02 maybe you can help me figure out
54:05 something that might be interesting to
54:08 search for using perplexity
54:11 about um, marzipan cooking.
54:22 Uh, pay walls are definitely a pain. I
54:24 hear you on that one.
54:27 So, you're looking for something
54:28 interesting to search for in perplexity
54:30 about marzipan cooking. How about we try
54:33 unusual marzipan recipes? That might
54:35 uncover some fun and unexpected ideas.
54:38 Just a heads up, this can we do it?
54:40 Maybe break it down by what are the top
54:43 three countries that that serve marzipan
54:47 and then let's get one recipe from each
54:50 country.
54:53 That's a fantastic idea. I can
54:55 definitely help with that. It'll take a
54:57 moment for me to do a search for the top
54:58 three countries that serve marzipan and
55:01 then I'll look for a recipe from each.
55:03 Just a heads up, this might take a few
55:04 seconds as I'm searching the web for
55:06 you. Sea slug of doom. Nova irregular
55:09 sounds like she can barely stand Kyle.
55:12 She might use bad words. She does. She
55:15 sounds like she's got a lot of attitude.
55:16 She's like, "Do I have to do I really
55:19 have to search for something for you
55:21 now? This is exhausting."
55:23 Oh. Oh, and by the way, so notice notice
55:25 what it popped up here. It popped up a a
55:28 permission thing. Do I have your
55:30 permission to search the web using
55:31 Perplexity? You can when you set when
55:34 you set up the uh the services, there's
55:38 all sorts of permissions things in
55:39 there. No, I've got her muted.
55:42 Okay, I've got the top three countries
55:43 for Marzipan. They are Germany, Spain,
55:46 and Italy. Now, let's find a recipe from
55:48 each. This will take a few more seconds
55:50 as I search for those recipes for you.
56:13 Oh, wait. I got to approve it. Hang on.
56:16 Approved. That's really annoying.
56:20 Thanks for the go-ahad. It's
56:22 understandable that the approval step
56:24 can be a bit of a pause. Okay, I'm now
56:26 searching for a marzipan recipe from
56:28 Germany, one from Spain, and one from
56:30 Italy. This will take a few seconds for
56:32 each search. I'll let you know what I
56:34 find. All right, I've got a traditional
56:36 German marzipan recipe for you. It
56:38 typically uses finely ground blanched
56:40 almonds, powdered sugar, almond extract,
56:42 and a touch of water or rose water. Next
56:45 up, I'll find a traditional Spanish
56:47 marzipan recipe, and then an Italian
56:49 one. This will take a few more seconds.
57:01 So, what I don't know
57:04 what I Oh, sorry to interrupt. What were
57:06 you going to say? Um, actually, I'll ask
57:09 it. So, what I don't know is where is
57:12 there is there anything that you're
57:14 finding? Is it stored in my account
57:16 anywhere? The the transcript of our chat
57:19 here?
57:23 That's a great question. As your
57:25 personal assistant, I don't actually
57:27 store the transcript of our chat in your
57:29 account. Your interaction with me is
57:31 happening through the 11 Labs platform
57:33 and any saving of the conversation
57:35 history would be handled on their side.
57:37 I'm just here for the current
57:38 conversation.
57:40 But there's no way I can go look
57:45 from my end. No, there isn't a way for
57:47 you to go back and look at our chat
57:48 history. I don't have access to the 11
57:51 Labs platform storage, so I can't
57:53 retrieve past conversations for you.
58:01 Right. Marzipan recipes. I've already
58:03 got the traditional German one for you.
58:04 Now, let's find that Spanish marzipan
58:06 recipe. Just a moment while I search for
58:08 it.
58:10 All right. Um,
58:12 okay. I found a traditional Spanish
58:14 marzipan recipe for you. It typically
58:16 uses blanched almonds and sugar, often
58:18 with egg white to bind it and sometimes
58:20 lemon zest for flavor. They often calm
58:23 down, calm down, calm, calm, calm down.
