Author: indigo
YouTube: INDIGO Digital Mirror
Guest: Guo Yu, former senior employee at ByteDance / retired programmer
This article follows the thread of the conversation, moving from ByteDance memories to the transformation of the engineer's role, from Agent OS to the layering and commodification of intelligence, and finally landing on a topic that seems distant but has already arrived: when work is no longer a mandatory option in life, how should we—especially young people—live?
In 2015, Guo Yu saw a product driven by a neural network for the first time at ByteDance; ten years later, he single-handedly used Agents in Tokyo to build more than thirty products, and then fell into burnout. In this episode of INDIGO TALK, we start with how Douyin unexpectedly won, and talk about the decision-making burden behind Vibe Coding, Agent OS, the five layers of intelligence, and why software will become as cheap as milk. When software can be generated at any time and discarded after use, what are we really delivering? Finally, when work is no longer a mandatory option, how should young people choose?
01. In 2015, seeing an "AI-driven company" for the first time
Long before the emergence of large language models, ByteDance was already a company driven by neural network products, and even the engineers themselves could not fully explain why the system operated the way it did.
Guo Yu's tech career began at Alipay in 2011. In April 2013, he first encountered Bitcoin at an internal sharing session at Alipay, when the price happened to be around 100 USD. That year was also the starting point of the real explosion of China's mobile internet: feature phones migrated to smartphones, a large number of practitioners from southern tech cities poured into Beijing, and many of his colleagues at Alipay also jumped ship to do mobile application development, or joined smartphone brands such as OPPO and vivo.
In 2014, because of some turning points in his personal life, Guo Yu made up his mind to go north and join a friend's startup in Suzhou Street in Zhongguancun. The team was working on Duoshuo and Tuchong, and at night they slept directly on the office floor. Interestingly, part of this company's investment came from the Weibo fund, and at that time Indigo was responsible for platform business and investment at Weibo. The two did not know each other then, and only years later in this episode did they connect this intersection. It was also precisely because Weibo had also invested in ByteDance early on that this relationship became the opportunity for the later acquisition.
In December 2014, Guo Yu received a call from Shen Zhenyu, saying that Toutiao wanted to acquire them. His first reaction was: "What is Toutiao?" At that time no one knew the name ByteDance. Back then Beijing was full of interesting startups, and 36Kr and the then just-starting Keep were both fancier than Toutiao; although Toutiao had just completed a 100 million USD Series C financing (Indigo added that Weibo had also participated in this round at the time), it was not used much among young people. After the merger, some employees in the team even left after a few months, feeling that this company was "not very reliable."
What truly made Guo Yu realize that ByteDance was different was a small story. Zhang Yiming asked employees to recommend their own products to relatives and friends. Guo Yu had a relative who installed Toutiao, particularly disliked a certain type of news, and clicked "not interested." But a few days later, similar news still appeared on the homepage. Guo Yu went to the recommendation algorithm engineer to report this bad case, and the other party's explanation greatly surprised him: they were actually also parameter-tuning engineers, and the recall strategy set for each experiment could not block this type of content 100 percent. They themselves could not clearly explain how the system actually operated internally.
"Some code is already running inside a black box, just like the large language models commonly driven by neural networks now. You do not know exactly how it works. You can only improve it continuously through experiments, to approximate the effect you want to achieve." [05:45]
This is exactly the situation everyone faces today with large models, except that ByteDance had already become accustomed to it ten years earlier. Among Alipay, where Guo Yu had worked, and the startups he knew about, not one had a product driven by a neural network; at ByteDance, however, products and business were split into a powerful business support department (called the middle platform back then) and a powerful commercialization department, with the recommendation algorithm as the engine running through everything.
Indigo inserted a Weibo anecdote here, which formed a sharp contrast. During the same period, Weibo was also improving its feed, but it still used a relatively manual rule-based algorithm. Weibo's fundamental difficulty was that it prioritized the relationship network, and like today's X, follow relationships were very important. Once recommendation algorithms were used to filter content, users would complain: "Why can't I see what the people I follow posted?" Weibo struggled back and forth between the recommendation feed and the following feed for a long time, and ultimately failed to do it well. To some extent, the divergence between the two companies was already destined at that time: one started from social relationships, the other from algorithms.
However, Guo Yu also reminded that even if text- and image-based recommendation products rely on neural networks, the audience they can cover and the effects they can achieve are far inferior to later video recommendations. 2015 was actually a very difficult year for ByteDance: there was no Douyin yet, Toutiao's DAU, MAU, and retention had all entered a bottleneck, and a large number of employees left.
02. Douyin was an accident
Douyin's success was not the result of a bet made in advance, but rather, in long-term internal competition with Huoshan Xiaoshipin, it won through excellent "user generalization."
Many people think ByteDance bet on Douyin from the very beginning, but Guo Yu says this was not the case. What Zhang Yiming initially wanted to build was a more mass-market video recommendation product, so a large amount of resources was invested in Huoshan Xiaoshipin, which was positioned for the general public. When Douyin started in 2016, it was only a very small team, and its name was not Douyin either, but Z, with a product form very similar to musical.ly.
