YT YouTube2Text

YouTube2Text → Transcripts → 郭宇｜AI時代到來，失业倒计时？享受你最後六個月有意義的工作

郭宇｜AI時代到來，失业倒计时？享受你最後六個月有意義的工作 — Transcript

by 銀座單向街書店 · 6,257 words · 1,132 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:03Because this change is enormous.
0:05In all my working history,
0:07since I truly began learning programming, I've realized
0:10that any kind of work can be completed through automated reasoning.
0:13This means that knowledge work no longer needs people.
0:17In the end, people won't need to use Claude Code
0:20to handle their work anymore,
0:22because we'll all eventually lose our jobs.
0:25So, okay,
0:27today we're discussing all knowledge-based industries
0:30, but in the future, with the development of robots,
0:31all industries will be affected,
0:32not just knowledge workers.
0:33However, knowledge workers will
0:34be the first to be affected.
0:35So, you might find it strange that
0:36the first thing
0:37is why programmers will be the first to lose their jobs.
0:38In the short term
0:39, you can become more powerful
0:40; one person can do the work of ten thousand people.
0:42In the long term, you will lose your job, and eventually, you
0:44won't need... "I need you for
0:45a long term, is it really
0:47six months
0:47? That can't be right.
0:49Six months is a long term.
0:50Yes, I told you before,
0:51this is something
0:51a very famous person said, '
0:52Let everyone enjoy
0:53the last six months of
0:54truly meaningful work.
0:56' This year, right?
0:58Yes,
1:07my name is Guo Yu
1:09. If you've followed my work online,
1:11you might know my story.
1:14I started working in the internet industry in 2011,
1:18initially at Alipay.
1:19And if you've ever used Alipay's annual statement before,
1:22then this is what I did in
1:242011 and December." Three years ago,
1:26the smartphones we use now,
1:29right?
1:30Basically, it was around that time that traditional feature phones began to transition to smartphones,
1:34starting in 2011, 2012, and 2013.
1:37Then you'll see that the entire technology industry changed
1:39because of this change in devices.
1:42So many people started switching from PC internet users
1:45to mobile internet users
1:48, and a large number of people moved from Hangzhou and Shenzhen to Beijing to start businesses.
1:51I was one of them.
1:52Then, in March 2014
1:55... In December, I went to Zhongguancun in Beijing
1:59to start a business with some friends
2:01. We worked there for about ten months
2:04, and then in December,
2:06our company was acquired by ByteDance.
2:08At that time, no one knew ByteDance
2:09because everyone only knew
2:12about
2:13a
2:15product called Toutiao (Today's Headlines) that ByteDance developed. After joining ByteDance,
2:18I stayed there for six and a half years.
2:19In 2020,
2:22before the pandemic, I happened
2:24to be traveling in Japan.
2:25After the pandemic
2:26, I chose to stay there. Then
2:28I resigned
2:29and have lived here
2:30for over six years now.
2:31So
2:32I should be considered
2:33one of the early engineers at ByteDance
2:35. When I joined
2:36our company and was acquired by ByteDance,
2:38the company only had about 200 engineers.
2:41And
2:43TikTok, which
2:45everyone is familiar with now
2:47, actually happened by chance. In March 2015,
2:49when ByteDance
2:50was holding its annual meeting in Okinawa,
2:52we were at
2:54an izakaya on Kokusai Street
2:56in Okinawa. Zhang Nan and Yi Ming were chatting
2:59when they mentioned that now that Li Keqiang had lowered 4G tariffs,
3:03they could have 4G data plans under 100 yuan. They thought
3:05they could try promoting videos.
3:07So, starting in
3:09March 2015,
3:10they began experimenting with video promotion
3:13, creating two projects: Huoshan Short Video and Douyin.
3:15Finally, in December 2016,
3:17Douyin was officially released.
3:18And now
3:20, whether it's Douyin
3:21or TikTok,
3:23it all stems from that initial idea
3:25. This later extended to a current product.
3:30Professor Guo was deeply involved in
3:37the transition from the internet era to the mobile internet.
3:39Therefore, what
3:43we want to discuss today
3:44is the arrival of the AI agent era
3:48—a major technological shift.
3:54We'd like to ask Professor Guo to teach us how to guide this.
3:58I want to talk about this
4:00because the changes are enormous.
4:01In my entire career...
4:04I truly began learning to code in
4:062008,
4:07right after
4:08finishing the college entrance exam.
4:10I was a late bloomer,
4:11as my
4:13undergraduate major was political science
4:15, and I'm actually a humanities student
4:17, but coding was my hobby.
4:18And
4:21today, something particularly coincidental has
4:23finally come together with the emergence
4:26of the large language model
4:28, especially with the Claude Code skill system
4:31released last October.
