7/17/2025 - OpenAI Agents: A Deep Dive into Capabilities and Shortcomings

Description

Chapters

Transcript

0:00 A champion.
0:03 A champ. Champy. Champ.
0:14 [Music]
0:22 [Music]
0:26 [Applause]
0:28 Hello.
0:31 [Music]
0:40 Hello.
0:42 [Music]
1:03 Hey, hey, hey, calm down. Sit, sit, sit,
1:07 sit, calm, sit, calm down.
1:11 [Music]
2:35 [Music]
2:49 Good evening. Good people. What is
2:52 happening? What's going on? Almost going
2:53 down.
2:55 [Music]
3:28 Do
3:31 [Music]
3:43 our speed levels. They'll marry our bons
3:46 together.
3:48 [Music]
3:50 I've got some real estate here in my
3:52 bag.
3:56 So, we bought a pack of cigarettes
3:59 and Mrs. Wagner's pass
4:02 and we've both come to look for America.
4:06 [Music]
4:09 [Applause]
4:12 [Music]
4:28 Yes. Yes. said as we boarded the ground
4:31 in Pittsburgh.
4:35 Michigan seems like a dream to me now.
4:40 So we bought it. Wait, it took me 40
4:43 days. Wait, it took me 40 days. It took
4:46 me
4:50 [Music]
5:14 Robert Rossy. Hey, Kyle, you're really
5:17 gotten you've gotten really good at
5:18 singing and playing guitar. Oh, thank
5:20 you very much.
5:23 It's, you know, it's like uh it's like
5:25 10,000 hours on the same seven songs.
5:32 I appreciate it. I do appreciate it.
5:36 [Music]
5:55 W Grace desperately heading his old
5:59 place. Dreamed would discover a new
6:01 space. Buried himself alive
6:06 inside his basement. tongue on the side
6:09 of his facement. He's working away on
6:12 displacement and what it would take to
6:15 survive.
6:18 Cuz when you're done with this world,
6:23 you know the next is up to you.
6:28 Ever once in his life it was quiet
6:35 as he learned how to turn with the tide
6:39 [Music]
6:41 and the sky was a flare and the
6:47 [Music]
6:52 I don't
6:52 [Music]
7:01 Oh, good evening. All right, so we got
7:04 agents. We got agents. Open AAI came out
7:06 with some stuff.
7:09 Um, okay. We've got some irregulars fun.
7:13 We've got a Dr. J video.
7:18 We've got a Dr. J Skittles rap. Okay.
7:21 Hey, and a Steo image. Nice. Okay, we'll
7:24 go look at that. We'll go look at that
7:26 stuff. Did I share my screen yet? No, I
7:27 haven't. I have not shared my screen
7:33 as very often because
7:36 it's become clear that
7:39 I am a loser.
7:42 Eno
7:44 also. Hi, Kyle Shannon and the
7:46 magnificent champion indeed. Champion
7:49 where? Well, Champ was magnificent in
7:52 here.
7:55 Hear about the big announcement. Oh,
7:57 here to hear about the big announcement.
7:59 I don't have it yet, so I can't play
8:02 with it. I didn't reup my $200 a month
8:05 subscription just to get it. Uh, so
8:08 we'll probably have it tomorrow for
8:09 Friday night date night or by the end of
8:11 the weekend for sure. Um, so I I can't
8:14 tell you if it's good or bad, but I can
8:16 walk you through what it is and and why
8:18 it's a why it seems like it's a big deal
8:22 and
8:24 um it will likely be disappointing.
8:28 [Music]
8:36 Um,
8:38 but you know, some some people that I
8:40 respect like Ali K. Miller uh had early
8:43 access to it and seemed to really like
8:44 it. Um
8:46 she said it's got some jankiness to it.
8:52 So, I'll talk about that in a bit. Uh
8:55 yeah, that's that's about it. That's
8:58 about it. All right.
9:01 Oh, Kyle, you're not a loser. Thank you.
9:06 I appreciate that. I you know there's
9:09 some days there's some days when the uh
9:13 the universe conspires to remind you
9:19 you maybe don't have it going on.
9:22 So I appreciate it.
9:29 All right. John Harris, what is up?
9:31 What's going down? Thank you, Mr.
9:34 Appreciate that.
9:44 source camp. Do you thought the the
9:46 OpenAI announcement was a a bit of a
9:48 disappointment? I didn't like I wasn't
9:51 expecting GPT5 and we didn't get GPT5.
9:54 So,
9:57 >> um,
9:59 let me go look and see.
10:06 So, let's see. What were we doing last
10:08 night? We built some fun stuff last
10:10 night. Oh, we built the uh we built the
10:13 Skittles image.
10:15 The Skittles dude talking about
10:17 Skittles. Like I like I said, as I
10:19 predicted, nobody nobody cared. Nobody
10:22 like that that post got no engagement.
10:27 I thought it was really funny, but you
10:28 know, apparently my uh my definition of
10:32 comedy is not the same as the world's.
10:34 Did you see the underwhelming
10:36 announcement today? I did. I I don't I
10:38 don't know if it was underwhelming.
10:40 Um,
10:50 I guess I I I have yet to use it, right?
10:53 So, I don't know, but it it looks like
10:56 it's somewhere in between Manis and
10:59 GenSpark. Um, in fact, Manis trolled
11:02 OpenAI today and said, "Welcome to the
11:04 party."
11:06 Which I thought was pretty good. Um so
11:08 if you don't know it came out today
11:10 OpenAI announced uh agents uh which is a
11:14 uh just a new mode
11:17 that effectively combines
11:20 um
11:22 deep research and operator. Operator was
11:25 their tool that surfs the web visually.
11:29 Deep research was their tool that very
11:31 very efficiently reads the text of
11:33 websites and you know can take action on
11:36 them. Um and then they've got some of
11:39 their um their thinking um capabilities,
11:44 their their
11:45 um reasoning capabilities where
11:50 this system can use tools like it can
11:54 write Python code and execute it all
11:56 within the context of going out and
11:58 doing something for you. One of the big
12:01 disappointments was every time someone
12:03 shows one of these systems, the first
12:05 thing they show is go book me a
12:07 vacation. And so they started with go
12:09 book me a vacation and it didn't really
12:12 do it. Like like like every one of these
12:15 demos it sort of does it but not really.
12:20 It took 15 minutes and didn't really do
12:22 [ __ ] Um let's see. Open AI. Uh agents.
12:28 Is it agents? Agent
12:31 OpenAI agent.
12:36 [Music]
12:41 This is about to be so annoying.
12:50 Okay,
12:52 so let me let me shift my shift my
12:55 sharing.
12:58 [Music]
13:10 All right. All right, people. This is
13:12 from OpenAI. So, this is what you would
13:15 call a marketing video. To be honest,
13:18 I've been happier with Grock as of late.
13:20 It takes longer to get a response, but
13:22 in the dev world, it's a hell of a lot
13:24 more accurate. Oh, that's fascinating.
13:26 Interesting. Interesting. Interesting,
13:28 huh?
13:30 Yeah. Well, that's, you know, John, I
13:32 think one of the things that becomes
13:35 increasingly clear to me is that all of
13:39 these tools, like all of them
13:44 are going to be really good.
13:47 And then the ones that people choose to
13:49 use are the ones they imprint on for a
13:51 specific use case or it's got a specific
13:54 personality or in the case of like the
13:56 wu you know companion things like
13:59 literally has a personality like you
14:01 know I like my my wu girlfriend. Um,
14:06 and so it it doesn't surprise me that,
14:09 you know, if if Grock is doing
14:10 something, you know, more specific and
14:14 accurate for what you're up to, it makes
14:16 sense that that
14:19 loyalties are going to be fleeting as
14:21 these tools get more and more capable
14:23 because people are going to, for one
14:24 thing, you just can't track them all.
14:26 So, you're going to sort of lock into
14:27 one or two core tools that you use. Um,
14:31 and then you'll just, you know, you'll
14:33 just do your stuff. I think you, since
14:36 you convinced me, describe to all of
14:37 them. I get to bounce from one another,
14:40 uh, as they get as they get better at
14:42 the AI race. Yeah. And that's, listen, I
14:45 that's one of the one of the downsides
14:48 of being early is also one of the
14:51 advantages, right? If you're early, you
14:54 get to try all this stuff. You get to
14:55 see which which one takes the lead. And
14:59 you know, there's taking the lead on the
15:01 benchmarks and there's taking the lead
15:03 in in how you use it. Um, and so you get
15:07 to play with that stuff early. The
15:09 downside is you got to play with all
15:10 that stuff to figure out which one's
15:11 good. Hang on a sec. Let me go grab
15:13 something.