The turning point came after the acquisition of musical.ly. Alex (Zhu Jun) joined, and both the product form and the recommendation engine behind it were significantly strengthened. The competition between Douyin and Huoshan continued repeatedly for four or five years, and in the end Douyin accomplished one key thing: user generalization.
Generalization, in Guo Yu's words, means slowly turning a product loved by young people from 18 to their twenties into a product usable by everyone from 7 to 70. When Douyin accomplished this, the historical task that Huoshan had originally been meant to undertake was also completed by Douyin, and the two eventually merged. In December 2016, Douyin and Huoshan, which the team had worked on for more than ten months, were launched one after another, and only by 2017 did ByteDance's overall situation truly improve substantially.
During this period, Guo Yu was always in the middle platform, and the businesses he supported changed very quickly. He worked on live video and participated in the once-popular quiz live-streaming product "Million Heroes." ByteDance invested a large amount of resources in programs related to CCTV's Spring Festival Gala, where users could answer questions in real time on their phones to win a million-yuan prize. Later, he was long responsible for the so-called Micro Runtime, that is, the mini-program engine running inside Douyin and TikTok. Games and various JS Bundles could be dynamically published there. You scroll through videos, scroll to an ad, and then scroll further and there might be a mini-game you can play directly. This system was first built by Guo Yu.
He positioned himself as a technical executor: during his many years at Toutiao, he never led a team, and did not want to.
03. A 50-yuan customer acquisition cost and an accidental settling in Tokyo
When even an "app factory" like ByteDance found it difficult to acquire new users, Guo Yu judged that the mobile internet cycle had ended; a pandemic unexpectedly led him to begin a new life in Japan.
In 2019, one figure left a deep impression on Guo Yu: the cost of bringing in a new user from outside had already reached more than 50 RMB. Even for an "App factory" like ByteDance with excellent retention, it was hard to acquire new users anymore. He realized that he had basically done all the things he could do in this era.
In mid-2019 he decided to leave, and the handover dragged on for about half a year. During the 2020 Spring Festival, he went home and then traveled to Japan, but the pandemic broke out, international travel was interrupted, and he was never able to return to Beijing. Even the computer handover was completed only when his colleagues returned to Beijing half a year later. Just like that, he obtained a long-term residence visa in Japan, later got permanent residency, and has lived there ever since.
Actually, as early as 2019 when he wanted to leave, Guo Yu had already made a table comparing various destinations. Singapore was the choice of many former ByteDance employees, mainly to change tax residency, but he grew up in Shenzhen and did not like tropical climates. Canada was also an option he seriously considered early on. If he had chosen Canada at the time, he would most likely have taken the rapid immigration pathway during the pandemic (Indigo added that at that time almost everyone passed). In the end, he chose Tokyo because its overall score was higher, and because by then he had already decided to leave the tech industry and live the life of an ordinary person.In his view, the opportunities in the tech industry were mainly Beijing's opportunities, and that mobile internet cycle had basically ended. It was not until the AI wave erupted again at the end of 2022 that he actually had only two or three years of real rest, and those two or three years happened to be covered by the pandemic.
"Leaving ByteDance was quite a coincidence. It wasn't such a complete experience, because the pandemic interrupted it." [11:13]
Indigo met Guo Yu during exactly that period, on Clubhouse in 2021 to 2022. Indigo recalls that during the pandemic he also began writing on X, and the strongest feeling at the time was that the whole era suddenly accelerated: remote work, new treatments, and especially SpaceX's rocket recovery. When his partners invested in SpaceX in 2019, it was purely for the stars and the sea, and they even felt that "this thing is definitely not very reliable." Later, on November 30, 2022, ChatGPT was born, the U.S. stock market was suddenly lifted by AI, and the two also began interacting more on X over U.S. stocks and Crypto.
04. From executor to "everyone," the real cost of Vibe Coding
After AI took over coding, engineers did not become more relaxed. One person has to shoulder the decision-making burden that used to belong to an entire small team, and that is the root of burnout.
Indigo sorted out the trajectory of these years: 2022 to 2024 was the Chatbot era; in mid-2025, Vibe Coding, proposed by Andrej Karpathy, began to become popular, with people writing code in Cursor, and then moving toward fully automated programming. He noticed that Guo Yu quickly entered a high-intensity Coding state, so he asked: what is the essential difference between this round of "implementation" and when he was an engineer at ByteDance?
Guo Yu's answer went straight to the core: the most important change is identity. After coding ability was taken over by AI, engineers stepped down and became observers.
In traditional tech companies, a small team decides what to do, with one to two weeks as an iteration cycle, setting goals, developing, designing, testing, releasing, and observing experimental results. With Agentic Coding, such a cycle can be compressed into one or two people and half a day.
It sounds like liberation, but in reality it is added pressure. You now have to be an engineer, an architect, a designer, and most importantly a product manager at the same time. Design, architecture, and technical execution can mostly be handled very well by AI, so you must become someone good at asking questions.