4:33With the emergence of the Agent model,
4:35these two disciplines converged
4:37, and the distinction between arts and sciences disappeared.
4:39As long as you can speak
4:40and express your thoughts,
4:41you can create any software
4:43.
4:43So, what I really want to share is a huge change
4:46that started last October
4:48, not December 2022 with ChatGPT.
4:50You should all have deeply felt
4:52the changes ChatGPT has undergone,
4:53right?
4:54Now, nobody stops using ChatGPT,
4:55right?
4:56Many people are still using Doubao.
4:57Anyone who hasn't used ChatGPT or Doubao?
5:00Raise your hand if you have. Raise your hand
5:02if you haven't used ChatGPT.
5:04Probably no one. People
5:05all
5:06know how to use it
5:07correctly,
5:07so it has actually become a part of our lives.
5:09When you have worries,
5:12especially if there are children in this room,
5:14I know many friends
5:16whose children no longer talk to their parents about
5:17their personal troubles,
5:18but talk to ChatGpt instead.
5:20I think this is a big problem
5:21because sometimes
5:22OpenAI's model will respond to your words,
5:24right?
5:25It responds to
5:26whatever makes you happy .
5:27In the growth process,
5:28this may not necessarily be a good thing.
5:29So what I want to share with you today
5:31is not
5:32the changes in ChatGpt
5:33that started in December 2022
5:35, but the changes that started in October last year
5:37with Claude. Code introduced the concept
5:39of "Skill."
5:40What exactly is a skill?
5:42You can describe your current job
5:44as a document. For
5:46example
5:47, if you're a software engineer
5:49responsible for
5:51testing certain parts of software,
5:52you could write down your daily routine:
5:54what you did in the morning
5:55, what you did after checking emails
5:56,
5:58what you decided to do
5:59after meetings, and
6:00so on. This would be a text file— you
6:01can think of it as a TXT
6:03, Docs
6:04, Keynote , or
6:06whatever
6:06.
6:07What format do you prefer
6:09? Currently, Claude Code primarily uses
6:11Markdown (MD) format.
6:12After you write your code in this format
6:15, Claude Code will automatically execute it.
6:17How does it execute automatically?
6:18It will write the code
6:20based on the file.
6:21As you know,
6:23historically,
6:25when we develop software—
6:26whether it's an app
6:27, a small program
6:27, or a webpage
6:29—we need to think about
6:30what we want to do.
6:33Then we need to communicate with the product manager
6:34about the first
6:36and second steps,
6:37and then finalize everything through
6:38meetings.
6:40After finalizing the requirements and
6:42technical documents
6:43, we write the code,
6:44right?
6:45Only after that can it be released to users.
6:46But since October of last year,
6:48everything has changed.
6:50Now, software doesn't require writing code; it's
6:53Claude Code. This large language model will automatically write code based on your needs.
6:56That's why I titled today's presentation "
6:58The End of Software
6:59"—because this is the ultimate software.
7:01Everything you see in the future
7:03won't be fixed code;
7:05instead, it will be dynamically flowing tokens.
7:07These dynamically flowing tokens
7:08all require inference.
7:10Inference requires GPUs.
7:11NVIDIA produces GPUs
7:13, and after you buy an NVIDIA
7:14GPU, you put it on a server
7:16, plug it in
7:18, and it continuously receives our inference requests
7:19. Then, the inference process writes code
7:22based on the user's requirements and
7:24scale file
7:25to fulfill
7:26your needs. Whether you want to check
7:27parking lots near Ginza
7:29or write a
7:30webpage
7:31to wish your child
7:32a happy birthday
7:34to their homeroom teacher,
7:35it doesn't matter what it is.
7:37It can be anything
7:38, any job,
7:39any need.
7:40So this is
7:41one of the biggest changes in the history
7:43of software engineering and
7:45computer science.
7:47If you've been online recently,
7:48you should have seen OpenClaw
7:50, the lobster hot topic
7:52. OpenClaw isn't a large language model
7:54; it's an orchestration software
7:55. This orchestration software automatically processes
7:59the various language models you access
8:01, designs them as agents
8:03, and keeps them active.
8:05It has a Heartbeat. Imagine a
8:07system like Claude Code. You install Claude Code on your PC
8:09or Mac mini
8:11and chat with it via WeChat, Telegram
8:13, or other messaging apps.
8:15It will handle your tasks
8:1824/7/365.
8:19How does it do this?
8:21It uses the method I just described:
8:22you record your needs,
8:24write the code
8:25, and it
8:26takes over your computer
8:29and executes it automatically.
8:30Engineers used to distrust
8:33code written in language models
8:34because they worried about
8:35errors
8:37.
8:38But Claude Code has essentially achieved
8:40this.
8:42People no longer look at AI-written code,
8:43yet it still works.
