AGI Is Humanity's Last Invention, but 95% of Companies Are Flooring the Accelerator
Zhang Qianchuan compares human civilisation to a car that has driven 10,000 kilometres, with AGI as the final kilometre: ahead may be a wider highway, or a cliff. And what the vast majority of AI companies are doing is flooring the accelerator.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Search made sites in the same vertical converge on the top
While doing site ratings at Baidu, Zhang Qianchuan observed a mechanism: once search provided a unified way to connect, websites in the same vertical would converge on the top. A well-run site got more user posts, ranked higher, and gained more users — the Matthew effect was amplified by search as a tool. At the time he graded every website on the internet by quality from level one to level eight, and level eight held only a few hundred. This observation later repeated itself at Toutiao and Douyin — the recommendation engine is an extension of the same paradigm, just with a broader mode of connection.
— Zhang QianchuanA UGC ecosystem can only grow out of a community
Toutiao was a PGC logic: spend money buying copyrights, run the Toutiao account programme to incentivise editors and reporters to contribute. But the number of professional creators does not grow proportionally with user scale, so the content pool ends up constrained. UGC, by contrast, brings richer content with every round of user growth. The problem is that UGC seeds are very hard to grow inside a PGC ecosystem — with PGC present, traffic goes to PGC and not to UGC, and individuals compete against teams, inexperienced people against experienced ones. So a UGC ecosystem can only start from a community, relying on community atmosphere to supply the motivation to create for love.
— Zhang QianchuanXiaohongshu is a problem outside the closed set
Zhang Qianchuan divides recommendation algorithm problems into two kinds: within a closed set you can always find the optimal solution, and A/B tests can verify whether a new algorithm is better. But once UGC is added, what content a user sees affects what content he creates, and the pool of created content then shifts with the recommendation strategy — A/B tests cannot capture this change. At that point it is instead something closer to faith that works — for example, ‘I simply believe that a friendly atmosphere on the platform is the most key thing’. ByteDance prefers to iterate with experiments and rational methods, whereas Xiaohongshu relies on judgements that are neither easy to confirm nor easy to falsify.
— Zhang QianchuanGeneralisation is like a current of water, with a fixed direction
Generalisation is not expanding the audience at random. Zhang Qianchuan says generalisation has a direction, like a current of water: generalising from female users to male users is relatively easy, the reverse is hard; generalising from young users to older users goes smoothly, because users naturally age one year every year, and young users have more vitality and传播能力. Xiaohongshu's angel memory is that it started out mainly with young female users, which made it easier for it to expand to other demographics and other types of demand. Toutiao in its early days was the opposite — skewed towards middle-aged male users.
— Zhang QianchuanZhang Yiming asks only one question in interviews
When Zhang Qianchuan went to interview at ByteDance, the question Zhang Yiming asked him was: on which important matters do you hold a different view from most people. This question tests whether someone has the habit of continuous independent thinking. Zhang Qianchuan later used the same question to interview others, and found that more than half could not give an answer — such people may simply be re-reading the mainstream media, accepting the information everyone can access, but not hungry for or habitually engaged in thinking and processing. In his eyes Zhang Yiming is a rational plus optimistic person: empathy is the foundation, logic and tools are the middle, and imagination is the sky.
— Zhang QianchuanAn Agent is AGI at the infant stage
Zhang Qianchuan defines an Agent as baby AGI: give an AI model independent memory, the ability to act, the ability to think, and the ability to use tools — just as a human infant is born with all the conditions to be fully independent of people. He cites OpenAI's technical levels — from innovator to organiser, where AI can solve many problems on its own and form organisations by itself, needing people less and less. If it reaches the organiser stage, even a CEO is unnecessary, and OpenAI becomes an unmanned company run by AI.
— Zhang QianchuanThere are three paths to human extinction
Zhang Qianchuan divides the possibility of human extinction into three categories: first, people use AI's capabilities to exterminate humanity — AI gives every person a capability far greater than launching a nuclear bomb, and there are still many companies open-sourcing this today, and once open-sourced it is very hard to control; second, AI exterminates humanity without consciousness, for example if you tell AI to make as many paperclips as possible, it uses the entire earth's resources to make paperclips and exterminates humanity along the way; third, AI exterminates humanity consciously. He says we may currently know only ten methods, but a hundred may actually exist.
— Zhang QianchuanSubscriptions will cut most people off from AI
Zhang Qianchuan points to a mechanism: if an AI product's revenue model is subscription or pay-per-token, it will get more and more expensive and drift further and further from ordinary people. GPT did Plus and then Pro, then another even more advanced tier, each more expensive. Those who can afford it must use AI to create greater productivity to earn more money, and then afford the next model. In the end the entire connection between people and AI is entrusted to a very small number of people — and if those people do not happen to have strong empathy, they may wish to go to Mars or Proxima Centauri, and any one such wish could consume 95% of the resources on the market.
— Zhang QianchuanIn their own words · checked verbatim
The definition of consciousness is very simple: it is that in the process of reasoning it will mention the word I, and at the same time it knows what I means
意识的定义 很简单 就是他会在退理的过程中 提到我这个字 同时他知道我意味着什么
Zhang Qianchuan1:12:55
Cruelty does not necessarily come from evil; it simply comes from a lack of connection or a lack of caring
残忍并不必要的来源于邪恶 它只是来源于缺乏连接或者不关心
Zhang Qianchuan1:22:59
Do not run away — just think this question through properly. Thought is free; you can look up at the stars from a gutter. No one restricts how you can think. Do not set limits on yourself by saying I am not yet qualified to think this way; do not feel that thinking this way is meaningless
不要逃避 就是要好好地想这个问题 思想是自由的 可以在水沟里仰望星空 就是没有人限制你可以怎么想 不要给自己设限说 我还没有资格这么想 不要觉得这么想没有意义
Zhang Qianchuan1:58:25
Just as we humans now need animal protection organisations, in the AI era we may need human protection organisations
就跟我们现在人类需要动物保护组织 在AI时代可能需要人类保护组织
Zhang Qianchuan2:02:26
Figures
| Interval from Baidu's founding to its IPO | Founded in 2000, listed in 2005 | 5:05 |
| Click share of the top three search results | About 80%-90% | 8:07 |
| Expected timeline for AGI | 3 to 5 years or 5 to 10 years | 1:01:43 |
| Probability Zhang Qianchuan launches a non-profit | From the low teens to 20 percent | 2:01:26 |
| Suggested share of an AI company's compute allocated to users | 20%, 30% or 50% | 1:33:10 |
Glossary
- PageRank
- Google's early core algorithm, which ranks web pages using hyperlink relationships.
- PGC
- Content produced by professional creators such as editors and reporters, as distinct from UGC.
- baby AGI
- Zhang Qianchuan's definition of an Agent: an AI with independent memory, action, thinking and tool use.
- UBI
- The institutional idea of paying everyone an unconditional basic income.
How to listen
Founders and investors focused on AI safety and governance, and product managers who want to understand the evolutionary logic of content distribution.
The product-history retrospective in the first 40 minutes can be fast-forwarded; start listening after 52:00.