"Essentially, you have to constrain the language space of the language model. You need to know what kind of language to use to communicate with it so that you can get the result you want; you need to know what kind of language structure can keep it pushing this matter forward." [16:30]
There is a very good analogy here: in the past, writing code was like laying bricks with your own hands; now it is like standing on a construction site directing a construction crew that never gets tired, but the blueprints, priorities, and which wall to tear down and rebuild all have to be decided by you. And because the Agent keeps running in the Loop, you also need to know how it adjusts its direction on its own based on experimental results. Writing code has thus changed from a relatively simple mental activity into a much more complex mental activity.
Why can't AI make these decisions for you? Guo Yu believes it is not only because it lacks context, but more fundamentally because it has no active intent. Without intent, it cannot make all decisions for you. Whether the currently hot Personal Agent, after having sufficiently rich context, can move from "giving suggestions" to "making decisions proactively," he is not sure, and at least right now it still cannot.
The result is a surge in psychological burden. Guo Yu observed that from October and November 2025 to February and March 2026, many engineer friends around him experienced Burnout: they worked day and night for two or three months and completed the workload of the past one or two years.
"The human brain consumes very little energy, and it also cannot keep up with AI's speed. We face too many decisions to make, which affects our autonomic nervous system, making the whole person very anxious and unable to calm down to truly think about problems." [18:30]
After burnout, many engineers, including Guo Yu, are rethinking two questions: what is truly worth doing? At what rhythm should we collaborate with AI?
Clash of views: should you make a plan?
Here the two had an interesting disagreement. Indigo shared his experience from the past month or so: he wanted to turn his knowledge management, trading, and information management tools into products and release them, and found that the method of writing Specs and PRDs from when he led a product team at Weibo was still useful. Documents do not need to be as refined as before, but after clarifying your thinking and then handing it to AI, the usability of the delivered result will be much better than just letting it run directly. So he is used to first discussing an execution plan with AI, then letting it execute, with Planning first.
Guo Yu went further. He mentioned that Claude Code recently canceled plan mode, because Agentic Coding itself is already strong enough. In the past, you might first spend half an hour planning all the Specs in great detail, then run it all at once; but now people hope more that it can run 24×7 without interruption, and this continuous operation itself can cover everything you used to want to plan, "because what we can think of is actually not as much as AI." In the past six months, he has rarely used plan mode. As SOTA models and Harness continue to evolve, human engineers also have to keep adapting, and this itself is another source of burnout.
Indigo admitted that if he does not discuss the implementation plan with AI, he will lack a sense of control. Some things can be abstracted up one level, and he does not care at all how the code is written, but he must be clear about the logic at the Business level. So he often lets AI finish, then asks it to explain its thinking to him, "let it teach me." Guo Yu said he has recently let go even more. These two postures actually represent two rhythms of current human-machine collaboration: one is to stay in the Loop and understand the system, and the other is to retreat to the goal level and trust the process.
05. Software is no longer hard-coded
Recommendation engines made content personalized for everyone, while Agents make software itself personalized for everyone. When software can be generated instantly, the "distribution" of software loses its meaning.
Guo Yu believes that an even more important change than the shift in roles is that the nature of software itself has changed.
He told a story about an all-night chat. In January or February of this year, his good friend Dongxu could already control hundreds of virtual machines at the same time, running all night continuously to accomplish one big thing, consuming billions of Tokens a day. Dongxu said at the time that he felt that with this kind of matrix-style Agent network, any software could be made.
This made Guo Yu realize that not only have the form and structure of software changed, but coding itself has also become an instant behavior: writing for the sake of completing a certain task or achieving a certain goal, and it can be discarded after use, or used only by yourself.
"Previously, with AI-driven recommendation engines, we personalized the content of software we had already made for everyone. But now even the software itself can be completely Customized. Your software is your software, my software is my software, and our software may not be compatible with each other, and there is no need to make it compatible." [22:50]
This contrast is very ingenious. Ten years ago, everyone used the same Toutiao, only seeing different content; today, people may not even be using the same App.
Guo Yu himself spent three or four months Vibe Coding more than thirty products, many of which were not released. At first he felt there was no difference from being at ByteDance. Back then, as an app factory, many products were also stillborn, or were cut after launch for failing to meet ROI. But he later discovered that he was still following the thinking of the previous era: making software for people who do not know how to write code.
This is being dismantled by what is called First Class Harness. Guo Yu's definition is: the first layer of Harness built personally by SOTA model labs, such as Claude Code, Codex (Guo Yu mentions it has now been merged into ChatGPT), and DeepSeek Harness (DSH). Their progress has already internalized the functions that a large amount of software needs to complete. If you want to use it to build a CRM or various management systems, it is a matter of two minutes, five minutes, or ten minutes. For this kind of need, making the software well and putting it into the app market for everyone to use has already become somewhat meaningless.
06. Agent OS, when Agents become first-class citizens of the operating systemThe next step for operating systems is to shift from being "application-centric" to "Agent-centric"; the computing resources we rent in the future will likely all be Virtual Agent Servers operated by Agents.
If the meaning of software is changing, what is the next step for the internet and operating systems? Guo Yu noticed that some people are already building entirely new distributions based on Linux, treating Agents rather than Applications as first-class citizens and letting Agents operate the entire system. However, for a new operating system to fill in the basic OS APIs and the functionality of complex applications, it takes a long time to accumulate.