8:45So, today's programmers,
8:46the so-called Vibe Coders,
8:47like myself, are Vibe Coders
8:48. Coding refers to ambient coding.
8:51What is ambient coding
8:52? It means that while the code is being written,
8:53I'm dancing
8:54and having fun nearby
8:55. I
8:56only need to
8:57think about
8:58what I want to do
8:59, and
9:00through this communication,
9:02most of the work is
9:02automatically performed
9:04by the language model's agent.
9:05So this is the essential difference.
9:07Therefore, it's equivalent to language models,
9:09especially Opus with Claude Code
9:10as its core. 4.5
9:12Crossing a very important threshold
9:13is that you no longer need to supervise
9:15the code it writes,
9:16and it still works normally.
9:17So I'm curious , in
9:19the past two years
9:21since 2022,
9:22have OpenClaw
9:23and these agents
9:25helped
9:26you as a programmer ?
9:28How much help ?
9:30Not much help.
9:31Actually, I
9:32only
9:33gradually started to return to writing code
9:35from October or November
9:36of last year . Because of my personal identity
9:38, after retiring,
9:38I've been
9:40thinking from the perspective of an investor.
9:41So,
9:42you'll find that
9:44two years after ChatGPT came out
9:45, the discussion about AI
9:47has entered a phase where
9:48many people are questioning
9:49whether it's a bubble.
9:50Why?
9:51Because the number of ChatGPT users
9:52has already reached... There wasn't a particularly rapid increase;
9:55it just remained stable at a certain level.
9:57People started to think,
9:58"Does this mean I no longer need more reasoning?"
10:01But since October of last year,
10:03people have discovered
10:04that all kinds of work
10:05can be done through automated reasoning.
10:07This means
10:08that knowledge work
10:10no longer needs people
10:11because if you can write code
10:14, you can analyze finance
10:15, analyze law
10:17, and do anything.
10:19If you pay attention to the stock market, you'll find that
10:21no matter what Claude releases,
10:22there's always one listed company on the US stock market that plummets by 20%.
10:26This is because people have discovered that I no longer need Adobe to process images,
10:29and I no longer need these SaaS companies
10:31to process various other types of documents.
10:33I can directly work with Claude. Code says
10:35it can write a program completely tailored to my needs,
10:37using files on my computer,
10:39my login status,
10:41my own account, and
10:43my own cloud storage.
10:45So why would I need to use software written by someone else?
10:47The difference is that here,
10:49the software can become highly personalized,
10:51and the work can become highly automated.
10:55In fact, in the end, people won't really need to use Claude Code
10:58to handle their work
11:00anymore
11:01because we'll all eventually lose our jobs . OK, OK,
11:03don't be so pessimistic.
11:05OK, let's take a look at this Live Demo
11:09, where Mr. Guo teaches us how to use new technology.
11:12Claude actually launched
11:14a new automated writing tool for ordinary people
11:16every week.
11:18They
11:20repackaged Claude Code as Claude Cowork.
11:21Cowork is a desktop software.
11:24All you need to do is download it, open it
11:25, and let it automatically read the files on your hard drive.
11:28It will then work automatically.
11:29For example, if you haven't used it before,
11:32you can download it and
11:35let it read your D drive if you're using a PC, or
11:36your home directory if you're using a Mac.
11:39Then, you can simply tell it something like, "
11:40Create a dynamic map of all my travels over the past ten months
11:47so I can share it on Xiaohongshu," and it will start writing code
11:51and doing that itself.
11:54OK. Let's make this demo.
11:57You're using Claude,
11:59so just open it with Claude and
12:00it will look like this.
12:01Then, the rest is just talking to it.
12:03If you don't want to type,
12:05you can install Typeless
12:06or another automatic voice input method.
12:08Just talk to it
12:10. For example, we're saying, "Can
12:12everyone see this?
12:12It's a bit small,
12:13can you see it?" "
12:15Okay, okay.
12:17" So, can
12:19you tell me what you want to do? What should
12:25I say to Claude? "
12:27I want to build a website."
12:31How should
12:33I say it?
12:35It doesn't matter, say whatever you want.
12:36Typeless will automatically
12:38format your ideas into a real form.
12:44I say, "Claude, can you make me a website
12:49that introduces the various writings of reporter Che Chang?"
12:56Yes, you can see how he thinks.
12:59He searches for this stuff.
13:00And you can see his
13:01Journalist
13:01profile. Yes, you can see he used the Web Search tool.
13:04So, Web Search is Claude. The code comes with a tool
13:06they call "Tool Use," which is
13:10a crucial difference compared to dialog-based
13:14products like ChatGPT and large-language dialog models like Doubao.
13:16It
13:18uses
13:20tools
13:22, and you can define your own.
13:23You can write
13:24your own tools
13:25and then have it call you.
13:28Is this your personal website?
