“Liang Wenfeng’s Conversation Transcript from the Group” (Part 1)

As well as our other colleagues in our company: when we first came to build this company, our original intention was not to think about how much money I would ultimately make, or to take it to the capital markets, to list it, or how we would do that. So we didn’t have those ambitions.
At the very beginning, when we were only a few dozen people, nobody thought like that. If he had thought that way, he wouldn’t have come. So overall, we did this with a very big goodwill toward the world, and we thought it was useful for humanity—something beyond money.
Of course, later on, after we discovered how much profit there was, there were other temptations—there’s another story there. But our starting intention, our vision, and the way we’ve kept that vision to this day are not about doing things in a manner that maximizes commercial interests. I think that’s the key point.
About twenty years ago, in terms of management, the person I admired most was Jack Welch, the former CEO of GE. When I look back now, most of what he said may already be wrong, but his most important point was right: a company’s most important thing is its vision.
Managing a big company doesn’t rely on your rules and regulations—it relies on vision. What is vision? Vision isn’t slogans hung on a wall. Vision is not what you say—it’s how you actually operate. Anyway, I don’t remember Welch’s exact wording, but it was something like that.
So how did we manage all these people, and how did we organize? In fact, we didn’t organize in the usual sense—we’re organized by vision, using vision to organize. We don’t have “organization” in the traditional sense.
This has advantages and disadvantages. In the future, we’ll find ways to leverage strengths and avoid weaknesses, but this is our distinctive feature. We don’t do it in a way like, “I need to achieve certain KPIs, with no assessment”—it’s only vision.
Even this vision isn’t written down. It isn’t something we wrote. We haven’t written anything down. The vision exists in the way we do things, and our attitude toward the world. Maybe each person in our company understands this vision differently, and maybe each person’s vision differs. But in a broad direction, they’re consistent.
I think we’re still doing this with very big goodwill toward the world—wanting to do a bit of something. We used that to organize ourselves.
Next, I’ll talk about it first. After I finish, everyone can ask questions. I may focus on talking about the things that come next around this vision. This vision is real; it wasn’t invented. We genuinely think this way and genuinely do this. Otherwise, you wouldn’t be able to explain many of our actions.
Why are we so persistent about open-sourcing? Because this vision itself requires open-sourcing. Without this vision, you can’t organize people.
For example, Zhipu also open-sources, but their open-sourcing is different from ours. Zhipu’s open-sourcing feels somewhat forced—they feel this isn’t their original intent. But for us, this is precisely our original intent.
And in terms of open-sourcing, we thought very clearly from the start. First, the vision itself. Second, we believe that to make AI a commercial success, open-sourcing has benefits.
This sounds a bit contradictory, a bit against intuition, because historically, open-source has conflicted with commercialization. But I think AI is different from the past. In the past, in a software company, the market in a year might be tens of billions of dollars. If you don’t open-source, you’re left with only a few hundred million or a few billion dollars.
But AI is so big that in the end it may account for ten percent of human society’s GDP. If so, that’s a very huge number. If one person monopolizes it, it’s impossible—you can’t monopolize something like that. You must share it with others; otherwise, you definitely won’t survive.
This is different from open-sourcing a piece of software before, because that software market isn’t that large. But AI is simply too big. If we wanted to monopolize the benefits, history would leave us behind. I think this is the most important objective law—an outlook on history.
It’s not that if I open-source, I lose the market; and it’s not that if I don’t open-source, I can monopolize it. Theoretically, that doesn’t match objective facts. You will definitely encounter a lot of resistance, and there will definitely be other methods to stop you from achieving that goal.
In this situation, I think you don’t necessarily need to follow traditional commercial thinking. You need a set of mechanisms to ensure that the benefits you can gain are limited—then you can make it. I think restraint is needed.
If we want to make AI succeed in our hands, first of all, I think restraint is needed. You can’t think that “the whole percent of human GDP should belong to me,” or “the percent of China’s GDP should belong to me.” The more you think that way, the less you’ll be able to do it.