So OpenAI took another path: it bought a large number of Mac minis to train the next generation of models to operate these Macs. He believes this path is quite reasonable, because the Mac itself is a very good Unix distribution, with extremely mature tools and software ecosystems in fields such as creativity and film and television production.
From this, he painted a picture of the near future: what you use on your phone is actually connected to a remote Mac, perhaps the one at home, connected back home through OpenAI's Tunnel or your own Cloudflare Tunnel, or connected to your server in Tokyo. On a longer scale, say five years, this remotely operated computer will definitely become specialized, turning into some kind of dedicated Linux distribution. Today, when you rent a server in a data center, you have to install Ubuntu yourself and install a bunch of software; in the future, all computing power rental may become Virtual Agent Server, with one Agent operating the entire system and completing what you need to do for you.
Indigo added a sense of realism with the Grok Bot he had recently been playing with. The Linux environment behind it has a strong flavor of an Agent OS: there is a small browser, a file manager, and a terminal. If you want to see what it is doing, you can go in and look, but most of the time you do not need to open it at all, just tell the Agent. All Agents share this one Computer. He also did an experiment: he installed clients on both his home MacBook and Mac mini, and the computer in the cloud could know that these two devices existed. By issuing a command on one, it could access it through the client on the other. Thus, the cloud virtual machine acted as the overall coordinator, the two local machines participated in execution, all data was connected, and he did not have to type a single command himself.
Guo Yu shared two remote capabilities that Codex has supported since February: one is traditional SSH port control, and the other is penetrating the intranet through Codex's own WebSocket connection, allowing you to control all computing resources from any machine or even a phone. OpenAI did not promote it as a selling point, but the capability has always been there. Guo Yu's own data center has several bare-metal servers, and now he can even delegate buying machines to an Agent: after purchase, the service provider sends the initial root password to Gmail, the Agent reads the password, logs in, installs its own key pair, deletes the password, and after that everything is operated and maintained by itself.
But the first time he did this, what was on his mind was something else:
"If OpenAI is attacked, or its connection is hijacked, it actually becomes the world's largest botnet. Because you can operate hundreds of thousands or even millions of different computing resources through one Agent." [29:55]
When everyone manages computers through Agents, the centralized Agent channel itself becomes a huge security single point of failure. This is the dark side of the Agent OS picture that cannot be ignored.
07. The Five Layers of Intelligence
Not everyone needs the strongest model, but the most suitable model. Intelligence is being layered by scenario, from personal convenience, personal productivity, enterprise collaboration, entertainment generation, to scientific research exploration.
Indigo brought the topic back to the user interface: the interaction layer is becoming LUI (Language User Interface), where everyone asks the model to generate the interface they need for themselves. But there is a Gap here: people who have done product and technology know how to describe requirements, but ordinary people may not.
Guo Yu believes that needs are layered, and gave the example of Meta's recently released Muse (currently only available to U.S. users). What it needs to solve is ordinary people's simple needs: automatically calling to book a restaurant, booking things while traveling in an unfamiliar country, and managing various kinds of information. It does not need a separate App, but is integrated into Meta's existing communication channels such as Instagram, WhatsApp, and Messenger, stringing all communication together around the Agent. In the past, you had to piece together a workflow using a bunch of software; now you do not need a workflow. State your temporary need, and it keeps processing until the goal is achieved.
Guo Yu has deep personal experience with this. In March and April of this year, he made a product that made calls on people's behalf, and after release quite a few Portuguese- and Spanish-speaking users used it (Indigo said he had also used it and recommended it to others); later he made another product intended to integrate phone calls, email, and all kinds of communication and hand them over to an Agent for automatic management. But by nearly October, super apps had already built these basic functions in. This once again confirms the judgment from the previous section: distribution for specialized software is becoming unnecessary. Of course, more complex creation still needs to be done around SOTA models such as Claude Code and Codex.
Following this line of thought, Indigo gave his five-layer framework, starting from a judgment: people do not need the strongest model, perhaps only the most suitable model and the most suitable scenario.
The first layer is personal convenience, like Muse, integrating daily connections and trivial matters.
The second layer is personal productivity, writing code and creating, using models such as Claude and Codex that have stronger interfaces.
The third layer is enterprise productivity collaboration, which is more professional than personal use and requires Agents to integrate into enterprise workflows, such as letting models directly learn to use Mac minis and Linux, connecting systems within the enterprise, with higher requirements for models.
The fourth layer is entertainment generation. Indigo believes this is ByteDance's strength: Seedance paired with the Hongguo short drama platform, AI-generated and human-consumed, has already formed an internal loop; BytePlus has also made tools for overseas creators, packaging the sale of video generation models together with a full set of creation services.
Small games and interactive games may no longer need people to make them. Brands and professional creators generate content and distribute it through social platform feeds; as head-mounted display devices become widespread, ByteDance and Meta, both companies betting on the metaverse, will occupy an entry-point advantage. "The previous metaverse was just born too early."