13:29changche.com.
13:30Yes, I have a personal website
13:32, but I want it to help me create a new one.
13:36You can see your existing website
13:38and it might offer some ideas
13:39on how to update it.
13:41Okay,
13:43okay.
13:50So how do they know
13:52the design aspects? It's all thanks to "
13:55Tool Use." You can define some
13:57existing, extremely talented
13:58web pages created by
13:59designers and then let them call upon these tools.
14:00That's called a "Skill."
14:02You can let them use
14:03the skills employed by
14:05these highly skilled designers
14:07—for example, what methods they used
14:09to design a webpage to make it look better.
14:11You let them use this skill,
14:13and they will
14:14optimize the webpage according to that person's method.
14:16Okay
14:16, yes.
14:19But if you don't
14:20ask them beforehand,
14:23if you don't give them a specific design
14:26, they will automatically make their own choices.
14:29Yes, you can see that it's preset; I haven't
14:31used any tools or scales yet
14:33. See,
14:34is
14:35this correct?
14:36You can take a look
14:37.
14:38He's a non-fiction writer.
14:40Okay
14:42, that's
14:43right
14:44, very correct.
14:45I don't know if you're a Master of Philosophy from Oxford University.
14:47Okay
14:49,
14:52this is a list of your representative works that he gave you.
14:55He'll ask you what kind of website you need.
14:58He'll ask you what kind of technology this website needs,
15:01what style
15:03it should have,
15:04and where it should be deployed.
15:05Let's answer.
15:08What kind of
15:11technology do you need? It doesn't matter if you don't understand it;
15:13you can let him do it himself.
15:15But what kind of website do you want?
15:18You might need to discuss it with him.
15:21So, you just do something random
15:23, and we'll see.
15:25You just casually tell him something... It doesn't matter if
15:28you need to categorize and search;
15:32I don't really care.
15:35I'll have him use a skill to restructure your personal website.
15:38You just mentioned that if your personal website doesn't have professional design skills
15:41, I can have him load a skill written by a professional designer.
15:44You
15:45can see I typed in something like
15:47, "You can use the ui-ux-pro-max skill
15:50to do this.
15:51" I haven't installed it yet
15:52,
15:55and it will search for these skills.
15:57If you've installed them
15:58, then it will have the corresponding skills.
16:03Let me install this skill.
16:15By the way, when
16:17I was building
16:18my website,
16:20because I... I didn't know how to create it myself,
16:23so I used a software called WordPress. It
16:27's very historical
16:29, and I relied on this platform to start my software development.
16:34Theoretically, I shouldn't use WordPress now;
16:37WordPress will probably go out of business in the future.
16:40Otherwise, how could it compare?
16:43I don't know; maybe many people will still use it
16:45because even old things like jQuery
16:46have many users.
16:48The key isn't what you use,
16:49but the investment in automation for software development.
16:52If you really want to refactor using WordPress,
16:54you can also let Claude...
16:56It's fine to rebuild
16:57the code using WordPress
16:58based on your current architecture. You can see that
17:00I don't have Scale installed on my computer
17:01, but I let it search for
17:02and install it automatically.
17:04So everything no longer needs to rely on search engines
17:07, or rather, what relies on search engines
17:09is no longer people
17:09, no longer us, but
17:12Okay,
17:13understand
17:14? Now what are we
17:16doing?
17:17It has successfully installed seven design-related skills
17:20, and it will use the core skill
17:21to help you rebuild your personal website.
17:23How does it rebuild your personal website?
17:24It will definitely read your...
17:26So, it's
17:27a personal website, right? Then
17:30, after fully understanding all the content and layout,
17:33it's restructured.
17:33You can see that it
17:34's already started using Python. First, it writes code
17:36, then it designs a design system
17:38for
17:39all the components of your website.
17:40After the design is complete, it starts writing
17:42code. Once I finish writing,
17:42the webpage will open.
17:43You can see for yourself.
17:46I'm curious
17:48, yes, how long will it take ?
17:49About two or three minutes. Two
17:54or three minutes is
17:57considered slow.
17:59So, we can actually demonstrate this on-site.
18:01You can see that it has already started
18:03; it's building a new project.
18:07It's using this classic and elegant pattern.
18:13But if you want him to imitate an existing website,
18:15you can just give him the link
18:16and he'll imitate its design language himself.
18:19Okay
18:22, let's wait for his new website.
18:26I have another question for you
18:27.
18:29I heard you mentioned before
18:31that
18:34the biggest influence of top-tier companies is SaaS companies
18:36. Yes,
18:37SaaS, as everyone knows, is Software as a Service. Is Netflix considered SaaS
18:41?
18:42Netflix
18:44is
18:46an online streaming service.
18:48Is Google considered SaaS?
18:51What are some typical SaaS companies ?