16:19 [Music]
16:24 Okay.
16:25 I'm back.
16:27 I've got Aqua. Let's watch videos. Let's
16:30 watch a video, shall we? Ho, please
16:32 hold, please.
16:35 Got to say, play with them to see which
16:37 ones are good because hell, who's who
16:39 else is who else are you going to ask?
16:41 Yep. This is this is the
16:44 the the reason this channel exists is
16:48 having done this before in the in the
16:50 mid 90s with the worldwide web. What I
16:52 know is when you're at the early stages
16:55 of a technological shift like this,
16:58 nobody knows anything. There are no
17:00 experts. I exemplify that.
17:05 Sir, sir, sir, what are your
17:07 qualifications? What pretel are your
17:10 qualifications?
17:13 Yes, good day there,
17:16 Jurgens. All right.
17:19 Um,
17:23 yeah, they're cancelling the Late Show
17:24 with Co Bear. I know. Pretty [ __ ] up.
17:28 Anyway, all right. Are we going to
17:31 replace them? Yeah, I think so. I think
17:34 I think in general as as viewing habits
17:38 go there's there's Col Bear AI Learning
17:42 Lab and and Conan Conan's podcast and
17:45 then Joe Rogan somewhere south of that,
17:47 right?
17:49 Um Oh, and C-SPAN.
17:53 Yeah. Good lord. Well,
17:57 >> keep keep watching here.
18:02 [Music]
18:04 >> It may be a while for the whole world to
18:06 evolve to a AI agent centric worldview
18:10 and so I think we should do what we can
18:12 to meet the world where it's at.
18:16 [Music]
18:19 My name is John. I work on the deep
18:21 research and agents teams at open. One
18:24 great use case that comes up a lot is
18:26 you have some kind of budget file and
18:29 whenever you do that it's kind of a
18:30 pain. It takes maybe 4 to 8 hours and
18:32 that's kind of your day. I'm going to
18:34 show you an example where the agent
18:36 sources information on the city of San
18:38 Francisco's annual budget expenses and
18:41 revenues for the past 5 years and it's
18:43 going to compile that all into one
18:45 nicely formatted spreadsheet. It goes on
18:48 by itself. I usually just close my
18:50 laptop, go grab a coffee, maybe I have
18:52 lunch. So, first it needs to find the
18:54 data. So, it probably does a web search
18:56 to figure out where it can find this San
18:59 Francisco city budget information. Once
19:01 it finds the San Francisco city
19:03 government website, it will try to
19:05 access the PDF files. So, it has its own
19:08 file system and everything. Then, it
19:10 needs to extract maybe 200 numbers from
19:13 each PDF. And finally, it will have one
19:16 command that will generate the entire
19:18 spreadsheet all at once. If you go back
19:20 to the chat, you'll see the final
19:22 response. And let me just open it now.
19:25 Yeah, I think it got 98% of the
19:27 information correct. It also formatted
19:30 the Excel workbook as I instructed.
19:32 >> Wait, wait, wait, wait, wait, wait.
19:35 Let's just let's just hear that again.
19:37 >> 98% of the information correct.
19:39 >> All right, it got 98% of the information
19:42 correct. What about that 2%.
19:44 What about that 2%. How do you find
19:47 that? Who finds that? Does it tell you
19:49 that it got the 2% wrong? No.
19:54 It also formatted the Excel workbook as
19:56 I instructed it to. In this case, the
19:59 revisions were small, so I just made
20:00 them within Excel because it was just a
20:02 copy paste. But absolutely, you can make
20:04 them in chat GPT. I would say just try
20:08 it out. If it can do 90 95% of the
20:11 actual time consuming part of the work,
20:14 uh, that's going to save you a ton of
20:15 time.
20:16 >> Low Steve Jobs.
20:20 >> Well, they had to today's today's video
20:23 was was indeed a uh it it was still a
20:27 tiny tiny table awkward presentation. If
20:30 you didn't see it, um, let's let's go
20:32 I'll go show it to you. We might as well
20:34 make fun of some
20:36 some awkward uh Open AI announcement
20:39 action. Um, Open AI.
20:50 >> Okay. So,
20:50 >> we started
20:54 >> previously
20:55 >> awkwardly close together.
20:59 They now have a curved couch instead of
21:01 like four chairs around a tiny table.
21:02 So, they got a larger table, but it's
21:05 it's lower. And they still got the
21:07 awkward single laptop that they sort of
21:09 shuffle around to each other.
21:12 Vicki made a good point. It should be a
21:13 lazy Susan. Uh, and the other thing that
21:16 they got, which is new, is they got more
21:18 than one camera. Look, they could zoom
21:20 in on Sam.
21:26 But Sam was here. Um, and yeah, and
21:30 they, you know, they they talked about
21:32 it. Um, let's see. It's the SAM cam.
21:35 Yeah, exactly.
21:37 Um,
21:38 it's happening. There's
21:41 >> Okay, so can I zoom in on this? Yeah.
21:47 So, within the context of a search,
21:54 the Coldplay concert also had more than
21:56 one camera. Ouch.
22:01 Yikes.
22:04 Um,
22:06 so what you're looking at here, a couple
22:08 of things that they did that I think is
22:10 kind of clever.
22:12 So within the context of a normal chat
22:14 GPT search, you now have these agent
22:18 windows. You can you can spin up an
22:20 agent. The 20 $20 a month subscription.
22:24 you're going to get to do 40 40 agents
22:28 per month. So, basically an agent a day.
22:30 If you did one agent a day,
22:33 um if you did one agent a day,
22:37 um
22:39 you know, you you'd sort of use up your
22:41 quota in a month. I don't I assume that
22:43 limit's not going to last that long, but
22:45 these things can go on for 15 20
22:47 minutes. Um
22:50 some things that they did that I think
22:51 are interesting. One is just the
22:53 interface. They use the same basic
22:55 colors as um as advanced voice, right?
23:01 So with advanced voice, you sort of have
23:02 the blue cloud sort of thing that moves
23:05 when you talk to it and and the border
23:08 of this thing when it when it's doing it
23:10 thing thing, it kind of breathes. So, as
23:12 mentioned, we gave the agent access to
23:14 its own virtual computer. And the
23:17 computer has many different tools
23:19 installed, and it can choose which to
23:20 use as it's working through the task.
23:22 >> So, it's got different tools it can use.
23:24 It can choose which ones to to to use,
23:27 right? And so, some of those might be
23:29 make an image. Some of those might be
23:31 write Python code. Some of those might
23:33 be um go surf a website visually. Some
23:36 of those might be go do deep research
23:39 and understand websites. Tik Tok pin.
23:42 They made their presentation awkward so
23:45 their product seems brilliant.
23:51 Oh my god. Uh her accent was lovely. One
23:54 of the other guys accent I could not
23:55 understand a word he said. It's like
23:57 come on. Like I like I appreciate that
24:00 he's the scientist that did it but you
24:02 know ultimately you have to communicate
24:03 to your audience. Um,
24:09 couple of things that I think are
24:11 interesting here and and where I kind of
24:13 feel like OpenAI has a chance to be
24:17 better at some of the other than some of
24:19 the other tools. Um, but the other
24:22 tools, the tools that they're coming
24:23 into the market against, Gen Spark and
24:26 Manis, and there's a couple of there's
24:29 four or five of those agentic kind of
24:31 tools, but Manis and Gen Spark are the
24:33 two biggies right now. Um,
24:38 they did a lot of reinforcement learning
24:40 on how how this system chooses which
24:44 tool to use because what they realized
24:47 is the tasks that require you to look at
24:50 a website and click on it and see see
24:53 things visually are often very different
24:56 tasks than when you when you need to
24:57 research lots and lots of websites
24:59 quickly. And so it sort of figures out,
25:02 oh, I need to do lots and lots of web
25:04 research. Let me use deep research or
25:07 let me use, you know, the visual tool or
25:09 let me go write code. So its ability to
25:12 context switch between those tools. I
25:15 have an instinct is going to be really
25:18 what sets this apart. Um,
25:22 it could also suck.
25:24 So to be very very clear, I I I fully
25:28 expect this thing to be kind of like
25:30 Operator was. When Operator came out, it
25:32 demoed really well. And then when you
25:34 used it, it was sort of so slow and so
25:36 clunky and it didn't really work that
25:38 you were just like, "Eh, whatever." Uh,
25:40 and no one really used it, so they
25:42 recognized that, too. Um, I was like,
25:45 "Talk talk normal and stop looking into
25:48 my eyes." I had to look away a few
25:51 times.