The fifth layer is scientific research exploration, which requires the most cutting-edge models, and such models may not be made public by model companies at all. Instead, they may sign annual framework agreements with pharmaceutical companies and research institutions and share profits, and defense orders are probably similar in form.
Guo Yu added two more blocks on this basis.
First, enterprise adoption is still awkward. Microsoft has just released a new form for enterprise work, lowering the hierarchy of Office below the Agent and letting the Agent operate Office. Copilot has actually become an operating system (the brand name is still retained). Microsoft wants to accelerate the implementation of Agent-centric automated workflows in enterprises, but enterprises are still hesitant: they know they can lay off staff to save money, but they do not know how to rebuild the entire organization's information flow and capital flow around Agents. One Person Company and small teams will naturally embrace it, while large enterprises (including Japanese enterprises) have many concerns, because no one can take responsibility for this matter. Only the CEO and the board can make the decision, and middle managers who see the opportunity do not dare to move. They even do not dare to use Chinese models such as DeepSeek, fearing that internal information will be distilled and leaked to other enterprises.
Second is what he calls the on-site AI layer: the cheapest and fastest models, used in highly automated scenarios, such as managing farms and factories. Such scenarios must rely on edge computing, because the amount of information a camera has to process in a day is too large. For example, if a ranch needs to identify which cow has run out of the pen, you cannot send all the images to OpenAI; you can only use extremely cheap models running locally 24 hours a day.
Putting these together, Guo Yu gave an investment-perspective positioning:"At this moment in October 2026, AI's impact on human society, organizational structures, and business is still very, very early. Although we are accelerating, what we can accomplish may be only 2%, 3% done." [39:30]
08. The 2% Pro User and the Commoditization of Intelligence
AI is a blank sheet of paper; it only amplifies people who have ideas. For the mass market, AI-generated content will move toward perfect competition, and intelligence itself will eventually become as cheap as milk.
Indigo observed that AI's user distribution also remains at an early stage: the first to adopt it are Pro Users, those active, learning-oriented, research-loving people. The reason is that AI does not place buttons in front of you like games or traditional software. It is a blank sheet of paper, and you must first have an idea to start it. Most people go to work, get assigned tasks by their boss, finish and leave, go home and lie down to watch TV. They have no ideas, and they have never thought about how to make requests.
As a result, AI shows a clear polarization: for Pro Users, it is a tenfold amplifier. One person can build more than thirty applications and manage hundreds of Agent Servers, equivalent to a company of fifty or sixty people in the past. But from various surveys by Gemini and ChatGPT, most users still treat them as more convenient search engines. By population distribution, perhaps only 2% to 3% of people can become Pro Users. The rest need another kind of AI to fill their time, and this is exactly what ByteDance is doing.
Guo Yu holds reservations about the commercial prospects of this path. Consumer entertainment has always been the core of ByteDance's core, relying on capturing attention and monetizing through advertising. But for products like AI short dramas, the profit model remains uncertain. Even if a closed loop is formed, you pay Seedance to generate videos, then upload them to Hongguo and let others pay to watch. In economic terms, it will also become a perfectly competitive market: there will always be someone using cheaper Tokens, spending more time doing more experiments, driving prices down until the selling price is only slightly higher than the generation cost. After deducting computing power and electricity costs, the profit is so thin that it becomes meaningless.
"If a market is a perfectly competitive market, its final pricing will be almost no different from cost." [42:30]
Indigo picked up the thread: this is what many people in Silicon Valley now believe, that intelligence will become a Public Good and be thoroughly commoditized. Guo Yu used a very apt metaphor:
"We buy milk cheaply at the supermarket because there are large centralized factories producing milk. If we view software as milk, in the future the cost of obtaining any software, anything that helps us achieve our goals, will become as cheap as milk." [42:50]
So Guo Yu's focus is shifting: from creating software to thinking about the purpose software was originally meant to achieve.
09. Confusion, and Waiting for the Environment to Mature
Under the new paradigm, rather than rushing to follow old thinking to make software, find users, and handle distribution, it is better to do what early ByteDance did: spend time observing and wait for the environment to truly mature.
Guo Yu admitted that he is somewhat confused now. A while ago in Singapore, he chatted with Evan You for half an hour. After selling his startup to Cloudflare, Evan also felt somewhat weary and was thinking about what to do in the next stage. Apart from returning to personal life and family, Guo Yu feels he needs to think more deeply, for example by observing for three to six months before deciding where to put his most important energy.
This reminded him of ByteDance's own history. ByteDance was founded in 2012 and spent two years validating its first idea: whether a recommendation system could work on text and images. In March 2015, at the annual meeting in Okinawa, the most important topic everyone discussed was that the overall environment had finally made people willing to use mobile data to watch videos, so video recommendation thus entered a mature period and was worth trying.