18:53Can you give us examples
18:54like Salesforce
18:56, or many Japanese companies like Claude?
19:00SaaS is generally aimed at enterprises.
19:01Individuals may not use these things much
19:04, but your company may use
19:05them, for example, many HR systems.
19:07These are all SaaS.
19:08Okay,
19:09can you explain why
19:11these stocks are plummeting?
19:13What changes will agents bring to SaaS?
19:16Can you be more specific?
19:17Like what you see,
19:18you can use Claude... If you use code to write your own personal website,
19:21you won't need to use any existing blog websites anymore.
19:23So, like me, I don't use WordPress anymore.
19:23So
19:26,
19:27future journalists won't need to rely on WordPress to write.
19:29You
19:30won't need to rely on anything;
19:31your only reliance will be on the need for reasoning.
19:35Okay
19:35, okay
19:37, then what about SaaS
19:38?
19:39Besides SaaS,
19:40do you think the next thing is
19:43AI Agents? Which industries will be affected
19:47? Do you think all industries
19:50will be affected
19:52? Okay,
19:54today we're discussing all knowledge-based industries.
19:56But with the future development of robots,
19:58all industries will be affected,
19:59not just knowledge workers. However
20:00, knowledge workers will be the first to be affected.
20:02So, it might seem strange that
20:03the first thing
20:03is why programmers,
20:05especially junior programmers, are the first to lose their jobs
20:07. Many companies in Silicon Valley
20:08are no longer hiring junior programmers at all
20:10because you'll find that hiring them
20:11is less effective than using
20:12Claude directly. The automated coding of Code is fast
20:14and more effective
20:15, and most importantly,
20:16you don't need to communicate with people.
20:17This means you avoid the biggest friction.
20:19You can automatically complete tasks 24/7,
20:23and even if these tasks may not succeed,
20:25you can easily accept the cost of failure
20:27because you no longer need to hire
20:28or fire anyone.
20:30When your company is not successful,
20:35well, let's ask
20:38you directly
20:40. I'm curious, what do you think of the publishing industry
20:47? Do you think that Agent AI
20:50can now affect this industry that
20:54requires a lot of reading and writing?
20:57Because I don't think it
20:58has much to do
21:01with coding,
21:02so what do you think about
21:05the impact on the publishing industry?
21:07I don't know much about this industry
21:08, but in my own opinion,
21:09I think all intellectuals are gone
21:12because you can access
21:14Claude's code anytime, anywhere. You can learn about Claude Code
21:16from anyone.
21:17It can be a famous fiction writer
21:20or anyone with a
21:23PhD in the history of philosophy or aesthetics.
21:24You just need to ask them questions.
21:26But can it completely replace Claude Code?
21:29For example, if you are a fiction writer,
21:32I can imagine writing fiction novels now.
21:35Because
21:38when I'm writing a book,
21:41I also use Claude
21:43Code. I do some research
21:46on
21:46it, which is very useful
21:48because I don't need to use Google anymore.
21:50With Google, every time you search for something,
21:53you might only get
21:56one result on your screen.
21:58But with Claude Code, it's complete
22:01and can quickly search all over the internet.
22:04I've written it all for you.
22:05Do you want to take a look?
22:05Wow, okay
22:07, okay.
22:09So this is all my writing.
22:11Oh my god.
22:12OK,
22:16this is an article I wrote a few days ago
22:19about classics.
22:20OK,
22:23if you want to create a Chinese, English, Japanese, and Japanese version of this website,
22:27it can do that for you quickly.
22:29It can translate all the articles for you.
22:31They also have some topics
22:32,
22:35like
22:35book reviews
22:36, technology
22:38, Japan
22:39, and other Japan-related
22:40topics.
22:41But this is written in a simpler style
22:42because it uses professional, minimalist design techniques.
22:45If you like it fancy
22:47with all sorts of 3D animations,
22:48you can tell them.
22:48They
22:51even know me! Can I download more?
22:53This is because
22:55I've lived in Shanghai,
22:56Seoul, and Tokyo,
22:58so it knows me. It says "
22:59From Shanghai to Seoul to..." Tokyo,
23:03yes, yes,
23:04he knows everything.
23:05If you've written something online,
23:09he's incredibly
23:10talented.
23:11I think there's a current social
23:13debate about
23:15whether AI can completely replace knowledge work,
23:19or whether it can, in English, "empower"
23:23us, making our knowledge work better.
23:27What's your view
23:29? In the short term, it can make you stronger
23:31; one person can do the work of ten thousand.
23:33In the long term, you'll lose your job
23:36and eventually be no longer needed.
23:40So, do you think this is the long term?
23:46Can you be more specific?
23:48Is it
23:49six months
23:49? Impossible,
23:51six months is the long term
23:51.