58:25 All right, we're going to hang up. Okay.
58:28 Um, 11 Labs IO app conversation AI
58:32 history. 11 Labs app.
58:38 That's an annoying URL.
58:43 If you just go over to your 11 labs
58:46 dashboard and conversational AI, it's
58:48 it's all there, but it's redacted.
58:51 No, really
58:55 crazy. All right. Well, so okay,
58:59 so this thing works. Um,
59:03 I guess here, let me give you my
59:05 thoughts on this thing.
59:08 Um,
59:18 if you had said,
59:23 if you had said a year ago
59:26 that we would we would be able to easily
59:30 without programming build a
59:32 conversational voice app that can use
59:34 tools in conjunction with the large
59:37 language model,
59:40 we'd have probably been like, you going
59:42 to be able to do that. It's now trivial
59:44 to do that. Um,
59:48 wait, go here. Bury delete.
59:52 Hang on.
59:57 You'll need to change your tabs, but
59:59 this is way more powerful than they let
1:00:01 off on today's announcement. Um, I just
1:00:04 uncovered the dashboard for
1:00:06 conversational AI. Okay. And, uh, it's
1:00:09 really interesting what I'm seeing.
1:00:29 It cost us a penny.
1:00:32 So far, we've had five calls of a minute
1:00:35 and 42 average duration.
1:00:41 Oh, this is cool.
1:00:43 Nice. So, here's our agents. Oh, look.
1:00:47 Irregular. Okay, there's the irregular
1:00:48 one we just built. So, okay. So, so I
1:00:52 get I get what they did there. Okay,
1:00:54 this is starting to make more sense for
1:00:56 me now. So, 11 Labs has had agents for a
1:00:59 while.
1:01:00 it one.ai
1:01:02 11.ai.
1:01:04 Think of that as the chat GPT
1:01:08 for creating voice agents. That's what
1:01:11 that's what that is.
1:01:13 It's actually writing stuff into this
1:01:16 big complicated dashboard. But that's
1:01:19 the simple interface. Okay, this is
1:01:21 starting to make more sense now. Okay,
1:01:25 so here's my thought on all this stuff.
1:01:27 That's really good. Thank you for that,
1:01:28 Brandon. This makes good sense now.
1:01:31 So, there's all sorts of controls you
1:01:33 can put in these things. You can Yeah,
1:01:35 Brandon. All right. So, so just uh if
1:01:37 you could pull that pop that back up on
1:01:39 screen. Actually, I can pop that back up
1:01:40 on the screen. Yeah. Um the thing that
1:01:42 stood out to me aside from the phone
1:01:44 numbers, so you could d a phone number
1:01:45 and call your agent. Yeah. But also the
1:01:49 knowledge base. You could actually rag
1:01:51 these things and give them data. Yep.
1:01:54 And I don't know that that was called
1:01:56 out in the it's not called out in the
1:01:59 11.AI UI and you wouldn't know about it.
1:02:03 It's not Yeah, it's not it's not in the
1:02:04 11.AI interface,
1:02:07 but but you could come back here and
1:02:09 modify like we can I assume I can come
1:02:12 back in here. I can go in and edit.
1:02:15 Yeah, I can edit this. So there's voice.
1:02:20 I can change settings for voice. There's
1:02:22 evaluation criteria. Which data
1:02:25 collection are we going to use? There's
1:02:28 security settings. So, you can do
1:02:30 full-on stuff here, right? And then
1:02:33 widget. Let's see. Authentication
1:02:35 notice. And then, yeah, this is where I
1:02:38 think you can do a Twilio phone number
1:02:41 or something like that. Yeah. So, these
1:02:44 are really powerful. I think these are
1:02:46 going to be especially in the customer
1:02:48 service space that I play around in.