"If we have not entered a mature environment, users do not know how to use our products, and we still follow the entrepreneurial thinking of more than a decade ago to make software, reach users, and distribute software to them, then does this itself not have much meaning?" [44:45]
The topic then turned to the next generation. Guo Yu said he envies today's children. They were born in this era, have no baggage, and would never think, "I want to make something and sell it to whom." Indigo also talked about his child who just started college: not much thought of making money, more enjoyment of personal hobbies. Previously, when interested in geography, the child studied it every day and made things every day, and had a great sense of community on Reddit. After life's pressures disappear, the next generation may follow interest communities more. Guo Yu added that after the bubble economy, Japan entered such a social state, and AI may bring the next generation of the whole world here. If it really is as Elon Musk says, that in ten to fifteen years people will no longer need money and robots will do all the work, then this state is very likely to come true.
10. Everything is Media
In an era flooded with AI content and with an extremely low signal-to-noise ratio, real people themselves become the scarcest differentiation; what we deliver is no longer software, but media.
Indigo shared his own thinking. This year in Silicon Valley, he met too many unreliable small companies, all doing the leftover edges of models. As Guo Yu said, if software no longer needs to exist, then these functions have no value. But he believes one thing will become increasingly important: the connection between people.
"The emotional connection between people will become stronger and stronger as AI intelligence becomes more and more advanced. I really want to see a real person here talking about their own things and sharing, and that will be very interesting." [46:50]
He set a small goal for himself: to productize himself, or rather to productize his Attention. A large amount of his thinking will be automatically organized by AI Agents into content that can be read and shared. Community members even hope he will directly open an API so that their Agents can read what his Agent writes, completing another transmission of information. What he wants to deliver is to deliver his attention to others and form connections through collisions of thought, just like this episode: two people concentrate their attention at high frequency and then release it to the audience.
"This is not software, this is media. Software is media, everything is media. Every time you produce a Vibe Coding product, that action and that result are part of Media. Users take it, digest it, feel satisfied, just like watching a short video, and tomorrow they still want to see you post another one." [48:10]
Such content will form a Community around a certain theme, such as US stocks and the primary market. People will still be willing to communicate with others in it, and KOLs will still emerge.
Guo Yu echoed this judgment with the example of Zara Zhang. She did not come from a programming background, yet she made many Keynote-related products and became very popular online. She posted a tweet: if you want a startup's product Demo to get more traffic, the only thing you need to do is have a real person appear on camera to introduce the product. Guo Yu's explanation is that there is more and more AI Slop, anyone can use AI to create whatever they want, and the signal-to-noise ratio has become extremely low. In an era of extremely abundant content, appearing on camera as a real person is itself the greatest differentiation, provided that you appear well enough and do not speak nonsense.
Guo Yu then told a classic internal ByteDance case that reveals the mechanism behind this differentiation. Every so often, Douyin would surface a completely unknown ordinary person and push them to the center of traffic across the entire network, often because they had unknowingly Hacked the part of the recommendation algorithm most aligned with human nature. ByteDance has a dedicated team to counter this behavior. The most classic technique is to deliberately write typos in video subtitles. Guo Yu initially thought it was to evade review, but later discovered that after adding typos, the completion rate would increase significantly: people would go back to check whether that character was actually written incorrectly. This technique was quickly identified and blocked, but engineers can never know how creators will Hack it next time, because although the algorithm is made by them, they are only parameter-tuning engineers and cannot intervene in its internal logic.
"Whether we distribute AI content or real-person content, in essence we are also Hacking the traffic of this mental model. Now there is more AI content and less real-person content. If you do something outstanding, the experimental effect in the early traffic pool will be better, and you will be placed into a larger traffic pool." [51:40]
From this, Guo Yu returned to his own core question: what exactly are we distributing? Is it software, or the story told through software, the personal brand established, or the business goal achieved through personal brand? The past order from A to B to C to D may now be completely reversed.
"If we can bypass software and achieve this goal, then perhaps we do not need software itself. If all our goals are achieved through intelligence, then having intelligence itself is enough." [52:35]He used the entertainment industry to illustrate: for most people, the goal is simply to pass the time. Whether watching short videos, watching Hongguo's AI short dramas, or scrolling TikTok, in the end it is all about consuming time; the product form in between does not matter. The reason Douyin grabbed a huge share from Tencent is that if you are playing games, you cannot watch short videos, and if you are watching short videos, you cannot play games. The same goes for long videos. All entertainment products look different in form, but their competition for users' time is direct. By extension, if work and life goals are all regarded as ends in themselves, then for To B software, factory software, and personal apps, what the intermediate software form is may not matter at all.
11. Industrial Revolution, or Renaissance?
This change altered the means of achieving goals like the spinning jenny; and when basic needs are met and time is released, the lifestyle of Japan's Heisei generation may foreshadow the next stop for all humanity.
Guo Yu compared this change to the Industrial Revolution, such as the 18th-century spinning jenny: the ultimate goal is to make clothes, and it does not matter at all whether human hands or machines are used, as long as the goal is achieved. Indigo then pointed out that people's hierarchy of goals is shifting upward: in the past the goal was to write software, but now you only need to make a request and let AI build the interface, or simply have no interface at all, and moving something from A to B is enough. So in an era when Agents are extremely advanced, what will people's sense of purpose and meaning become?