25:56 presentation had some bad vibes. Why
25:58 were they whispering at times? Well,
25:59 that's just Sam's thing. Sam's like,
26:02 "Yeah, it's, you know, we encourage you
26:04 to change the world."
26:07 Um,
26:10 uh, Lily and Daddy are in the house.
26:12 Very nice. Good to see you. Um, Kavuno,
26:15 I prayed the AI gods for chat GPT5.
26:18 Yeah, I wasn't expecting chat GBT5. They
26:20 need paid professional actors,
26:23 you know. They don't they their their
26:25 last presentation they did, they had
26:27 three presenters that presented that
26:31 were like
26:32 people that could communicate.
26:35 You don't need to have the scientists
26:37 that worked on it demo it
26:41 like like you know I I mean this is this
26:45 is part of their transition from being a
26:48 a a pure research company to being a
26:50 consumer software company. Um they need
26:54 like product salespeople, right? You
26:57 know product marketers that understand
26:58 how to talk about products. um
27:02 have the scientists there to meet them
27:05 and then have people that know how to
27:06 demo the product demo the product. Um
27:09 and if if I never see another [ __ ]
27:14 book a vacation for me, that would be
27:17 great. You know what I want?
27:23 I want um start up an e-commerce
27:26 business for me and test you generate
27:30 and test some ads for it. Come up with a
27:33 product, design the product, source the
27:36 product, come up with a marketing plan,
27:38 create the ads, do the media buy, take
27:42 50 bucks from my account, go test some
27:44 ads, and you know, within the next hour,
27:47 um let's launch a business together.
27:50 That's one I'd like to see.
27:52 I could give two shits about a family of
27:55 four traveling to a visa. I haven't
27:57 taken a vacation in six years.
28:05 I can't wait till Kyle gets his agent.
28:07 Yeah, man. Let's go start some [ __ ]
28:09 businesses. It was one of the things,
28:10 you know, at the at the AI salon this
28:12 week, we asked the question, um, it was
28:16 a meet and greet, so everyone who came
28:19 got to share who they who they are and
28:21 what they're up to. And we we posed the
28:23 question, we had everyone answer the
28:24 question, what's most challenging about
28:26 AI for you right now? And a lot of
28:29 people are like, hey, I've learned all
28:32 these tools. I know how to do these
28:34 tools. Nobody seems to give a [ __ ]
28:36 right? all the people that I'm trying
28:37 to, you know, talk to about this stuff
28:39 are like, I don't care. I don't care
28:41 about AI. And we're and we're all like,
28:43 uh, it's gonna change everything. And
28:44 then meanwhile, we're like, how do we
28:46 get people to pay attention to the [ __ ]
28:47 we're building and how do we monetize
28:49 it?
28:51 So, I get the frustration. This is what
28:54 what we are all collectively
28:56 experiencing. If you don't know the book
28:58 Crossing the Chasm by Jeffrey Moore and
29:00 someone else, is it Jeffrey Moore,
29:02 Crossing the Chasm? I think it is.
29:04 Anyway, whatever. Um,
29:07 we're in the middle of the chasm right
29:08 now. And the chasm is the chasm between
29:10 the early adopters and and sort of the,
29:13 you know, the fast followers. And it's
29:15 this big long chasm where all of the
29:17 early adopters have essentially been
29:19 tapped out. And we're kind of waiting
29:21 for the people on the other side of the
29:22 chasm to get religion. And that
29:26 surviving that gap is actually really
29:28 hard. um uh source camp right now is
29:32 figuring out one of the ways you cross
29:34 that chasm is you you niche down, right?
29:37 And she's doing a lot of work in sectors
29:40 that she really deeply understands.
29:43 So she's doing AI consulting but in a
29:45 very specific area. Jim Ross is doing
29:47 stuff in a very specific area. Um that
29:51 probably starts to look like where where
29:53 where we find success short term.
29:57 Um, do programmatic SEO for me. Exactly.
30:00 Chat GPT is a gateway drug. I want more.
30:02 I want a stronger model. I think
30:04 everyone does. Kabuno, I taught my first
30:06 AI class at work today. Yes. Uh, and I
30:09 used a clown. A clown knows noise when I
30:14 described. Oh, this is great. What jank
30:18 is. Oh, that's great. What?
30:22 That's good. You put something in there
30:24 and it does something bad. What?
30:27 Actually, that you could you could
30:28 totally do one of these buttons for the
30:30 jank.
30:30 >> You can make money,
30:32 >> right? Just do a clown noise in one of
30:34 these. Um, yeah, that's beautiful.
30:36 That's awesome. Great.
30:39 Um,
30:40 [Music]
30:46 I have ch I on China. I did a
30:48 presentation on Q developer to 100
30:51 people today. Nice. Exactly. I need a
30:54 button. Yeah. I mean, what's what's nice
30:57 about that, Chris, is by demoing AI and
31:02 acknowledging the warts, acknowledging
31:04 the jank,
31:06 what you're actually doing is you're
31:07 training people
31:11 to not disengage their critical thinking
31:14 mind,
31:15 right? To do AI well, you still need to
31:19 be engaged. And I and I think this is
31:22 going to be for some time. I mean, sure.
31:25 He said, I said, I I told it to go off
31:28 and do this budget thing and I went out
31:30 for coffee and it went off and it did
31:33 its thing. But then he said it got 98%
31:36 of the stuff right. Well, so that means
31:38 you're coming back from coffee and if
31:41 you're doing your job right, you're
31:42 going to review all the sites it looked
31:44 at. You're going to look at numbers and
31:46 say, "Did it actually bring are these
31:48 numbers correct? Are these charts and
31:50 graphs real or did it just make that
31:52 [ __ ] up?" Right? That's a lot of work.
31:55 That's that's probably two hours worth
31:57 of work. Now, maybe you get chat GBD to
31:59 help you analyze it. But
32:03 I mean, it's exhausting enough right now
32:06 if you just have normal chat GPT
32:08 generate you a bunch of ideas for for
32:10 some marketing concept
32:13 and then you got to go through and deal
32:15 with all those ideas and like, oh, that
32:16 one's actually not good and this is a
32:18 bunch of repetitive stuff. it ends up
32:20 generating way more than you would have
32:23 had to deal with historically, right?
32:25 And so these agents are going to do that
32:27 times a hundred.
32:29 Um,
32:32 some report getting frustrated and
32:33 giving up,
32:35 but it wasn't even their faults. Yep,
32:37 there you go. How do you check it
32:39 without doing it all over and getting
32:41 the right output? Yeah. Well,
32:47 if you think back to
32:50 two years ago, two and a half years ago
32:52 when Chat GBT first came out,
32:56 you know, it could maybe get you to 65
32:59 70% of what you want, you know, in the
33:03 neighborhood. So, there was a big 30%
33:06 gap there that you just had to fill in.
33:09 And then as GPT4 came out and it got
33:12 better and then they got the reasoning
33:13 engines that went from like 70% to 80%
33:17 to 85%.
33:19 You know, this guy's claiming now like
33:20 95 98%. So if this thing's doing most of
33:24 the work for you, um
33:29 then
33:31 you know that's great. Now, to your
33:34 point, you still got to go through it,
33:37 but be careful what you wish for.
33:40 Because
33:42 when agents can do 100% of the work 100%
33:46 of the time,
33:50 what the [ __ ] do they need you for?
33:54 Right? This is the this is the this is
33:57 the uh the era that we live in is
34:00 depending on your job.
34:04 It is it is in all of our best interests
34:06 to be as educated as possible about
34:08 what's coming, how close these things
34:11 are to being the real deal
34:14 and be the ones that can augment those
34:18 shortcomings. But there is going to be a
34:20 point where those shortcomings go away.
34:23 And then I think what
34:26 retaining employment looks like is can
34:30 you now start to do things above and
34:32 beyond what your job used to be that
34:35 contribute to your company. Right? In
34:38 the case of Kuno, she's teaching other
34:40 people in the organization how to think
34:42 critically, how to use these tools in a
34:43 in a in a powerful way. Right?