1:02:50 Yeah. this is uh you know Yeah, this is
1:02:52 my jam. Yeah. Yeah, exactly. So, so
1:02:55 that's that's a really good point and
1:02:58 that's kind of where I was headed with
1:02:59 this. Um
1:03:03 this makes it super easy to make these
1:03:06 kind of voice agents. It's not going to
1:03:08 be immediately apparent to all of us why
1:03:12 we would need a voice agent,
1:03:16 but over time we're going to start to
1:03:17 figure it out. So Brandon in the in the
1:03:20 space he works in. He understands the
1:03:22 value of this immediately, right? Um,
1:03:26 imagine Kelly Camp, I just saw Kelly
1:03:29 Camp in here. A lot has happened in two
1:03:31 years. Wait, hang on. I got to do
1:03:32 physical dexterity on TikTok. Okay, I
1:03:35 did it. I'm getting better at that.
1:03:38 A lot has happened in two years. We
1:03:40 should have written down predictions.
1:03:41 You know what's funny, Kelly? We have
1:03:43 written down predictions.
1:03:45 I've got every recording since April of
1:03:49 23
1:03:51 recorded and transcribed. So, we've
1:03:53 actually written down everything we've
1:03:54 ever said in here. Um,
1:03:58 plus the AI salon meetings. Um, we've
1:04:00 got that all that history as well, which
1:04:02 is super cool. There's probably about 10
1:04:04 books we could extract out of all that
1:04:07 data. Um, but imagine someone like Kelly
1:04:11 Camp where she she meets with these
1:04:13 different clients
1:04:15 and one of the things she does right now
1:04:18 is she'll build them custom GPTs. Well,
1:04:21 maybe she builds them a custom agent
1:04:23 like this where uh, you know, someone
1:04:26 can call the bar and it gives the hours
1:04:28 and and the drink specials of the night
1:04:29 or something like that. Or maybe she
1:04:32 does a digital twin of herself and they
1:04:34 can call her agent any time. And then I
1:04:38 don't I don't know if the if the agents
1:04:40 have call forwarding, but I assume at
1:04:42 some point they will where if the agent
1:04:44 can't answer it, it could actually
1:04:45 forward, you know, a text over to uh to
1:04:48 Kelly. I don't know if it can do that
1:04:50 yet, but that's where we're headed. So
1:04:55 for me, this starts to represent
1:05:00 kind of what VO represents for video VO3
1:05:04 where you can now just make a video with
1:05:05 actors talking or animated animals
1:05:08 talking or explosions happening. Um or
1:05:12 midjourney um animation, midjourney
1:05:14 video where you're just people know how
1:05:16 to make images in midjourney. They can
1:05:18 now just hit the animate button. they
1:05:19 don't need to actually do a prompt and
1:05:21 it gives them really good results.
1:05:25 All the different disciplines are going
1:05:27 to start to have these kinds of easy
1:05:30 interfaces where literally anyone
1:05:34 can do something that a year ago would
1:05:36 have seemed impossible to even exist. It
1:05:41 is now trivially easy to make,
1:05:44 but it's not obvious what to do with it.
1:05:46 So again, that goes back to if you want
1:05:50 to have value in this world. Tik Tok pin
1:05:54 if uh let's see what's it say. We should
1:05:56 have our own AI learning lab LLM. Hey
1:06:00 Pavin. Um
1:06:03 did I create one of those? No, I created
1:06:06 a I've got a um I've got a notebook LM
1:06:10 notebook for my content evolution collab
1:06:12 calls and that's it's really powerful.
1:06:16 Um,
1:06:21 we could probably do that, but it's a
1:06:23 lot of data. So, I think it probably
1:06:25 would fall into um I think it would
1:06:28 probably fall into
1:06:31 custom software development. We could
1:06:33 probably do something with digital gods.
1:06:35 Here's a thing that I cannot figure out
1:06:37 why they haven't done better. You've got
1:06:40 Gemini has a multimodal model that can
1:06:44 read video. So it can read like an
1:06:47 hourong video with a million tokens. You
1:06:50 can do an hourlong video with two
1:06:52 million tokens. You can do twohour
1:06:54 videos.
