No Matter How Powerful AI Gets, Humans Still Need Their Hands Back
Recorded in Chinese
Continuing from my experience with Replit, I consider why people need direct control over AI-generated interfaces and music, including judgments words cannot fully express. I offer possibilities for these tools while leaving the future open.
00:00 Continuing from the previous recording,
00:05 I lost track while I was talking
00:07 and forgot my original point.
00:10 Now I want to add it back in.
00:13 At the start of the last recording,
00:16 didn't I say I thought that one day
00:18 I'd become someone remarkable?
00:19 Then I looked for an example to prove it.
00:22 I already talked at length
00:24 about that example in the last recording.
00:27 So let me pick it back up.
00:29 It was about Replit.
00:32 The first time I tried Replit,
00:34 I gave feedback along these lines:
00:38 if, in the future,
00:41 when we use AI coding,
00:43 the interfaces it generates
00:45 let people drag things manually,
00:49 and not just drag them,
00:51 but change their colors,
00:52 resize them,
00:56 and do all the other things
00:57 that humans can do by hand,
01:01 if that became possible,
01:03 that would be ideal.
01:05 I also said in my feedback
01:07 that I thought it was a future trend—
01:10 the direction things would move toward.
01:12 Then Replit Agent 4 came out
01:16 and actually implemented
01:18 the exact feature I had wanted.
01:21 So I think this example shows that,
01:24 after I immerse myself
01:27 in experiencing something,
01:30 I can anticipate
01:32 how it would need to evolve
01:34 in order to become better.
01:37 It's like—
01:40 when I had just graduated,
01:45 I strongly resisted entering education.
01:47 Because I felt
01:49 so unhappy when I was a student,
01:52 so joyless,
01:53 I didn't want to participate
01:55 in that same system and witness
01:59 students being unhappy and joyless.
02:01 But by a twist of fate,
02:03 I really did enter the field.
02:04 And once I did, I realized
02:06 I was very good at this kind of work.
02:08 I could spend a long time
02:11 immersed in that system
02:12 and sense
02:15 what was making students so unhappy.
02:16 In class, I would try
02:18 to use my own approach
02:20 to bring them
02:23 some joy
02:26 and happiness.
02:30 Okay, let's wrap up that topic
02:32 and return
02:36 to what I was saying about Replit.
02:38 I think that
02:43 if AI is involved in life or work,
02:48 human control
02:49 will inevitably be present
02:52 and must be intentionally designed in.
02:58 In the previous recording,
02:59 I mentioned music.
03:02 I said I hoped for a tool
03:06 that would allow every music—
03:10 every music lover
03:12 to build music with blocks,
03:14 or like playing a game.
03:17 This also illustrates that—
03:21 aren't there now many
03:23 well-known music-making products?
03:26 For example, Suno.
03:28 Maybe not "software."
03:29 A tool—or product.
03:32 For example, Suno,
03:33 or other tools and products.
03:36 But I think they're missing something.
03:39 Take Suno, for example.
03:42 If you want to change part of the music,
03:44 you can only describe it in words.
03:46 You can't directly alter its elements.
03:50 So I think
03:52 the music tools of the future
03:55 should let people make
03:59 pixel-level adjustments.
04:02 The most direct example is—
04:06 I don't know
04:07 when this feature came out,
04:08 but doesn't Figma allow
04:12 Claude Code, through the CLI,
04:17 to sync its coded design into Figma,
04:21 make pixel-level adjustments there,
04:24 and then send it back to Claude Code
04:26 to continue the design work?
04:28 I think this trend
04:30 supports my judgment:
04:34 the right to control it
04:36 must be returned to humans.
04:39 Otherwise,
04:40 it can only remain
04:43 something made for fun.
04:44 AI is already so advanced,
04:48 and AI coding is so powerful,
04:50 that it already enables
04:51 many people to build tools.
04:53 But is
04:54 what they build truly usable,
04:56 and can it become
04:58 a breakout phenomenon.
05:00 That still happens only rarely.
05:03 But if humans could
05:06 have a larger role,
05:09 then those—
05:12 oh dear,
05:15 how should I put this?
05:20 Well,
05:23 I think
05:24 no matter how accessible a technology is,
05:30 even if all can use it,
05:31 that doesn't mean everyone
05:34 will find it
05:36 simple
05:39 and easy to learn.
05:41 For some people,
05:45 what they may need is—
05:53 for example, they describe it in words,
05:57 and the agent presents its design.