Guo Yu told a story from when he had just "retired" to Japan. He often drove out to have fun. One day he drove to Kujukuri Beach in Chiba and saw many people surfing, and that day was a weekday. Only later did he learn that quite a few young people in Japan no longer take formal jobs, but work as アルバイト (temporary workers). In their view, temp work is free, and the income is not much different from that of regular employees. This is mainly thanks to Japan's very small wealth gap, where income is similar no matter what job you do. So they work three days and rest four: work three days a week, and spend the remaining four days surfing or doing whatever they want to do.
"When your material desires are extremely low, or when all your needs are satisfied by this society, you can freely pursue anything you truly want to do. These things may be your interests, or they may, because of long-term interest, become your career." [55:50]
He gave the example of the Olympics: some Japanese athletes in events such as skateboarding are not professional athletes in the traditional sense. They simply accumulated professional experience and skills through years of interest since childhood, and then chose to compete in the Olympics. This is not like F1 drivers, where huge investment is required before there is any return; rather, it is the natural result of interest-driven development, and also a very important manifestation of the so-called "Heisei shut-in" society after many years of development.
Of course, Guo Yu also admitted that for the general public, most people will still think about how to pass the time. But within the public, there will always be a sufficiently large minority who use AI to greatly increase productivity. In absolute numbers, they will be far more numerous than in the era of traditional education, and more equal as well. In the past, a child born in a village could not receive a standard university education, and could not do software even if they wanted to; now, as long as they are willing to change themselves, they can get what they want.
He borrowed a philosophical concept that had always been discussed inside ByteDance: if you see Douyin as an information cocoon and immerse yourself every day in what you want to watch, then of course you will become useless; but if you see it as a window to a larger world and actively train your recommendation algorithm, then you can learn anything you want to learn. He remembered that on Douyin there was a child from a remote mountain village in the west who especially loved runway shows, made clothes out of plastic and rags, and walked the runway in the fields. By continuously posting videos, they eventually really went to walk in Milan.
"In traditional times, no matter how much you spread your ideas, you would only be considered crazy by your compatriots and fellow villagers. But with tools like this, you can use them to reach a bigger stage. AI is the same." [58:00]
Guo Yu believes that the childhood of our generation belonged to the last century, so it is inevitable that we have fixed ways of thinking; the next generation does not have these burdens. Proportionally, the public will still consume entertainment, but the changes brought by AI are too enormous. Even if only a small minority, they will create explosive productivity and creativity, and "we will see many things we never even imagined before."
Indigo, by contrast, prefers to call it a Renaissance. Just as in the 14th and 15th centuries, when productivity suddenly increased, nobles spent large sums to support retainers and artists, and science and art flourished together. Japan's Heisei generation has already experienced a similar state, except that the cause was the economic model rather than AI; and now, the next generation around the world will have more time because of the productivity gains brought by AI.
This is a double-edged sword: having more time may mean not being able to find a job. But the poverty floor is also constantly rising. "No matter how badly off you are now, you live better than a king 300 years ago. You have all kinds of electronic devices." Indigo believes that the matter of going to work has existed for more than a hundred years since the Industrial Revolution, and should also disappear after the intelligence revolution. People will diverge: some will have no desires and no demands, and can survive by scrolling short dramas and playing games, as long as they do not harm society; others, the Think Different people, and there are always such people in the human species, otherwise we would have gone extinct long ago, will seize this AI Renaissance.
"We now have tools of infinite possibility. As long as you have ideas of infinite possibility, you can do things of infinite possibility." [61:55]
He imagined that in another year or two, as long as you have enough agency in your mind and unique human creativity, mixed with AI, you will be able to produce a large number of works, that is, Artifacts, or entirely new media forms; with these works, people who identify with you will quickly gather into a Community and participate in creation together. The social structure may shift from "enter a company to work" to "Pro Users create media, many people build communities around the media, forming a circulating small world." Therefore, social media remains important.
12. Advice for young people: do not cling to a falling stock
Stop treating work as an iron rice bowl; short-term anxiety is an escape from a long-term decline.
Finally, Indigo asked Guo Yu, this person who had "been acquired as a startup founder, retired early, and then ran into the AI revolution," to give young people some advice.
Guo Yu first described two types of young people. One type is very anxious, feeling that they will soon lose their jobs, or simply cannot find work at all, because the cost reduction and efficiency gains brought by AI have reduced the number of positions in companies, and they feel the future is bleak. The other type is children who have not yet encountered this problem. They are full of interest in AI, and toys and games that they could not make before, they can now make themselves. These two generations happen to lie on the boundary, and are the two most symbolic generations in terms of AI's impact on real life. Further on, for example, when children who are now two or three years old grow up, AI will already be something utterly familiar, as commonplace as old people using WeChat and they using QQ.
For those who are being forced to look for work, his most important advice is: stop treating work as the treasure book of life and an iron rice bowl, and stop viewing work as a mandatory option. If you do not work, the choices you face are infinite. In the short term, this magnifies anxiety, because you do not know what you want to do.
"But the reason you chose not to work is actually that you are avoiding a long-term decline." [65:00]
He used investing as an analogy: when a stock is trapped, what you need to decide is not the specific selling price or how much you have lost, but two things. First, is there a better opportunity? Second, is it on a very, very long downward snow slope? If the job itself is on a long downward slope, then clinging to it is the greatest risk.