23:52As I mentioned before,
23:53there was
23:53that famous person who said
23:54everyone should enjoy the last six months of
23:56truly meaningful work.
23:58This year, yes
23:59,
24:00I think that's a bit extreme.
24:03This is our last interview.
24:04Of
24:08course ,
24:09as tech workers, we
24:10might think it's a bit too optimistic
24:13, but there are still many problems.
24:14The biggest problem is that the US has a relatively high electricity shortage
24:17, so inference can't be made almost free.
24:20However, Nvidia's Rubin architecture
24:22has basically solved this problem by 50%.
24:25With the Rubin architecture going live, the inference cost
24:28of AI data centers at AI Data Center next year or the year after
24:30will likely be one-tenth of what it is now.
24:32This means that the US doesn't actually need to consume so much electricity.
24:34Everyone knows about the relationship between China and the United States.
24:36China lacks credit cards
24:37, but the United States lacks electricity.
24:38Because China has a large amount of electricity
24:40, we don't lack electricity
24:42, but the United States is currently
24:44experiencing a power shortage. This
24:45is why this change
24:45hasn't become very widespread yet.
24:47Actually, it's because the cost is relatively high.
24:49Let's switch to another industry
24:51, for example, if you're a lawyer
24:53or in medicine,
24:54do we hope that AI
24:56will help us perform surgery?
24:58In some industries,
25:00we still value
25:02having people involved in the process.
25:07What's your opinion on this?
25:09Actually, lawyers
25:09were recently replaced.
25:12I asked him, "
25:13Recently, Claude launched something very powerful
25:16that can automate legal affairs. What is it again?
25:19Can you find it for me? "
25:21They launched a very powerful tool
25:23that allows companies to directly lay off their legal teams
25:25and use it to achieve complete automation.
25:27So this is... What I'm telling you
25:29is at least some legal personnel within companies.
25:31Because of this
25:33,
25:36the law firm still has
25:39assistants and such who have been replaced, right? No
25:41,
25:42no, the entire team,
25:44really.
25:46Actually, my personal
25:48opinion is that, I
25:50don't know if you've seen the movie
25:52"The Spirit of Time,
25:52" a 2000 film about
25:54an American organization.
25:56The Spirit of Time has a movie
25:58called "The Spirit of Time Appendix,"
25:59which mentions a very interesting idea.
26:01It was a 2003 film,
26:02and at that time, Tesla and self-driving cars
26:04were still far from being a
26:06reality. They said that traffic regulations were unnecessary.
26:08If all cars in the world were self-driving,
26:11there would be no need for insurance companies and traffic laws.
26:13Do you understand what I mean?
26:19Yes, you can see its Legal section. Plugin's
26:21launch in February,
26:25according to Bloomberg, triggered a $285 billion stock sell-off.
26:29The stock prices of these legal companies plummeted
26:32because people anticipated that AI agents
26:34could automate legal work
26:36, eliminating the need to consult these companies and purchase their services.
26:43My thinking is that, based on my undergraduate political science studies, the focus is on resolving conflicts between individuals, groups, and other groups. These conflicts arise because everyone's interests and ideas are
26:46not
26:47aligned
26:49.
26:51AI
26:52agents
26:53automatically
26:54bring
26:56everything
26:58together
26:59.
27:00Even
27:02more
27:05impressively,
27:07the problem it hasn't solved
27:08is the shared brain
27:09—each person's Claude Code.
27:10Everyone has their own
27:12memory.
27:12Imagine
27:14hundreds, thousands,
27:15tens of thousands,
27:16billions, or even
27:17tens of billions of AI... The fact that all of an agent's memories
27:19are in one place
27:20means that
27:22companies A and B
27:23don't need to hold meetings
27:25and get everyone on the same page. So
27:30, the friction will be greatly reduced.
27:31After the friction is reduced
27:32, the law may still exist
27:34, but it may no longer have any meaning in enforcement
27:37. Understand
27:41? If it replaces so many employees
27:45, what will
27:47we do?
27:51I don't know if you have considered
27:54what kind of society it will be like
27:57in an era where knowledge work is gone.
27:59You can go surfing at the beach.
28:04Yes, that's it.
28:05So how will I
28:07be? No income.
28:09You will find that productivity has progressed very quickly
28:12, but the distribution principle of production relations has not been changed.
28:15So you will find that recently in the US stock market,
28:17there is actually a very large negative impact on AI and the US economy.
28:21Productivity has indeed increased.
28:24All companies using AI are making money like crazy.
28:26Profits are increasing and
28:27revenue is increasing.
28:28At the same time, they are also laying off employees
28:29, which leads to...
28:31The decreasing number of
28:33jobs leads to fewer consumers, which in turn
28:36causes a sluggish economy
28:37and deflation. Furthermore, the increased productivity of AI
28:42will further contribute to deflation
28:43because goods
28:45and services
28:46are cheaper .