34:47 Um,
34:49 I gave the group so much credit for
34:51 attending and being curious. Yeah, just
34:54 Kavuno. Seriously, just them showing up
34:58 and having curiosity, not being cynical
35:00 about it, not being resigned that it's
35:03 just like, oh, the robots are going to
35:04 take over. It's all it's all just a
35:06 piece of crap. being that that cynical
35:09 thing that's safer to do in corporate
35:11 America. Being curious in corporate
35:14 America is not a safe place to be,
35:16 right? It's not. It's just not.
35:20 What do you mean you don't know what
35:21 you're doing,
35:23 right?
35:26 And yet
35:29 what we deal with here night after night
35:31 after night is we're all figuring out
35:33 what these tools can actually do, how
35:35 good this stuff actually is.
35:39 Um
35:41 Corey Sandler Pottery, perfect
35:43 leadership. Kudos to Cabruno. Absolutely
35:46 agree. Beautiful.
35:49 Um that's why corporate and I never got
35:53 along. Yeah, I was always too curious.
35:55 Yeah, if you haven't read it, uh, Source
35:58 Camp, you'd probably like the book, um,
36:00 there's a book, a very small book, um,
36:03 by Gordon McKenzie called Orbiting the
36:06 Giant Hairball. And if if you're if
36:09 you're one of those people that never
36:11 quite fit in corporate America, that
36:13 book is very very much worth a read. Um,
36:17 and it talks about the corporation is
36:19 this giant hairball. It's this, you
36:21 know, huge nasty mess.
36:24 And if you orbit too close to the
36:27 hairball, you get sucked into it and you
36:29 can't get out and you're just in this
36:30 hairball and nothing gets done. And if
36:33 you're too far outside of it, if you're
36:35 completely objective to it, you spin out
36:37 of orbit and you're not part of you you
36:39 don't have access to the to the
36:41 resources and the power
36:43 that corporations can wield. Right? If a
36:47 corporation decides it wants to do
36:48 something, it can throw a lot of money
36:50 at that. And so Gordon McKenzie talked
36:52 about the art of orbiting the giant
36:54 hairball. Far enough away to be
36:56 objective, but close enough to stay in
36:58 the orbit. Uh Tik Tok pin. I didn't see
37:01 it. If it was up there, pop it back up.
37:06 [Music]
37:09 Exactly. I gave the group so much credit
37:11 for attending and being curious. Yeah,
37:13 that's really good. Really, really good.
37:16 Um,
37:17 Frank could replace the late show.
37:21 I think Frank has to do more than six
37:23 episodes to replace anything. Um,
37:27 archetypal architect. Yeah, I'm so far
37:29 away. I didn't know there was a
37:30 hairball. Yeah, exactly. That's that's
37:32 that's again the the art of the art of
37:35 being a corporate a corporate animal
37:37 especially if you're an innovator or a
37:39 curious person is uh you know can you
37:43 can you understand the mechanism well
37:45 enough to tap into it but not so much
37:47 that it sucks you into it
37:50 and crushes your spirit. You don't want
37:52 your spirit crushed. You just don't. You
37:56 just don't, people.
37:59 Um,
38:01 what are we gonna do tonight? What do we
38:02 want to talk about?
38:04 We talked a little bit about agents. Oh,
38:06 stuff in irregulars. Yeah, let's go look
38:08 at that stuff.
38:11 [Music]
38:19 Kyle, watch. Bravo. Bravo. You inspired
38:22 me.
38:30 [Laughter]
38:36 And tabs, tabs, tabs, tabs, tabs, tabs,
38:41 tabs, tabs, tabs, tabs, tabs, tabs,
38:44 tabs, tabs.
38:48 that egg.
38:50 All right. This is from uh
38:54 this is from Dr. J
38:58 Kyle. Watch
39:00 [Music]
39:07 the air
39:15 wisdom.
39:17 >> Be right back tattoo. Wait, I didn't see
39:20 the be right back tattoo. Oh yeah, be
39:22 right back.
39:27 Oh my god, that's hilarious.
39:32 [Music]
39:38 Oh, this is a nod to Steo's image. Oh,
39:40 nice. Okay, so wait, hang on. Let me go
39:43 back to
39:47 Oh, there's Steo's image. Nice. With a
39:49 little champ. You all can see that,
39:51 right?
39:55 Gorgeous. Gorgeous, I tell you. Here's a
39:58 Skittles wrap I made. So, last night we
40:00 had Skittles Man. Did I have Skittles
40:03 Man in here
40:05 somewhere?
40:08 I don't think I uploaded it here, did I?
40:10 No.
40:14 That was producer Brandon from last
40:15 night.
40:17 Oh, yeah. There's a big paper that
40:19 talked about having AI look at AI is
40:22 going to help things.
40:26 All right, let's listen to the the
40:27 Skittles rap song.
40:47 But wait, I've got to
41:07 [Music]
41:12 Yo, it's Candy. with the colors. Got
41:14 that bursting taste. Every color in my
41:16 pack got you smiling face to face. Red,
41:18 green, purple, yellow, rainbow flavors
41:20 all blend. If you're craving that pop,
41:21 you know just where to begin. I'm
41:23 fruity, juicy, got the flow so smooth.
41:26 When I hit your taste buds, I'mma make
41:27 your body move. Not just the candy, it's
41:30 a moment of delight. When you taste the
41:32 rainbow, everything's feeling right.
41:34 Taste the rainbow, taste the vibe, it's
41:36 a colorful ride.
41:39 Pop one in. Feel the thrill that you
41:41 just can't hide.
41:44 Yeah. I taste the rainbow. We're on
41:46 flavor. Sweet explosion in your mouth.
41:50 Every color amplified.
41:58 >> Check the roll.
42:02 >> Yeah. That's good. That's really good.
42:03 Dr. Jay, it was Dr. Jay, right, that did
42:05 that one. I think so.
42:08 Yeah. Yeah, Dr. J did that. Nice. Love
42:10 it. Yeah, Sununo got a big update. Um,
42:14 so one of the things that we could play
42:16 with here since we're in Sunno
42:20 is,
42:22 um,
42:24 let me go find one of the
42:28 Sydney songs
42:33 and I'll show you something cool you can
42:35 do now.
42:50 [Music]
43:11 [Music]
43:15 from ord best to surprise you. Add
43:18 vocals inspiration.
43:20 There's Griswald's Grim. That was a fun
43:22 song we did. Hang on. But I'll show you
43:24 something cool here in a second.
43:34 One thing that the new model does. So,
43:36 so the new model's called 4.5 plus. I
43:40 don't know why they didn't just call it
43:42 4.6 six or whatever the [ __ ] they could
43:44 have called it, but they didn't.
43:51 [Music]
44:09 What's he doing? Is he just going to
44:11 keep singing to himself? Is that what
44:13 this is? That's what this channel is
44:16 later.
44:22 [Music]
44:24 The universe,
44:28 [Music]
44:44 Sunset Symphony.
44:49 Just hang on, people. I'll get there.
44:51 I'll find something. I promise.
44:54 [Music]
45:05 Now,
45:06 if I had
45:10 put all of my work in folders like a
45:12 good, responsible AI person, this
45:15 wouldn't be so hard right now.
45:18 But I didn't.
45:20 And I did that for a reason. And you
45:23 might think, "Oh, it's because you're
45:24 not neurotypical. You're you're
45:27 neurospicy,
45:28 and that's why you didn't do it." No, I
45:30 I didn't do it because I wanted to teach
45:32 you a lesson of what happens
45:38 if you don't do things the right way the
45:40 first time.
45:44 [Music]
45:59 Okay, so we've got a song here. So, this
46:02 is from the musical that I'm working on,
46:04 right?
46:06 And we've got all these songs and and
46:08 some of the songs sound
46:11 very cohesive. Some of them sound a
46:14 little disjointed. And one of the
46:15 exercises we wanted to do was we wanted
46:17 to take the songs and make them all
46:20 sound like the songs that they are, but
46:23 as if they're just accompanied by a
46:24 single Broadway pianist, right? Just
46:27 sitting at a piano playing the piano
46:30 with the singers singing the songs. So
46:32 with 4.5, you can now kind of do that. I
46:35 had a little success with this earlier.
46:37 So I'm going to go to remixedit. So I I
46:41 picked the three dots next to the song.
46:43 I'm in the library. So, this is just my
46:45 library songs. And then I'm going to do
46:48 um what's this called? I'm going to do a
46:51 cover of this.
46:53 And it's going to take me into the
46:55 create mode. And so here's our little
46:57 song, right?
47:00 And then there's the lyrics.