He told the story of his father. After reform and opening up, many people working in state-owned enterprises or work units wanted to go south to Shenzhen and Guangzhou to "jump into the sea" of business, but they feared a risk that people today find hard to understand: their years of service would disappear. So when his father went south, he spent several hundred yuan every month specifically to arrange a leave without pay to retain his position, worried that if his venture failed he would still have to come back. As it turned out, a few years later, that mine closed down. Friends in the Northeast feel this even more deeply; there were simply too many laid-off workers.
So the first step is to change your way of thinking. Do not treat work as the core goal. In the short term you will be anxious, but in the long term this choice is undoubtedly correct. Of course, this also depends on the content of the work: if it is going to a dull work unit to hold an iron rice bowl, or taking the postgraduate entrance exam for that purpose, he thinks there is no need at all; but if you have the ability to join a cutting-edge AI lab and become one of the people creating all of this, that job is worth doing.
As for how to explore the world after not working, and how to earn the income you want, Guo Yu gave a very specific piece of advice: go take a look in the south, and see how people in Shenzhen face all of this. When he was in university, he had two roommates from Chaoshan. Around 2010 and 2011, they never thought about going to work after graduation, because the family outlook was "rather be a small boss than work for someone else." They themselves could not quite explain where this outlook came from, but it had been passed down from generation to generation and had already become a mental brand stamped into their minds."He felt that working for others would never lead anywhere. I think this idea holds up very well in the AI era: be your own boss, be your own OPC. If you don't know how to be your own boss, go take a look at Shenzhen." [68:20]
Indigo added that the Bay Area and San Francisco in the US are also places like this. He particularly agrees with the stock metaphor: holding onto a stock that keeps sinking, what's truly expensive is the opportunity cost. Before giving up, first choose a direction that can rise, then switch over to it, "the most important thing in investing is not to lose out on opportunity cost."
Extended Reflection
Looking at this conversation as a whole, one hidden thread runs through it from beginning to end: black boxes, attention, and purpose.
In 2015, ByteDance engineers were already coexisting with a system they themselves could not fully explain, approaching results through experimentation; ten years later, everyone using large models is in the same position. The difference is that back then only recommendation engineers needed this mindset, whereas now everyone who uses Agents to get work done needs it. The disagreement between Guo Yu and Indigo about "whether to make plans" is essentially two attitudes toward the black box: one is to maintain understanding and control, the other is to trust the process at the goal level. It is foreseeable that as Harness continues to evolve, the latter will become increasingly feasible, but the "judgment" represented by the former will not depreciate because of this; instead, it will become the key to distinguishing Pro Users from ordinary users.
The second thread is attention. ByteDance's commercial essence is capturing attention; Indigo's Everything is Media is packaging and delivering one's own attention; the typo Hack Guo Yu talked about is a micro-sample of the battle for attention. When the marginal cost of both software and content approaches zero, what remains scarce is only the attention of real people and real people themselves. This may mean that in the future, "product managers" and "content creators" will gradually merge into the same role.
The third thread is purpose. From the spinning jenny to First Class Harness, the means are constantly being replaced, while people are forced to move up to the goal level. At the individual level this is an opportunity, but at the social level it brings a real problem: when most people's goal is "killing time," and only 2% to 3% of people can use AI as an amplifier, a Heisei-era-style leisurely life requires a prerequisite, namely a relatively small wealth gap and a comprehensive social safety net. The reason the Japanese sample holds is precisely because "income is about the same no matter what job you do." Whether the AI era can replicate this prerequisite, rather than further widening the gap, may be the key question for whether the "Renaissance" can truly benefit the majority.
Key Takeaways
AI has not made engineers' lives easier; it has compressed the decision-making burden of an entire team onto one person. The real bottleneck has shifted from writing code to asking questions and making judgments, and the human brain's pace cannot keep up with AI's pace, so finding a sustainable rhythm for collaborating with AI is just as important as learning to use the tools.
Recommendation engines make content different for every individual, and Agents make software different for every individual. When software can be generated instantly and discarded after use, the entire previous-generation startup logic of "build good software and then distribute it" needs to be reexamined. First ask clearly what purpose the software is meant to achieve; if the purpose can be achieved without building software, then there is no need to build software.
Intelligence will stratify, and it will also commoditize. Not every scenario needs the strongest model, and in a fully competitive content market, prices will eventually approach computing cost. Intelligence will become as cheap as milk, and what will be scarce is intent, judgment, and real people.
In an era flooded with AI Slop, appearing on camera as a real person, genuine thinking, and human-to-human connection become the strongest differentiation. Every piece of work you deliver is essentially media, and communities will grow around media.
For young people, the most important thing is not to find a stable job, but to judge whether they are currently standing on a long downhill ski slope. Short-term anxiety is the price paid to avoid a long-term slide; if you don't know how to be your own boss, go to Shenzhen, go to the Bay Area, and see how those who never intend to work for others live.
Note: This article is based on program subtitles; quotations are colloquial transcriptions and have been lightly polished.