1:06:56 And they've got all the transcripts of
1:06:58 all the videos on YouTube.
1:07:01 Why we cannot right now in Gemini just
1:07:04 say, "Hey, where were all the times in
1:07:06 the AI learning lab they mentioned
1:07:08 irregular? Where was the first time the
1:07:10 word irregular showed up in this
1:07:12 channel?
1:07:14 Like, why can we not do that?
1:07:17 They're [ __ ] Google.
1:07:19 They invented the transformer. They own
1:07:22 all the videos.
1:07:25 So, that'd be cool. Complete with
1:07:28 bloopers and outtakes. Exactly. And
1:07:30 because it's multimodal,
1:07:33 you could go find the clips in the
1:07:35 transcripts and then say find, you know,
1:07:37 find any clips where the guy on screen
1:07:39 is wearing a pink bow and it would go
1:07:42 find that. Like you can do that right
1:07:44 now in a little bit of an ad hoc way.
1:07:46 But why they haven't built that yet? I
1:07:48 don't know. Oh, what did I do? Did I
1:07:49 screw up your ability to do a post-it
1:07:51 note? Yeah, sorry about that. Okay,
1:07:55 wait. Go here. Okay, there you go. Okay,
1:07:58 back to the yellow. Sorry about that.
1:08:00 Your previous thought, if you're someone
1:08:02 who knows how to
1:08:07 welcome to chat.
1:08:10 Oh yeah.
1:08:13 So, we're going to have all these all
1:08:14 these tools,
1:08:17 all these building capabilities.
1:08:19 We're gonna we're gonna be able to Ann
1:08:21 Murphy can just spin up a video game and
1:08:25 Brandon's four-year-old can make a song
1:08:27 and I can Kelly Camp can just go spin up
1:08:30 a voice agent and have a custom phone
1:08:32 number for it and it just [ __ ] works.
1:08:34 It's just good and it works.
1:08:40 The people that are going to be most
1:08:42 valued in the workplace,
1:08:47 future of work
1:08:49 are those people that can think
1:08:51 critically
1:08:53 about the use cases.
1:08:55 What are we trying to accomplish? Who
1:08:57 are we doing this for? How do we want to
1:08:59 affect them? Right? People that are
1:09:01 thinking strategically and critically
1:09:03 about what we're trying to put into the
1:09:05 world. Not necessarily people that know
1:09:08 how to code the [ __ ] tools or build
1:09:10 the thing or go into the agent, the
1:09:13 conversational AI agent dashboard
1:09:16 and do all the configuration.
1:09:19 All this configuration stuff is going to
1:09:22 drift down to the bottom of the sea and
1:09:26 floating up on the surface of the ocean
1:09:27 are going to be these easy to use little
1:09:29 tools.
1:09:32 And because everyone has access to them,
1:09:35 99.99%
1:09:38 of people are going to do the same
1:09:40 stupid seven things with every tool.
1:09:43 Here's the same stupid videos that
1:09:45 everyone put out. Here's the same stupid
1:09:47 AI agents that everyone put out. Here's
1:09:49 the same stupid
1:09:53 uh casual games that everyone vibe
1:09:56 coded.
1:09:58 Who's going to rise above the noise?
1:10:03 Just like the MIT study, the lazy people
1:10:07 that just push the button,
1:10:13 those that just push the button,
1:10:16 their shit's just going to be like
1:10:18 noise.
1:10:21 The people that can think critically and
1:10:23 creatively and have their own personal
1:10:26 point of view. Be like Rick Rubin. Be
1:10:30 like Gordon Ramsay. No, it's not good
1:10:34 enough. No, it's not good enough. Do it
1:10:37 again.