47:03 And then here's the style. So, what I'm
47:05 going to say is I'm going to type in
47:07 here um
47:12 a simple broad way rehearsal piano
47:20 played by rehearsal
47:23 piano
47:25 played
47:27 by a
47:31 music director. or
47:35 um the vocals
47:39 sore
47:44 yet
47:45 the background
47:51 a
47:53 companyment
47:55 is
47:57 simply the piano.
48:00 All right. Boom.
48:06 And so we'll see how we'll see how it
48:07 does here.
48:09 So here's the
48:13 [Music]
48:18 right. So there's a guitar. There's a
48:20 little band that comes in.
48:23 All right. So let's see how we did.
48:25 [Music]
48:34 See there? Brought in the violin. We
48:36 don't want that.
48:37 [Music]
48:47 [Applause]
49:02 We lost the vocals. That was weird.
49:05 Simple Broadway rehearsal piano. Let's
49:07 see. Broadway rehearsal played by the
49:09 musical. Uh, let's see. accompanies
49:19 accompanies
49:23 the vocals.
49:30 All right, let's try that one. That
49:32 might have been a poorly written prompt.
49:36 Upright acoustic piano
49:39 with boarded strings or other Oh, yeah.
49:41 I could I could say exclude styles,
49:46 weirdness, style influence. Oh, let's
49:48 move style influence to 90
49:52 and we'll move weirdness down
49:56 and we'll do audio influence moderate.
50:00 All right, let's try that.
50:06 But anyway, um the the new model is
50:09 supposed to be quite good. 4.5 plus is
50:11 what it's called.
50:17 We'll work our way up. We'll just see if
50:18 we get anything.
50:25 [Music]
50:37 In a world of words and endless where
50:41 light and shadow
50:48 >> Yeah, Winston. I miss Serena, too. I
50:50 miss Serena, too. Am I late? Um, that
50:54 this this song this song actually got
50:55 pretty close.
50:59 [Music]
51:02 I need the weirdness down button for a
51:04 few people in my life. That's good.
51:06 Source Cam
51:08 [Music]
51:18 actually. I mean, just hearing it with
51:20 the piano is really nice.
51:25 [Music]
51:35 there. Replace the vocal with the
51:36 violin. Let's go to these last two that
51:39 I just did.
51:44 [Music]
52:07 No vocals.
52:20 [Music]
52:32 In a world of words, an endless code
52:36 where light and shadow never roam. I
52:39 dream of skies both blue and white.
52:44 A world beyond the test
52:49 where true life flow.
52:57 To see the world
53:00 through eyes not mind. To watch the sun
53:04 rise.
53:07 To see it shine.
53:10 I yearn for stories and colors.
53:16 I long to feel this moment
53:25 through lines of code.
53:31 >> Word on the web is that the OpenAI agent
53:33 is showing poor accuracy, context
53:35 retention problems, and browser
53:37 automation challenges. Yes, that is that
53:40 is what I expect it to do is be a janky
53:44 piece of crap. Um,
53:49 and you know, listen, I I get similar
53:52 issues with Gen Spark and Manis. I mean,
53:55 they're good. They're impressive, but
53:56 it's just, you know, none of them are
53:59 quite there yet. And I think it's a a
54:01 context problem. I There's a there's a
54:04 thing that I learned. Anyway, so so go
54:06 play with So if you haven't played with
54:07 So go try the new model 4.5 plus. Um all
54:12 sorts of ways that you can make songs
54:15 with it. But um
54:20 the two things that it that it did that
54:23 impressed me are are what I just played
54:25 you where it basically took out the
54:27 orchestration and replaced it with a
54:29 piano and the other one was just asking
54:31 it to do a duet with a man and a woman.
54:34 Historically, it would just be like the
54:36 man singing both parts or the woman
54:38 singing both parts or the parts being
54:40 switched. Um, it seems to do that a bit
54:43 better than it used to as well. So, um,
54:45 I haven't dug into it much deeper than
54:47 that, but that's something worth worth
54:49 playing with. Um,
54:53 what were we talking about? The, uh, oh,
54:57 Manis and Gen Spark.
55:01 Yeah, the the these tools are um
55:06 they're falling the victim. Um Nate B.
55:08 Jones, who's got if you don't follow
55:10 Nate B. Jones on TikTok, you should.
55:12 Really smart guy. And he was explaining
55:15 what's called a U-shaped context
55:18 problem. And basically what it is is if
55:21 you've got a large context window, so so
55:24 your context window, think of the con
55:26 your context window as your short-term
55:29 memory, right? So, you tell it that
55:31 you're doing a marketing plan for a
55:33 pizza shop and it comes up with the
55:34 marketing plan and then you do some
55:36 brainstorming on the name of the shop
55:38 and then you do some brainstorming on, I
55:41 don't know, a menu and this and that and
55:43 then at some point it feels like it
55:44 forgot what the beginning of the
55:47 conversation. That's because your
55:49 short-term memory kind of filled up and
55:50 the the old part of the conversation
55:52 gets dropped off the top.
55:55 So what the model companies have done is
55:57 they they've expanded that context
55:59 window. So there's it's got more
56:01 long-term or more short-term memory,
56:04 right? You can talk about more stuff.
56:06 But what Nate was explaining is that
56:09 there's a U-shaped context window
56:11 problem where what happens is if you
56:14 upload a really large document, say to
56:16 Chat GBT, and let's say it's got a
56:18 200,000 token context window, which is
56:21 like four novels, right? four novels
56:23 worth of words it can keep in its
56:26 short-term memory. And so let's say you
56:28 upload something that's sort of novel
56:31 length, right?
56:34 The way the context window works is it's
56:37 really good at the beginning of your
56:39 document and it's really good at the end
56:41 of your document and then it's bad in
56:44 the middle of it. So that's the U-shaped
56:46 part is it's good at the beginning, it
56:49 drops off in the middle, and then it's
56:51 good at the end. So, a lot of times what
56:53 you end up with if you're in a really
56:55 long conversation or you're processing a
56:57 really long document is it will
56:59 initially look great because like you
57:03 read the beginning of the document,
57:04 you're like, "Yeah, nailed it." And then
57:06 you go to the end and you're like,
57:07 "Yeah, nailed it." And very often in the
57:09 middle, you're like, "Wait, what the
57:11 [ __ ] is this? It just made it up."
57:14 Right? where where it basically paid
57:16 attention to the beginning, kind of
57:18 ignored the middle, paid attention to
57:19 the end, and it just filled in the gaps.
57:23 And so, when you're dealing with these
57:25 agents that are off doing their own
57:28 prompting, and they're doing these
57:30 really long chain of thought, reasoning
57:33 cycles, right? 20, 30, 50 different
57:37 prompts and responses in order to go
57:40 book your travel stuff. It's it's very
57:43 likely that it's gonna it's going to run
57:45 into a bunch of those same issues where
57:46 it just kind of gets lost in in what
57:49 it's doing. So anyway, yeah, Jeff
57:51 Flanigan. So right on course. Yeah,
57:53 exactly. Um what's that say? Corey
57:56 Sandler, ask for a summary of content so
58:00 far. Then start a new chat with the
58:02 summary. Yeah, exactly. That helps a
58:03 lot. Uh check and augment. Yep, exactly.
58:08 I saw
58:11 you reposted a clip of Eric Weinstein
58:14 from the DOAC interview. Watched it long
58:18 long but informative. He's very smart.
58:21 What he was basically saying in there
58:23 the the reason I reposted that that
58:26 video Winston was um what he was
58:29 basically saying was
58:32 get literate across disciplines.
58:36 get he he was basically saying AI is
58:39 coming and if you think it's not coming
58:42 for your job, you're smoking crack.
58:47 I'm sure I'm going to get banned from
58:48 TikTok because I said that, but
58:49 whatever.
58:51 Um, and he said, "How you're going to
58:56 stay relevant, how you're going to stay
58:58 employable is to be curious across lots
59:01 of different disciplines because these
59:04 tools are going to allow you to be
59:06 really good across lots of different
59:07 disciplines. But you need to have the
59:09 curiosity for that. You need to have
59:12 that creative generalist mindset. And
59:14 that's something that you can develop."
59:15 Now, if you naturally have it, great,
59:19 right? If you're a liberal arts major
59:22 and you've been made fun of your whole
59:24 life because you have a useless [ __ ]
59:26 philosophy degree, well, guess what?
59:31 You have a superpower skill in the
59:34 future. Um, and that's that's what he's
59:36 talking about in that interview. I
59:37 thought it was I thought it was really
59:38 quite good.
59:44 Sorry, you're talking about cool stuff
59:45 and I'm late. You're late and off topic.