06:01 But then,
06:04 they can't describe the next feeling.
06:10 Yet if they have design skills,
06:14 they see the interface
06:15 and know exactly what to change.
06:18 Some things simply cannot
06:20 be expressed through language.
06:22 Only through their hands,
06:25 their vision,
06:26 and their judgment.
06:29 Their eyes,
06:31 judgment, and hands
06:34 can act directly
06:37 is the interface designed by the agent.
06:41 They can manipulate it directly.
06:45 That's—
06:48 mm.
06:49 I've always hoped that
06:51 no matter how powerful AI becomes,
06:53 the human element,
06:54 or human agency, should be respected.
06:59 As I complained in an earlier video,
07:04 why did the social app Elys
07:09 take away part of people's agency?
07:13 That is something I find
07:16 very hard to understand.
07:17 I know the product team
07:19 has its own values,
07:21 but those values
07:22 conflict with mine.
07:24 That means their values
07:26 screened me out as a user.
07:29 Still, I respect that team.
07:31 After all, many teams
07:35 build products for users who fit—
07:38 or who share their values—
07:42 rather than trying to serve everyone.
07:46 Right.
07:49 Ah.
07:51 Mm, okay.
07:52 Back to Figma and Claude Code.
07:57 I haven't tried it myself,
07:59 because when I tried using Figma,
08:02 I couldn't figure it out.
08:04 With almost no design background,
08:09 even my Photoshop skills are basic.
08:16 At present, I still can't grasp
08:20 how Figma actually works.
08:22 So I've never understood
08:24 what the software could do for me.
08:28 After trying it a few times,
08:31 It felt off.
08:32 and gave up on it.
08:33 So even now that it
08:36 integrates with Claude Code,
08:38 I still haven't used it.
08:39 But I think this trend
08:41 will absolutely, inevitably
08:43 extend to Claude Code itself.
08:45 One possible future
08:47 is that Claude Code
08:49 will make a UI
08:50 we can edit ourselves.
08:52 The same goes for Antigravity.
08:55 The path it is taking
08:57 resembles Replit Agent 4.
09:00 My own guess
09:01 is that AI coding—
09:03 if it definitely—
09:06 wait, no.
09:07 When AI coding
09:09 reaches a certain stage,
09:13 natural-language description is one level.
09:16 But many things
09:17 cannot be described in natural language.
09:20 An agent cannot fully understand
09:22 what humans convey beyond words.
09:25 Some things simply
09:29 cannot be put into words.
09:31 So how can we put those things
09:34 into the work itself?
09:36 Only by having a person
09:39 shape the product directly.
09:42 Every detail they change
09:45 will affect the product in some way.
09:49 And those effects may be things
09:52 AI cannot perceive—
09:53 the parts that lie beyond language.
09:58 I don't know when this prediction
09:59 might become reality,
10:02 or whether, as AI keeps developing,
10:06 something else will emerge
10:08 that I could never have foreseen.
10:11 Mm.
10:13 We'll just have to wait and see.
10:15 Today's recording, or record,
10:18 is only a bold guess.
10:22 I simply want to see if it comes true.
10:25 And if it does,
10:27 what will it look like?
10:30 The times are developing so quickly.
10:34 I can't predict the future either.
10:41 Speaking of pattern recognition,
10:47 I used to think AI
10:50 was exceptionally good at it.
10:52 But the things AI can recognize
10:56 are those that follow rules,
10:57 have order, and are easy to compress.
11:00 The irregular parts,
11:01 the unordered, hard-to-compress parts,
11:03 are precisely the work of humans.
11:07 For example,
11:09 I don't know whether
11:11 anyone has tried this.
11:13 Suppose
11:14 someone used Replit in December—
11:18 last December,
11:18 or November.
11:22 Could an AI arrive at
11:25 the same point I raised back then
11:27 and reach the same conclusion?
11:29 I don't know.
11:31 Perhaps it could.
11:33 But at least for now,
11:37 I haven't encountered similar claims
11:41 about how Replit at that time
11:45 should have approached its next iteration,
11:51 which way to go.
11:52 I haven't seen anyone express that.
11:55 But through immersing myself
11:57 in it at the time,
11:59 I identified a future direction.
12:01 I think this is one of
12:03 the most precious human abilities:
12:07 through immersive experience,
12:11 we can predict how something
12:14 could become better—
12:16 what it could become,
12:17 or what else it could become.