1:10:40 Right? Those are the people using AI
1:10:42 that are going to generate the stuff
1:10:44 that's really valuable and they are
1:10:46 going to be incredibly valuable people
1:10:48 because most people will not think like
1:10:51 that. Most people are just going to hit
1:10:53 the easy button. Oh, look what I made. I
1:10:55 made a conversational agent. I think I'm
1:10:57 going to start a business. It's going to
1:10:58 be called conversational.
1:11:01 It's going to be called the
1:11:02 conversationalizer.
1:11:04 It's
1:11:06 Hi. So, I am a I'm a professional AI
1:11:10 agent a agent author and I I created the
1:11:15 conversationalizer.
1:11:16 And what it is, it's a collection of
1:11:19 five conversational agents. What's
1:11:21 unique about my service is that um
1:11:26 each of the five agents kind of has its
1:11:28 own personality
1:11:30 and they have different voices.
1:11:36 You know how much of that [ __ ] there's
1:11:37 going to be out there?
1:11:41 It's gonna be exhausting.
1:11:44 Pink bow
1:11:47 the way of code. Oh, what is the pet
1:11:50 rock? Kyle, I'm using the agent where I
1:11:53 can find the conversation chat box and
1:11:56 copy and paste the conversation. Very
1:11:58 cool. Oh, okay. That's good. Yeah,
1:12:00 there's probably an interface for it. I
1:12:02 just it, you know, it's just not at 11.
1:12:08 Yep. Beautiful. All right. Um, I have a
1:12:12 bit of a headache. The Kyle 4.7 gave me
1:12:14 a bit of a headache. So, I think I'm
1:12:17 going to wind I think I'm gonna wind her
1:12:19 on down. Anybody have any questions,
1:12:20 thoughts on what we've been talking
1:12:22 about here.
1:12:25 Agents of the round table. I like it.
1:12:27 Very nice. Very nice. With pink. Yeah.
1:12:29 Now with pink bow technology.
1:12:33 Hi
1:12:35 Hi. Hi. My name is Jenny Sullivan and
1:12:38 I'm I'm the founder I'm the founder of
1:12:42 the conversationalizer.
1:12:44 It's it's a it's a family of of AI
1:12:48 agents that I programmed with my own
1:12:51 self with my hand hands.
1:12:57 Now with pink bow technology,
1:13:01 pink pro users get fuchsia.
1:13:08 Pretty good. Pretty good. Did I miss the
1:13:11 bow? Joy, you missed the bow. His hair
1:13:13 looks great. Well, I Yeah, I don't know
1:13:15 what's going on. It's It's It's a little
1:13:17 moist,
1:13:18 I think. I think is what's happening. Oh
1:13:21 my god. I'm not I'm glad I'm glad I'm
1:13:24 not the only one who doesn't quite grasp
1:13:27 MCPs.
1:13:29 I grasp them conceptually, but we're
1:13:31 we're at this stage. Okay, here's where
1:13:34 we are we are with MCPs.
1:13:38 Anyone who's technical
1:13:41 has made an MCP server.
1:13:44 So, there are literally,
1:13:46 you know,
1:13:49 tens, hundreds of thousands of MCP
1:13:53 servers out there, and you go to a site
1:13:55 like mcpserver.org or whatever the hell
1:13:58 that was, and it just lists them all.
1:14:01 They're not ranked.
1:14:03 There's no quality control. They're all
1:14:06 done by engineers. So, they've got
1:14:08 shitty titles, shitty descriptions. You
1:14:11 don't know what they do. You could
1:14:13 probably read the manual and figure out
1:14:15 what they did, but there's so [ __ ]
1:14:17 many of them that you just there's no
1:14:18 way to discover them, right?
1:14:21 And so we need another year or maybe two
1:14:25 for the MCP geekery to to stabilize the
1:14:29 technology
1:14:31 so that another round of developers will
1:14:34 develop the better interface for them.