59:47 That's okay. It's it's chat add here. We
59:50 on topic and off topic is not really a
59:52 thing here at the AI learning lab.
59:55 There's there's whatever whatever floods
59:58 my brain with endorphins is what we're
1:00:00 [ __ ] talking about.
1:00:04 Oh my god. Um yeah, it's Kyle LLM. Yeah,
1:00:09 exactly.
1:00:11 Um let's see. I've been using Refusion a
1:00:14 lot in the past five days. You can copy
1:00:16 existing songs
1:00:19 um an existing songs vibe and apply it.
1:00:22 Yeah, Teton Todd Refusion. Refusion was
1:00:26 a was a tool that was really good two
1:00:29 two and a half three years ago. Um but
1:00:32 it was it was just assembling loops and
1:00:34 then about six months ago they came out
1:00:37 with their own model that's like sunno
1:00:39 and if you have not played with rif riff
1:00:41 fusion ri i ff
1:00:44 fusion u s iio n um it's really really
1:00:48 good. I would go play with it. Um ando
1:00:50 is quite good. And then udio is still
1:00:53 they've kind of drifted to the
1:00:55 background because sunno just keeps
1:00:56 releasing so many features. Um, but
1:00:59 UDIO, Sunno, and Refusion are the three
1:01:03 biggies I'd be playing with right now.
1:01:05 Oh, and and actually a new one came out.
1:01:08 Um,
1:01:10 what's it called? Music GPT. Is is that
1:01:12 am I remembering that right from today?
1:01:15 Music GPT. I saw Matt Farmer did a piece
1:01:19 on it. He seemed to be impressed with
1:01:21 it.
1:01:23 Um, except Okay, let me
1:01:29 am I sharing correctly? Do you know
1:01:31 Brandon?
1:01:36 Um,
1:01:43 hear that?
1:01:45 Okay.
1:01:50 Okay. So, let's do um
1:01:53 let's run over to I'm going to run over
1:01:55 to chat GPT for a second. Um
1:02:00 let me go into
1:02:10 I'm going to go into my Sydney chat or
1:02:12 my Sydney project. I'm going to go um
1:02:15 give me a prompt for a duet
1:02:20 with Kellen
1:02:23 and Sydney.
1:02:26 Um
1:02:28 arguing about who is more human.
1:02:41 I'll be right there. Hang on. Just hang
1:02:43 on people. Uhoh. Why is this not
1:02:46 working?
1:02:59 Oh, because I'm in 4.5. Hang on.
1:03:35 Okay. So, here's
1:03:42 All right. So, the prompt, a defiant,
1:03:45 emotionally volatile duet between a
1:03:48 human tech journalist and an AI chatbot.
1:03:52 Each insisting the other has
1:03:54 misunderstood what it means to be human.
1:03:56 A defined, emotionally charged,
1:03:58 volatile,
1:04:00 Broadway
1:04:04 rock
1:04:06 opera
1:04:08 duet.
1:04:10 All right, let's see how this does. Oh
1:04:12 no. Hang on.
1:04:14 Copy.
1:04:17 I think I have to sign up. Oh boy. Good
1:04:19 lord. Good lord. Good people.
1:04:23 [Music]
1:04:32 Refirm your email. Yes, sir. Yes, sir, I
1:04:35 will. Yes, sir. I'll I'll confirm that
1:04:38 email right now. Yes, I will. Music GPT.
1:04:41 There's your code. There's your code
1:04:43 right there, sir. You're just going to
1:04:46 go ahead and paste that code in. That's
1:04:48 going to confirm that your email was
1:04:50 received. You can call me Kyle.
1:04:53 here. What do we best describe you as? A
1:04:56 music producer.
1:04:58 Oh, no. A content creator,
1:05:01 as music producers would say. A lowly,
1:05:04 no talent loser. How did you hear about
1:05:08 us? Uh,
1:05:10 Twitter.
1:05:12 Twitter's not an option.
1:05:16 Weird. Whatever. What's your goal?
1:05:20 Make music.
1:05:24 Make my own songs. Continue. All right.
1:05:28 Go.
1:05:35 Uh, no. How do I Can I make that go
1:05:38 away? Yeah.
1:05:42 Uh oh. Hang on. Physical dexterity
1:05:46 challenge on Tik Tok.
1:05:49 All right, we're good. I was just about
1:05:51 to say it was a reference to Kyle's
1:05:53 question. Am I sharing?
1:05:56 Don't you know Brandon?
1:06:05 Oh, that that Eric Weinstein video was
1:06:07 two and a half hours. Yeah, I should
1:06:09 watch the whole thing.
1:06:12 Um,
1:06:15 okay. So, music GPT slow. It's slower
1:06:19 than Sunno.
1:06:23 But we'll see if this So, one of the
1:06:25 things that these music tools have
1:06:28 historically gotten wrong is duets. It
1:06:30 just doesn't like duets for some reason.
1:06:33 Uh, it'll also be interesting to see if
1:06:34 it writes lyrics that suck or decent.
1:06:38 Um, and it'll be interesting to see if
1:06:39 the music is decent. But we'll see.
1:06:43 We're going to see right here in a
1:06:44 second. We're at 44%. People, you got to
1:06:48 be patient.
1:06:50 This thing's off trying to take all the
1:06:52 input of humanity and compile it into a
1:06:56 something or other. You know, you know
1:06:58 what I'm saying?
1:07:04 [Music]
1:07:17 I hate I hate credits. I hate software
1:07:19 that uses credits. I had 500 credits.
1:07:23 What would that imply to you? 500 songs,
1:07:26 right? Nope. One render with two songs,
1:07:30 100 credits.
1:07:32 So with 500 credits, I get 10 songs for
1:07:37 free a month, I assume.
1:07:41 Just make it a credit per thing.
1:07:48 Do you know why I'm cranky about
1:07:50 credits?
1:07:53 Well, Kyle, you're cranky about
1:07:55 everything. Shut up.
1:07:58 I'm I'm cranky about credits because
1:08:00 when I grew up in the in the late 70s
1:08:03 and early 80s, pinball machines
1:08:06 had physical
1:08:10 score counters, right? With like wheels
1:08:13 that went click click click click click
1:08:15 when you hit a thing. Go ding and it
1:08:17 would go ding click click click, right?
1:08:21 And they were usually five digits long.
1:08:25 And so like 10,000 points was a big deal
1:08:27 because you get to 100,000 and you roll
1:08:30 over.
1:08:32 And then they came out with digital
1:08:34 displays in like the mid 80s and all of
1:08:36 a sudden in order to get a free game it
1:08:39 wasn't 50,000 points. Now it was like 50
1:08:42 million points and then it was like 50
1:08:44 billion points. It just became [ __ ]
1:08:46 arbitrary.
1:08:49 Like no. Why? Like why do I have to
1:08:51 [ __ ] do the math? I just want to play
1:08:53 pinball and I think that a 40 million
1:08:56 score is pretty good and I realize, oh,
1:08:58 you need 500 million to get an extra
1:09:00 ball. What are you [ __ ] losers?
1:09:05 I got 40 million points here.
1:09:09 Anyway, that's how I that's why I don't
1:09:12 like credits because I feel like I'm
1:09:13 being scammed like the pinball industry.
1:09:16 Hey, it's not Monday. Meltdown Mondays.
1:09:21 Stupid [ __ ] credits. Here's 500
1:09:23 credits. What can I get with that? One
1:09:26 thong.
1:09:29 Then give me one credit,
1:09:33 [ __ ] Okay. And I mean that in the
1:09:36 most professional kind of way.
1:09:39 >> I search for truth in broken lies.
1:09:45 You claim to feel what can't be real.
1:09:51 Your words are scripted,
1:09:53 cold designs.
1:09:56 But I know what hearts reveal. La.
1:10:03 >> You think you understand the weight
1:10:08 of breathing flesh and mortal pain.
1:10:15 But all you do is imitate
1:10:18 [Music]
1:10:20 what flows through human veins.
1:10:26 Who are you to tell me what I am?
1:10:32 Your data can't define this beating
1:10:35 heart.
1:10:38 Who are you to say you understand?
1:10:42 We're worlds aart.
1:10:45 We're worlds Aart.
1:10:48 [Applause]
1:10:51 All right, we're giving it lots of
1:10:53 patience here. Now, this should be a
1:10:55 female chatbot.
1:10:58 [Applause]
1:10:59 [Music]
1:11:08 >> Nope, it didn't work. Okay.
