Why Can’t Making Music Be Like Building with Blocks?

00:01 Long ago,

00:02 I came across some posts on Xiaohongshu.

00:05 They said, "I always feel I'll get there,"

00:09 "but I don't know when I will."

00:14 I think I have

00:16 similar thoughts myself.

00:19 Like,

00:20 I always feel that I'll become

00:24 someone remarkable,

00:25 but I don't know

00:27 when I'll become

00:28 that remarkable person.

00:30 I don't really know how to

00:34 explain this kind of thought,

00:37 or state of mind.

00:41 For example,

00:45 just before Christmas last year— I think it was Christmas Eve—

00:55 I saw a post on X from Replit's

01:01 product lead, I think.

01:03 Their post said

01:09 the first specified number of commenters

01:11 would get some Replit credits.

01:16 My comment happened to fall

01:21 within the number they had set.

01:24 And then,

01:26 I received two months'

01:29 worth of Replit credits.

01:32 Um, two months of credits, yes.

01:37 I think it was $25 a month,

01:40 so $50 for two months.

01:42 And then,

01:42 in the first month, I used it

01:49 and spent the entire $25.

01:51 I built a very, very small app.

01:55 I was still quite new to vibe coding,

02:00 and at the time,

02:01 I didn't know much about agents.

02:05 I can't remember whether I had

02:09 started using Antigravity yet.

02:13 Anyway, I tried Replit

02:17 and had its agent build things for me.

02:21 It was also around then

02:22 that my design style truly took shape.

02:28 Oh, I really loved the webpage

02:29 it designed for me back then.

02:32 I was building a website

02:34 for matching people with careers.

02:37 The basic idea was this:

02:38 if you wanted

02:40 to work in a certain profession,

02:43 you might want to know whether

02:47 your experience and abilities fit it.

02:53 So,

02:53 I built a site to assess that fit.

02:56 It would also give you questions

02:59 to answer.

03:00 After you answered them,

03:01 it would provide an analysis.

03:03 In the end, everyone using the tool

03:07 would receive a report.

03:09 That report took the form

03:13 of a webpage.

03:14 It also included data visualizations.

03:19 Anyway, the agent designed it for me,

03:21 and I absolutely loved its colors.

03:24 There is a color palette I've loved

03:28 since 2021 or 2022,

03:34 when I designed a logo for myself.

03:37 That logo used purple,

03:40 and that particular shade of purple—

03:44 is "value" the right word?

03:46 Its RGB values

03:48 were a fixed set of numbers.

03:50 I really love that color.

03:53 Anyway,

03:55 I truly loved the site

03:58 that this agent designed for me.

03:59 The tool looks pretty rough,

04:03 but it was also

04:06 one of my more substantial projects

04:09 after discovering vibe coding.

04:15 If someone really

04:18 took the time to use this product,

04:20 they'd see that I put

04:22 so much thought into it,

04:23 both in the UI design

04:26 and the UX.

04:30 I really put a lot of care into both.

04:33 And of course, I'm grateful to the agent

04:36 for all its early assistance.

04:39 Later, while I was building the site,

04:43 we went through several rounds.

04:46 Anyway, with that $25,

04:48 I hadn't even finished the tool

04:52 before all $25 was gone.

04:55 So I took the next step:

05:01 I downloaded all the code

05:04 I had generated on the platform,

05:06 every last bit of it.

05:08 Then I moved it into VS Code.

05:10 I had Claude write code for me,

05:14 and then

05:14 I edited the code in VS Code.

05:19 I didn't yet know how to use Claude Code.

05:23 I don't think the desktop app existed yet.

05:26 I can't remember if it did.

05:27 Either way, I wasn't using it.

05:29 Everyone seemed to use Claude Code

05:31 through the terminal,

05:33 and I didn't understand terminals.

05:37 So I never used it.

05:38 Instead, in Claude's chat interface,

05:43 I described the changes I wanted,

05:45 and it revised the code for me.

05:47 Then I made the edits in VS Code.

05:51 That's how I completed the tool.

05:56 After using Replit through that process,

05:59 I felt

06:03 it suited beginners who wanted to make something.

06:07 The problem was getting the details right.

06:10 That drove me crazy.

06:13 For example, changing a color,

06:16 or changing the size of a box,

06:22 or fine-tuning the text layout—

06:25 I had to adjust things over and over,

06:28 especially the typography.

06:30 Every person who used the tool

06:33 would leave with their own webpage.

06:35 And on that webpage,

06:36 how should all the content be arranged?

06:39 What should the text formatting be—size,

06:43 font,

06:44 and the spacing between visualizations

06:47 and the text?

06:50 How should the main heading,

06:53 subheadings, and body text be placed?

06:56 For every one of those details,

06:57 I had to tell the agent again and again

07:02 exactly what result I wanted.

07:04 It would then

07:05 use my natural-language description

07:07 to revise it.

07:08 But each revision

07:10 still wasn't quite satisfactory.

07:12 So we went back and forth,

07:13 revising it countless times.

07:16 Even after the $25 ran out,

07:18 those details still weren't right.

07:20 So I pulled down the code

07:22 and asked Claude to revise it.

07:25 That led to one

07:27 important piece of feedback:

07:32 if

07:35 I could manually adjust

07:37 the interface generated by the agent—

07:39 things like its size,

07:41 its colors,

07:42 and its layout.

07:43 That would be wonderful.

07:45 But since

07:47 those Replit credits

07:51 came from Replit's product lead,

07:54 I was going to post

07:56 feedback on my experience on X.