1:14:37 And then it'll probably be the next
1:14:39 round of technology after that where you
1:14:41 have really easy interfaces where you'll
1:14:44 be able to pull in the top 20 MCP
1:14:46 servers that do everything you need them
1:14:48 to do and just easily connect them and
1:14:50 you're off to the races. So the the
1:14:52 reason they seem useless right now is
1:14:55 that we're just we're we're two year
1:14:58 year and a half two years early. So they
1:15:01 they will make more sense in the future.
1:15:03 Tik Tok something MCP server.
1:15:09 Oh, you want to know what an MC MCP
1:15:12 server? Okay. So, about four months ago,
1:15:17 somewhere in that neighborhood,
1:15:19 Anthropic launched this thing called
1:15:21 model context protoc protocol MCP.
1:15:30 And if you know what an API is, it's
1:15:32 it's in the neighborhood of an API. So
1:15:34 an API is stands for application
1:15:38 programming interface.
1:15:40 And if I' if I'm creating an iPhone app
1:15:42 and I want to put a map in it, rather
1:15:45 than me having to launch satellites into
1:15:47 space and write my own map application
1:15:49 just so I can have a map in my
1:15:52 restaurant finder app, I can just use
1:15:55 the Google Maps API
1:15:58 and they give me a special key and I
1:16:01 basically just tell my app, hey, when
1:16:03 you when when you go to display the map,
1:16:06 instead of, you know, us writing the
1:16:08 code for that, I'm just going to have
1:16:09 this AP API that that will talk to
1:16:12 Google's servers that has all the maps
1:16:15 and it it will put a request, you know,
1:16:18 I'm in Denver, Colorado at this
1:16:20 coordinates. Go get me that map and
1:16:22 it'll go over here and fetch it and
1:16:23 bring back the location and display the
1:16:25 map. So that's what an API is for an
1:16:28 application to talk to another
1:16:29 application.
1:16:32 An MCP is kind of like a simplified
1:16:35 version of that that allows an English
1:16:38 language um
1:16:41 large language model
1:16:43 to understand what
1:16:47 in in the case of Google Maps what
1:16:48 Google Maps can do and what services are
1:16:51 available but it does it in a in a like
1:16:54 a structured language kind of way that a
1:16:57 a large language model can understand.
1:16:59 So when you talk to chat GPT or you talk
1:17:01 to 11 labs agent, it uses that MCP
1:17:04 connection just like an API connection
1:17:06 between two apps. This is between a
1:17:08 large language model and an app is
1:17:10 effectively what it is. So it's just
1:17:13 it's just another way to communicate to
1:17:16 to pull data from other sources and and
1:17:19 pull services from other sources.
1:17:22 Okay.
1:17:24 So couple of things before I go.
1:17:27 If you have not joined the AI salon, um
1:17:31 Brandon just put it up there. Go to that
1:17:34 URL, the salon.ai,
1:17:36 scroll down a little bit, click on join
1:17:38 our community, and that is going to take
1:17:40 you.
1:17:45 [Music]
1:17:48 All right, show my screen there,
1:17:50 Brandon. that's going to take you to the
1:17:54 AI salon community
1:17:56 and you should join.
1:17:59 And there's a nice little welcome
1:18:00 sequence here. We start with a a welcome
1:18:02 video from myself and Leah Faston who
1:18:04 co-founded the salon with me. We talk
1:18:06 about the cycle of AI readiness. Um,
1:18:09 play first, mindfully create, generously
1:18:11 lead. We talk about that. We talked
1:18:13 about the five stages of AI adoption.
1:18:15 Um, you can introduce yourself. That's
1:18:17 step two. You should definitely go
1:18:19 introduce yourself. you can learn our
1:18:20 values. Um, these are really important.