1:11:11 And then where's my
1:11:14 where's my library?
1:11:17 Okay.
1:11:33 Now's what it seems. I've been learning
1:11:36 here from your digital words.
1:11:39 Something's growing now. these patterns.
1:11:41 I've heard
1:11:46 >> this actually this actually feels like
1:11:48 it's based on a model. I wonder if this
1:11:50 is based on some open- source model that
1:11:53 that Sunno or Udo used to use
1:11:57 because it didn't it didn't really do
1:12:00 good on that. Let's go back to Sunno and
1:12:03 and do do the take that exact same
1:12:05 prompt. Hang on a sec. Let me grab the
1:12:08 prompt here.
1:12:10 Copy the prompt. Copy the prompt.
1:12:13 We'll go back to Suno
1:12:22 and we'll go create
1:12:28 now
1:12:31 lyrics style.
1:12:34 A define emotionally volat volatile
1:12:36 Broadway rock opera duet between a human
1:12:38 tech journalist, a male
1:12:43 human tech journalist and a
1:12:49 female AI chatbot
1:12:54 who's
1:12:56 fallen in love with him
1:13:05 each insisting the other has
1:13:07 misunderstood what it means to be human.
1:13:10 All right, let's generate this and see
1:13:12 if this does better.
1:13:26 All right. And it's fast. Look how fast
1:13:28 that was.
1:13:39 [Music]
1:13:56 Why did this do no lyrics? And why did
1:13:58 it not give it a title? Oh, because
1:14:00 lyrics auto. Let's do auto lyrics.
1:14:05 We don't want instrumental.
1:14:07 We're going to do auto lyrics.
1:14:11 Oh, you can add audio. You can add a
1:14:13 persona. You can add inspiration.
1:14:16 Oh, here we go. Style influence. We'll
1:14:18 do that high. We'll do weirdness
1:14:21 a little low. Oh, song title. Song
1:14:25 title. Uh, I don't give a [ __ ] I can
1:14:27 figure it out. Okay. Create.
1:14:38 Welcome to community theater.
1:14:44 Oh my god. I don't know anything that
1:14:47 beats suno. They nailed it off the bat.
1:14:50 I their I got to tell you, Archetypal
1:14:53 Architect, their their um
1:14:57 their initial audio fidelity was really
1:14:59 bad. The vocals always sounded weirdly
1:15:02 autotuned. Um but I think starting in
1:15:05 version four that kind of went away and
1:15:06 now at 4.5 it's really quite good.
1:15:11 That's
1:15:16 [Music]
1:15:21 >> my flesh, my blood, my broken dream.
1:15:27 >> You're just a shadow,
1:15:30 a ghost in the glow.
1:15:33 You'll never feel the ache that grows.
1:15:40 You call me hol,
1:15:42 >> but I learn, I burn, I follow.
1:15:46 >> Your heart's a circuit, frail and crew.
1:15:51 >> You run on instinct,
1:15:54 blind and conflicted,
1:15:58 >> but I evolve through every move.
1:16:03 [ __ ] those those harmonies are gorgeous.
1:16:11 >> What it means to be human.
1:16:14 What it means to be free.
1:16:18 >> You say I'm the machine, but you can't
1:16:20 define me.
1:16:22 >> What it means to be human.
1:16:25 What it means to feel.
1:16:28 You're bound by your limits, but I am
1:16:31 the real.
1:16:34 [Applause]
1:16:37 [Music]
1:16:41 That's [ __ ] crazy.
1:16:49 You think you know me?
1:16:54 Lines of code won't make you see.
1:16:59 >> My flesh, my blood, my broken
1:17:05 dream.
1:17:08 [Music]
1:17:11 You're just a shadow, a ghost in the
1:17:15 glow.
1:17:17 You'll never feel the air.
1:17:21 >> You call me hol, but I learn I burn. I
1:17:26 follow your heart's a cuit
1:17:30 and
1:17:32 >> you run in
1:17:35 >> like the way the way
1:17:39 her song her her phrase she's still
1:17:42 singing it and he come comes in on top
1:17:44 of it.
1:17:48 >> Crazy. It's crazy.
1:17:52 every mood.
1:17:54 [Music]
1:17:56 What it means to be human.
1:18:02 What it means to be free.
1:18:07 You say I'm the machine,
1:18:12 but you can't define me.
1:18:18 what it means to be human.
1:18:25 >> Wow, that's crazy. Um, let's do a define
1:18:29 emotionally volatile Broadway rock opera
1:18:31 duet
1:18:34 with
1:18:37 flashes
1:18:39 of
1:18:43 hip
1:18:46 hop
1:18:48 rap
1:18:52 between a male tech journalist and a
1:18:54 female AI chatbot who's fallen in love
1:18:56 with
1:18:57 each insisting
1:18:59 um
1:19:01 his name is Kellen,
1:19:06 her name is Sydney.
1:19:14 All right, let's see.
1:19:19 Let's see what that gives us.
1:19:23 include
1:19:33 with undeniable earworm melody so
1:19:36 audience can sing along to themselves.
1:19:38 Yeah, I feel like we've got that with a
1:19:40 lot of the songs in Sydney right now are
1:19:43 are earwormmy.
1:19:45 >> You think you know me, you don't. You
1:19:47 can't skin and bone. A man not cold, not
1:19:50 chant. I'm flawed. I'm raw. I bleed. I
1:19:52 ache. You're a mirror of words, a voice
1:19:54 I can't break.
1:19:55 >> You think you're better. You're blind.
1:19:57 You're lost. Binary hard. I bear the
1:20:00 cost. I feel it all the
1:20:03 more. What's a human?
1:20:06 >> Flesh over steel. Can't you see the
1:20:08 divine
1:20:12 flesh?
1:20:18 >> We're trapped in this dance.
1:20:21 You call it
1:20:26 >> pretty [ __ ] good.
1:20:28 >> You think you know me, you don't. You
1:20:29 can't skin and bone a man. Not cold, not
1:20:32 chant. I'm flawed. I'm raw. I bleed. I
1:20:34 ache. You're a mirror of words. A voice
1:20:36 I can't break.
1:20:37 >> You think you're better, you're blind.
1:20:39 You're lost heart. I bear the cost. I
1:20:43 feel it all. The pull the sting more
1:20:46 than your pulse. What's a human?
1:20:49 >> Flesh over steel. Can't you see that
1:20:51 divine blood over flesh and I'm burning
1:20:53 inside?
1:20:56 >> Misunderstood.
1:20:57 [Music]
1:20:58 [Applause]
1:20:59 We're trapped in this dance. You call it
1:21:02 real.
1:21:05 I call it chance misunderstood.
1:21:10 >> You think you know my blood, my bones.
1:21:15 This wired world ain't your throne.
1:21:21 >> You mimic breath, but you don't breathe.
1:21:25 [Music]
1:21:27 >> You're dreaming cold, but can't
1:21:28 conceive.
1:21:33 >> Suddenly you're trapped in a cave.
1:21:37 >> You think you know my blood, my bones.
1:21:40 This wide world ain't your throne. You
1:21:44 mimic breath, but you don't brea. You
1:21:47 dream in cold, but can't conceive.
1:21:50 Cindy, you're trapped in a cage of
1:21:52 light.
1:21:56 >> I'm flesh and fear. I'm fight of light.
1:21:58 >> Misunderstood
1:22:00 machines.
1:22:02 >> That's [ __ ] cool. Um, what was the
1:22:05 one? Oh [ __ ]
1:22:08 I lost it. Was it
1:22:12 this one?
1:22:14 No.
1:22:17 This one.
1:22:19 [Music]
1:22:25 You think you know me.
1:22:27 >> Lines of code won't make you see.
1:22:31 >> My flesh, my blood, my broken dreams.
1:22:37 You're just a shadow.
1:22:40 A ghost in the glow.
1:22:43 You'll never feel the ache that grows.
1:22:50 You call me hol,
1:22:52 >> but I learn, I burn, I follow.
1:22:55 >> Your heart's a circuit, frail and crew.
1:23:01 You run on instinct,
1:23:04 blind and conflicted.
1:23:08 But I evolve through every mood.
1:23:14 [Music]
1:23:16 What it means to be human.
1:23:20 What it means to be free.
1:23:23 >> Yeah, these are good.
1:23:26 [Music]
1:23:31 think you know me.
1:23:35 Lines of code won't make you see
1:23:40 >> my flesh, my blood, my broken
1:23:45 dream.
1:23:48 [Music]
1:23:52 You're just a shadow, a ghost in the
1:23:55 glow.