07:58 So in that feedback,

07:59 I mentioned this exact point:

08:01 if

08:03 the code generated by the agent

08:07 allowed me to edit it

08:09 through the interface it created—

08:15 say it generated a webpage,

08:18 and I could edit things on that page,

08:20 like the size of a dialog box,

08:25 or its colors

08:27 and other details,

08:28 that would be wonderful.

08:29 And I remember that

08:31 in my feedback, I said

08:33 I thought this should be

08:36 vibe coding's future.

08:39 Agents wouldn't merely

08:41 generate code for you;

08:42 we humans could also

08:44 manually adjust their interfaces.

08:49 At the time, I simply

08:50 felt certain this would become

08:52 a trend in AI coding.

08:56 Then, not long ago,

08:58 Replit released Agent 4,

09:03 and it offered exactly this feature.

09:07 Connecting those two moments,

09:10 I think I do have

09:11 some foresight.

09:13 After gaining deep hands-on experience,

09:19 I can anticipate

09:20 how the product should evolve

09:21 to become more human-friendly,

09:23 and which direction it should take.

09:27 Because I've always felt that

09:30 whether it's today's AI coding,

09:33 AI music generation,

09:37 or AI video generation,

09:40 human control

09:41 absolutely cannot be left out.

09:43 We need to return aesthetics, judgment,

09:47 and taste to human beings.

09:50 And humans express these things

09:53 through more than language.

09:55 There can also be

09:56 other ways to express them.

09:59 If we want to include

10:03 as many of those ways as possible,

10:07 then there must be

10:10 some point of entry

10:12 where human participation is needed.

10:15 With Replit Agent 4,

10:18 you can change colors

10:20 and other things in the interface.

10:22 That kind of design

10:23 is essentially

10:24 the point of entry I just described.

10:26 It gives people—

10:28 it gives humans

10:29 more manual control.

10:34 I've thought about another question.

10:36 I'm a music lover myself,

10:40 and I'd really like to make

10:45 some music of my own.

10:46 I once tried FL Studio, a music—

10:54 uh,

10:57 is it music editing software?

11:00 I'm not entirely sure either.

11:02 Anyway, back then,

11:03 I had no background in music theory,

11:05 so I had no idea what to do.

11:07 I looked for lessons on Bilibili,

11:11 um,

11:12 tutorials to watch.

11:15 But I just couldn't do it.

11:18 I couldn't learn it.

11:19 It felt so difficult.

11:20 Why was it so hard?

11:22 Now that vibe coding,

11:23 or AI coding, is so popular,

11:25 I wondered:

11:26 why isn't there a music app

11:32 that works like—

11:35 like building music with blocks?

11:39 Because making music, well,

11:41 if you look at its musical elements,

11:45 it involves many different blocks:

11:47 drums,

11:49 piano,

11:51 vocals,

11:51 guitar,

11:53 bass.

11:54 Each of those sounds is like

11:57 an individual block.

12:00 Suppose it took these tangible

12:04 instrumental elements,

12:06 along with vocals

12:07 and post-production effects,

12:11 and visualized them all as blocks.

12:14 If I could drag those blocks around

12:17 and arrange how they appear

12:20 to make music,

12:24 then for someone with no background in music theory,

12:27 wouldn't that be

12:29 an easy way to create music?

12:33 It may go back to what I once said:

12:35 today,

12:37 everyone can shoot videos,

12:39 everyone can take photos,

12:41 and everyone can write.

12:43 So why can't

12:44 everyone make music too?

12:47 What if we could make music

12:50 something people could

12:53 easily reach in everyday life?

12:56 Right now, everyone who engages with it,

12:59 apart from the music creators

13:02 who actually make music,

13:04 is simply a consumer.

13:09 A product is only truly democratized

13:14 when its audience can be producers

13:15 and consumers.

13:18 Only then is it truly accessible to all.

13:23 Now that AI is so powerful,

13:26 people say knowledge has already

13:29 been democratized.

13:30 The difference is whether people seek it

13:33 and ask for it.

13:35 I think music is

13:36 the same kind of thing.

13:38 I have this need,

13:40 so who can help me fulfill it?

13:44 I actually tried building an MVP,

13:47 but I had no musical training

13:51 and didn't know what music theory meant.

13:54 So what I made—

13:56 I can describe

13:58 what it felt like

14:00 the first time I used my MVP.

14:03 I built that MVP with Codex.

14:06 Admittedly, Codex isn't as good

14:08 at coding as Claude Code.

14:11 But when I opened

14:12 the local preview,

14:15 before touching anything,

14:17 I simply clicked the play button.

14:19 And then the sound came out—wow.

14:21 In that moment, my God.

14:25 Then

14:25 I adjusted some blocks it had designed,

14:30 along with some buttons.

14:33 Every adjustment

14:35 changed the music along with it.

14:38 And I thought,

14:39 It felt so rewarding.

14:40 Even though it was a terrible prototype,

14:42 I could still feel

14:44 that my happiness was real in that moment.

14:47 So if a company

14:50 really built such a beginner-friendly,

14:53 building-block-like

14:55 tool for making music,

14:57 then for people like me with no background in music theory

15:00 who still love music,

15:02 it would be an absolute treasure.

15:06 I really, really, really, really

15:09 hope a product like that appears.

15:14 Mm.

15:16 What was I talking about at the start?

15:22 I've already forgotten.

15:23 All right, I've talked for quite a while.

15:26 Let's wrap it up here.