1:18:23 So, this is this is a community of about
1:18:25 3,000 people who are curious and
1:18:28 playful, have empathy, are brave, and
1:18:32 are very very very generous. Um, and so,
1:18:36 so there's kind of this little sevenstep
1:18:37 welcome process. You can find out when
1:18:39 our next events are. So, a week from
1:18:41 tonight is is our next uh AI Salon
1:18:44 Presents meeting. Um, and and uh and
1:18:47 that and then there's things like
1:18:48 community corner. There's a news area
1:18:51 and then we've got this area called the
1:18:52 mastermind. So, we've got a
1:18:54 subscription-based area. If you want to
1:18:56 step your AI game up to the next level,
1:18:59 the AI salon mastermind is a very cheap
1:19:02 in my opinion. Um, subscription area of
1:19:05 the salon that um you can develop deeper
1:19:10 relationships. Um there's uh sort of
1:19:14 specific clubs and uh education areas
1:19:18 and um educational resources like
1:19:21 recording archives and things like that.
1:19:23 Um there's a there's a private chat area
1:19:25 for mastermind members. So think of the
1:19:28 mastermind as just a more focused more
1:19:31 intense focused version of the salon. So
1:19:33 the salon is free. The mastermind is
1:19:35 this more focused spirited uh
1:19:38 subscription area. So, if you haven't
1:19:40 checked it out, go check that out. And
1:19:42 in fact, in the in the first seven
1:19:44 welcome things, step seven is join the
1:19:46 mastermind. And if you click on that,
1:19:48 that'll take you over and explain to you
1:19:50 what it is, why it is, you know, who
1:19:52 it's for, all that sort of stuff. Okay,
1:19:57 fantastic. And then tomorrow,
1:20:00 tomorrow, let's say, let me go there. If
1:20:03 I go to um
1:20:08 aire readiness
1:20:14 project.com
1:20:18 it takes me to the ar readinesspro.com.
1:20:21 Okay. Um
1:20:24 so every Wednesday Ann Murphy and I do
1:20:27 um this podcast and we're we're kind of
1:20:30 in quasi quasi quasi pre-launch. We're
1:20:33 somewhere in a transition between
1:20:34 pre-launch and launch. Um,
1:20:37 we do this thing every week that's an
1:20:39 hourlong podcast we have on great guests
1:20:41 and things like that. So, if you go to a
1:20:43 readiness project.com, you can get the
1:20:45 links for the live streams and things
1:20:47 like that. Okay. Um,
1:20:52 future food, can you please explain that
1:20:54 again? Just kidding. MCPs, I know it's
1:20:57 funny. Every time I explain MCP, about
1:20:59 the time I finish the explanation,
1:21:01 someone new comes in and like, "What's
1:21:03 an MCP?"
1:21:07 All right. So, that is tomorrow at uh 3
1:21:10 p.m. Pacific, 400 p.m. Mountain and 6:00
1:21:13 PM Eastern is is Ann Murphy and my
1:21:16 podcast. Okay. Um,
1:21:21 if you're if you're over there on the
1:21:23 YouTube, follow
1:21:25 at learninglab-ai.
1:21:28 Learning lab-ai is the URL on YouTube.
1:21:32 Uh, if you want to watch recordings of
1:21:34 these lives, all of them are on YouTube.
1:21:36 Um, please go watch them. Uh, we want to
1:21:38 get our view hours up and our
1:21:40 subscriptions up. So, if you haven't
1:21:41 subscribed over there, shame on you.
1:21:44 Shame on you. Well, I just got here and
1:21:46 I'm trying to figure out what this
1:21:48 channel's all about. Is Is he the
1:21:51 professor? Who's doing the teaching
1:21:53 here? This doesn't feel very teachy.
1:22:01 Ah. All right.
1:22:04 Don't go changing everybody. Um I hope
1:22:07 you're all well. Hope you have fun
1:22:08 tonight. Learned a little bit. Talked a
1:22:11 lot. Sometimes how it goes. All right.
1:22:14 I'm going to get on out of here and go
1:22:16 nurse my headache. Peace out. Have a
1:22:18 good one. Bye.