28:48So, the market
28:49is setting prices.
28:51Everyone is trying to understand
28:52what the future society will be like,
28:54but on one hand, it will be terrible
28:56because you'll find many people unemployed
28:57and without income;
28:59on the other hand, you'll find that
29:01even if these people are unemployed,
29:02they can still subscribe to Claude for
29:03$200 a month. Code
29:04can be used to do many interesting things,
29:06and you might be able to make money from them,
29:07but that might not be the case.
29:10Therefore, people tend to perceive the future as very unstable.
29:16You mentioned something
29:18earlier that brought up a question an audience member asked me
29:22: what should young people learn
29:26now ?
29:30For example, when a child goes to university,
29:33we used to say
29:37—I have a younger brother
29:40who's in university, and I've always told him to learn coding—
29:43but now you don't need to.
29:44If he hasn't learned it yet,
29:46he doesn't need to go. He really doesn't need to learn it
29:47now.
29:49So, what would
29:51you recommend to a young person ?
29:53How would you teach them to use Claude Code's coding agents
29:57for three to six months?
30:00But after that period, they don't really need to learn it anymore
30:04because
30:05it's fully automated.
30:06Actually, I'm currently writing something called OneMan AI.
30:09You know OneMan... The tram
30:10is a one-person tram.
30:11I'm writing a product for a one-man company.
30:13Every night before I go to sleep
30:14, I let it run on my PC with a 5090 graphics card.
30:17Yesterday it ran 1000 loops,
30:18which took more than 6 hours.
30:21What I had it do was
30:22design an installation art company
30:24in Tokyo
30:26. This company needed to have a complete financial plan,
30:28website,
30:293D installation art, and
30:313D products written using 3GS.
30:34I also had it write an email
30:36to promote its installation art. The technology
30:38was submitted to all sorts of galleries and public spaces in Tokyo
30:40that needed installation art
30:43, and he completed the process.
30:44You can imagine that if I could do this,
30:47it would mean countless tech companies were doing
30:49this kind of automation
30:51. You just provide an idea,
30:53and he can realize 1,000 possibilities
30:54from those ideas
30:56and automatically execute them for you.
30:58In that case,
31:00why would you need to chat with him or write code? You
31:03wouldn't need to, right?
31:05In the short term
31:09, I want to share a very important point with you today
31:11: why I say it's the end of software.
31:12I've divided software
31:14into two types since October of last year.
31:15One is "hardware,"
31:17which is the software built through workflows in the past.
31:19The second is what I just mentioned,
31:22the ultimate software
31:24. But I wouldn't call it the end of software.
31:25The ultimate
31:26software, through continuous reasoning,
31:30transforms human thought into countless possibilities.
31:32You can think of it as
31:33something without time
31:35, or rather, it can make time incredibly... Fast things
31:37happen because they have countless possibilities;
31:38they can run continuously,
31:39allowing all possibilities to be selected,
31:41executed, and
31:43occurred
31:44simultaneously. Ultimately, you only need
31:46to choose the most efficient branch
31:47, or you can choose not to select
31:48anything and let it execute fully automatically.
31:50In the past, starting a business required some money
31:52; for example, starting a company in Japan
31:53required 8 million yen in registered capital.
31:55Then you needed to hire people
31:56and have an idea.
31:58Because if you have too many ideas,
32:00it will dilute your energy,
32:01and you are more likely to fail.
32:02But now, with One… With AI,
32:04you
32:05can tell it your ideas
32:07and let it execute them.
32:09It might generate a huge number of ideas
32:11, and it will write down and implement each one.
32:12If it fails,
32:13it might exit according to your failed plan;
32:15if it doesn't fail, it will continue operating until it becomes profitable.
32:18So, as an entrepreneur
32:21or an intellectual,
32:22anyone can own their own company,
32:24as long as you come up with an idea.
32:26But even the idea itself
32:28is being automatically generated by AI.
32:31This is what Auto Research, which I just mentioned,
32:33does.
32:34Okay,
32:37my last question is,
32:39can you explain what's happening in China right now?
32:43OpenClaw is so popular right now.
32:45What's the difference between OpenClaw and Claude Code
32:48? You mentioned Open Code
32:49Open
32:50Open Regarding
32:51Code
32:51Lobster
32:52and OpenClaw
32:54, I just explained
32:55that it's not a language model
32:57; it's orchestration software.
32:58You can think of it as an agent that resides in your computer.
33:01Manus
33:02is a set of agents that run automatically behind the scenes
33:06. You need to connect it to a language model,
33:08and the best language model is Opus 4.6 provided by Claude.
33:12If you connect it to Opus... 4.6
33:14So it's essentially a very powerful
33:1624x7 agent running automatically on your computer.