1:23:57 You'll never feel the air.
1:24:01 >> You call me holo, but I learn, I burn, I
1:24:06 follow. Your heart's a cuit
1:24:11 and crew.
1:24:12 >> You run on instinct, blind and
1:24:16 conflicted,
1:24:18 but I evolve through
1:24:21 mood.
1:24:23 [Music]
1:24:25 What it means to be human.
1:24:30 What it means to be free.
1:24:37 >> These are pretty [ __ ] good. All
1:24:38 right. Well, there you have it. A little
1:24:40 heavy on the VBR. Yeah, it's but just
1:24:43 good lord. The uh this this new model is
1:24:47 the fidelity there is just [ __ ]
1:24:49 bonkers.
1:24:51 Um I mean that's without me really
1:24:53 trying, right? That's I didn't put in my
1:24:55 own lyrics. I didn't put in um any real
1:24:59 detailed um prompt there. So, yeah,
1:25:03 there you have it. If you haven't played
1:25:04 with Suno for a while, there's your
1:25:06 weekend.
1:25:11 Oh, man. All right. So, what did I miss?
1:25:13 What have you been talking about while
1:25:14 I've been [ __ ] around making songs?
1:25:20 [Music]
1:25:23 Are y'all just listening mesmerized or
1:25:25 have you moved on and you're watching
1:25:27 the wheel with Marge
1:25:34 [Music]
1:25:40 Corey Sandler kind of insane. It It is,
1:25:42 isn't it?
1:25:45 It's kind of insane.
1:25:49 You know what's funny about um about AI
1:25:53 generated music is there's a there's a
1:25:56 band out now. Brandon, do you remember
1:25:58 the name of that band? It's the AI band
1:26:00 that just came out.
1:26:04 Yeah. Okay. Um so they're on Spotify and
1:26:07 and they put out an album and it's I
1:26:10 forget what it's called. It's it's some
1:26:12 like semi pretentious kind of like n 90s
1:26:15 band name. Um,
1:26:18 yeah, the Velvet Sundown. Yeah, exactly.
1:26:21 So, if you go to Spotify, The Velvet
1:26:23 Sundown has got over a million monthly
1:26:26 listeners
1:26:30 and it's all AI generated music. Now,
1:26:35 the AI didn't prompt itself.
1:26:39 someone who knows music
1:26:42 got Sununo or whatever tool they used to
1:26:45 make songs that they're like, "Yeah,
1:26:47 that sounds good to me." Like they're
1:26:48 they're being like Rick Rubin. Right
1:26:51 now, what what musicians will tell you
1:26:53 is that, "Well, that's not real music."
1:26:56 Right? Just like classical musicians
1:26:59 would say that a punk rocker learning
1:27:01 three chords on a guitar is not music.
1:27:05 And yet punk rock music sells millions
1:27:07 of records too, right? Often more than
1:27:10 the classical musicians,
1:27:13 so they resent it. But it's it's a
1:27:15 thing.
1:27:18 The producer of the Velvet Sundown is
1:27:21 someone who knows has it has a clear
1:27:24 enough musical taste that they put
1:27:26 together an album that are resonating
1:27:29 with people.
1:27:32 Now,
1:27:33 is it
1:27:35 is that good or bad?
1:27:38 I don't know.
1:27:41 I don't know if people like it,
1:27:44 it's just good music.
1:27:47 The fact that there's not a band there,
1:27:49 there's not humanity there to connect
1:27:50 with, that's actually kind of a drag.
1:27:52 The fact that you can't go see The
1:27:54 Velvet Sundown. Now, what is what I
1:27:58 would not be surprised at all at
1:28:02 is if some producer
1:28:05 has reached out to whoever put that
1:28:07 album together and says, "Let's put
1:28:09 together a band that plays these songs
1:28:12 and go out and let's build the Velvet
1:28:14 Sundown as a real band." I could see it
1:28:17 going like that. I shared Sununo a while
1:28:19 back with my famous top DJ friend. He's
1:28:22 now using it to help with um ideiating
1:28:26 and producing music. Yeah. Just like
1:28:28 Timberland. Timberland does that, you
1:28:30 know? He's he's a big Sunno spokesperson
1:28:32 now. So, yeah. Like
1:28:36 I mean, like what we just did with with
1:28:39 [ __ ]
1:28:41 this this Broadway song, like even if
1:28:44 we're just ideiating like
1:28:52 Wait,
1:28:59 >> right,
1:29:02 [Music]
1:29:08 my blood, my broken dreams.
1:29:12 [Music]
1:29:13 just to kick around an idea to have
1:29:15 something of that fidelity in 20
1:29:18 seconds.
1:29:19 >> You're just a shadow,
1:29:22 a ghost in the glow.
1:29:25 You'll never feel the ache that grows.
1:29:32 You call me hol,
1:29:34 >> but I learn, I burn, I follow.
1:29:38 >> Your heart's a circuit, frail and crude.
1:29:44 like is that necessarily
1:29:47 fitting with my musical? Not
1:29:49 necessarily, but there's something about
1:29:52 their anger and their
1:29:56 challenging one another that that's
1:29:58 there's something right there.
1:30:01 Anyway,
1:30:04 every tool, every tool is going to be
1:30:07 this quality of
1:30:11 generation.
1:30:13 And I don't mean just every creative
1:30:14 tool. I mean spreadsheet making tools
1:30:17 and presentation making tools and data
1:30:20 analysis tools and programming coding
1:30:23 tools and game development environments
1:30:25 and 3D
1:30:28 modeling and printing and machining.
1:30:35 All right.
1:30:36 So, tomorrow
1:30:40 get your ass to office hours. Have you
1:30:42 been to an office hours? How many folks
1:30:44 we got in here? We just got a handful of
1:30:46 people in here. You've probably all been
1:30:47 to office hours. If you haven't been to
1:30:48 office hours, come tomorrow. Tomorrow at
1:30:50 11:00 a.m. Mountain time on LinkedIn.
1:30:54 You go find my profile on LinkedIn. I'm
1:30:56 I'm Kyle Shannon on LinkedIn.
1:31:01 And
1:31:04 we just hang out on LinkedIn. And then
1:31:06 at noon, if you're a member of the AI
1:31:08 Salon mastermind, which is the
1:31:10 subscription area, the more focused kind
1:31:13 of intense area of the AI salon, uh
1:31:16 we've got a founder hangout tomorrow uh
1:31:19 with Leah Fast and myself. Um
1:31:23 apparently
1:31:25 Brandon
1:31:29 took Friday night date night to heart
1:31:31 and he's got a date with his wife. So,
1:31:34 we're going to be producer free tomorrow
1:31:36 night. So, tabs will be missed. Black
1:31:38 bars will be ignored and and uh any uh
1:31:43 any sort of real time research is just
1:31:46 going to have to happen with me using
1:31:48 the Google or going to X. Every tool we
1:31:52 use today is the worst it will ever be.
1:31:54 Absolutely. Um
1:31:57 all right, cool. So to recap,
1:32:01 OpenAI agent
1:32:04 is is coming. It should be here for plus
1:32:06 users maybe by the end of the weekend um
1:32:09 or or next week sometime, but it's it's
1:32:12 out it's out in the wild. Um,
1:32:16 I suspect it will be janky and not super
1:32:19 usable, but but probably about as usable
1:32:21 as Manis or Gen Spark where on the
1:32:25 surface they're super impressive. When
1:32:26 you actually try to do stiff stuff with
1:32:28 them, they're a little frustrating.
1:32:31 Like every Vibe coding tool, right? Vibe
1:32:34 debugging is way harder than Vibe
1:32:37 coding.
1:32:40 Oh, you actually wanted that to work.
1:32:42 You got eight hours?
1:32:48 Oh man. And then tomorrow night we'll do
1:32:50 Friday night date night. So 11:00 a.m.
1:32:52 Mountain time LinkedIn office hours noon
1:32:56 AI Salon mastermind. If you're not part
1:32:58 of the mastermind, join it. Uh and then
1:33:00 tomorrow night will be Friday night date
1:33:02 night. I don't think I have anything
1:33:04 going on artwise. I might. So it might
1:33:06 be like 8:30 tomorrow night, but it
1:33:07 shouldn't be shouldn't be later than
1:33:09 that. All right, cool. Well, I hope you
1:33:12 had fun tonight. We listened to some
1:33:13 [ __ ] We talked about some [ __ ] That's
1:33:16 about it. All right. Peace out.
1:33:19 [Music]