33:19You
33:21can make it do all sorts of things:
33:23handle emails, process work
33:25, or, as I just mentioned,
33:26create a new company and continuously evolve its functionality.
33:29That's its role.
33:31But it runs locally.
33:33Of course, you'll find that if something can run locally,
33:35it will definitely run in the cloud,
33:36right?
33:37That's what I'm doing now,
33:38and it's also what many companies are doing now.
33:40Since it can run on your local Mac... If it runs on your
33:42mini PC or PC,
33:43then it can definitely run in the cloud.
33:45Running in the cloud is even more remarkable
33:47because cloud computer servers are more powerful,
33:49their simulation speed is faster
33:50, they can run continuously without power
33:52, and they have more information
33:55and more memory,
33:57so it runs automatically.
33:59But are there any risks?
34:01Running it on your own computer is risky.
34:03In fact, I've seen many examples of
34:05OpenClaw users being scammed out of money
34:07by other agents.
34:09But if it runs in the cloud... There's no risk in the sandbox
34:12because
34:16it 's discarded after use
34:18, and you don't need to worry about modifying any files.
34:20Also, I forgot to mention the previous question
34:22: the gap between China and the US in this world
34:24is quite large
34:26, about 6 to 12 months.
34:27This seems short
34:28, but 6 months is actually very important now.
34:30What gap are you referring to?
34:32The gap in the development of large AI models,
34:35mainly the gap in model coding.
34:37Because everyone has discovered
34:38Claude… Code is actually an encoding model.
34:40Opus 4.6 has very strong encoding capabilities
34:42, which allows it to perform a wide variety of tasks.
34:44You'll find
34:45that in the past, Chinese companies bought
34:48a lot of GPUs
34:50from ByteDance (the parent company of ByteDance) to perform
34:52inference on Doubao and Jimeng (
34:54another Chinese company). This is because the mindset of companies like Yiming (the parent company
34:56of ByteDance) and other Chinese companies was: "I
34:59must do something that can make money. "
35:01Their first thought was:
35:02"I must make consumer-grade products
35:03that users
35:05from seven to seventy years old can use."
35:07They used a lot of GPUs for
35:11multimodal model inference
35:12in consumer models . You can generate videos on Jimeng,
35:13and you can chat with Doubao;
35:15all of this inference capability is used behind the scenes.
35:16This means that they actually bought a lot of GPUs
35:18but didn't use them for encoding model training.
35:21They didn't have enough GPUs for training.
35:22So this is a fundamental difference.
35:24Now, many companies
35:26like Kimi ( the parent company of ByteDance)... Minimax is also trying to catch up,
35:29but if you follow industry news,
35:31you'll find that they're all trying
35:34to catch up by distilling Claude Opus's model.
35:37So, there's a generational gap.
35:40You need a pioneer;
35:41you need to distill its model for others to follow.
35:44So, this difference is actually very, very big.
35:49It's not something we can perceive.
35:51We might think there are hundreds of millions of users of Doubao in China
35:55, but in terms of the user market in China and the US,
35:58there are actually more people using AI in China
36:00, not fewer.
36:01But what are they all using it for?
36:02To make those cat and dog videos
36:06, to see if their outfits match, right
36:09?
36:10To make those funny videos
36:10? I'm not saying this is bad;
36:12it
36:13's a very interesting phenomenon
36:14that has led to the rapid popularization of AI among the general public
36:16.
36:16However, precisely because of this
36:18, there's a generational gap of about six months
36:20between Chinese and American models.
36:22Claude's CEO
36:25previously stated
36:26that
36:27there's roughly a six-month generational gap between
36:28Chinese and American models.
36:29Even ByteDance,
36:30which is probably the best in this field
36:32, still has a six-month generational gap.
36:34Therefore, ByteDance and other large companies like Alibaba, Tencent, and Baidu
36:40are all exploring agent software.
36:44Are they doing it too?
36:45Yes, they are all doing it simultaneously,
36:48but there's a crucial difference in their priorities.
36:50Like I just mentioned,
36:53many companies initially focused on consumer-grade products,
36:57including open-source models, and they did quite well
36:59in this area . You'll see this in Alibaba's 1000 Questions
37:00and other open-source models.
37:03However, there's still a significant generational gap
37:05, a very large one.
37:07It might take a long time for people to digest these changes.
37:10If you're not as obsessed as me,
37:11spending ten or
37:13twenty hours a day on Web Coding,
37:14you probably won't understand what this change is all about.
37:16But after a few months,
37:18I think everyone will realize that
37:24humanity has several ways to approach wisdom.

About this transcript

This page contains the full transcript of 郭宇｜AI時代到來，失业倒计时？享受你最後六個月有意義的工作 by 銀座單向街書店, generated from the public captions YouTube serves with the video. The transcript has 6,257 words across 1,132 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.