So from the beginning we felt we needed restraint. The more restrained you are, the more likely you are to pull this off. This is a commercial consideration—of course, it’s also a macro consideration.
I think this is consistent with intuition—at least with my intuition, or at least I truly think this way. We don’t have a lot of other advantages. We don’t have any special abilities. We’re not richer than others, and we don’t have people who are better than other companies. Actually, that’s not the case.
Think about it: when we founded this company two years ago, we didn’t have much money, we didn’t have many cards, and we didn’t have much reputation or appeal. We were just a group of very ordinary people.
[00:11:49]
I’m really just a group of ordinary people. If the kind of narrative you like is “a group of ordinary people did something extraordinary,” rather than “a group of geniuses did something extraordinary,” then that’s very related to our restraint. It’s in line with our restraint and with our vision—part of the same thread.
So, will open-source conflict with commercialization? I think with AI, if you don’t restrain yourself, you won’t be able to do it.
Open-sourcing is part of restraint. Our restraint isn’t only shown in open-sourcing; it shows in many other aspects as well. But overall, we don’t need to consider the issue of open-sourcing, and we don’t need to consider the issue of restraint.
The more restrained you are, the easier it might be to succeed—or at least so far it’s been proven and explained. Otherwise, we wouldn’t be able to explain why we were able to succeed: we don’t have weapons, our starting point is very low, and we have very few resources. Our people are basically a random bunch of ordinary people. And I myself am just a college graduate, not from the top schools either.
This restraint is also part of our vision. AI is too big, and the benefits are too big. We’re very restrained. As long as we can pull it off, the final benefits will be very large. If you just split a little, the benefits are already very large. So right now, you basically don’t need to think about which portion of those benefits you take, or how you take it—you don’t need to consider this. Because the benefits are big enough.
If you take just a little bit, it’s already more than enough. That’s why earlier we said we only earn a reasonable profit, and we look at your willingness rather than how huge the profit is. That’s different. This isn’t our API pricing. For our API pricing, we think a reasonable profit is roughly: we buy a batch of equipment from the market, and the payback period is ten months. That seems like a reasonable profit.
Given the current situation, considering you have risks, and you have upfront investments, etc.: if for a server we amortize it over three years or five years on the finance side, but from a business perspective we think the payback should be around ten months—that seems enough. OK, we think it’s enough. So this is the logic behind our current API pricing. Our V3.2 Flash and others are all based on recovering the equipment cost in ten months.
That’s our standard. It’s not really about maximizing profit. If you were maximizing profit, you should set the price higher. Because in that price range, user demand has no elasticity: even if I raise the price by half, or double it, the token consumption wouldn’t be meaningfully different.
If I double the price, then my total revenue would be close to double. Wait a moment—I’ll check. Wow, that’s great.
Let me tell you a story. Our DDCP model: at the beginning, we were worried that demand would be too high, so we set the initial price relatively high. The team wasn’t very happy at the time. Later I lowered the price again, down to one quarter, and everyone was very happy.
I think that’s our real thinking. It’s what I said earlier about the vision: we still want this thing to be useful to people, not to earn the most money. It’s about everyone being able to use it when we can earn a reasonable profit. I think what other people in our company thought back then was: when we lowered the price, many people in the company group were cheering. Everyone felt really happy.
Because that’s the purpose of all the effort and care we put into making this model. The purpose is to make it very affordable, with very good performance—so that everyone can fully use it. We felt really happy about that. That’s our motivation, that’s our vision, and it’s the consensus that lets our company come together to do this. I think that’s the internal consensus of our company.
This is probably a bit special. Because for our other competitors, a price cut like this is definitely not good news. They definitely wouldn’t cheer. Because your revenue and your ARR would drop by half. Yes, that’s the difference with us.
We think that’s enough. From inside our company: getting payback in ten months, commercially, I’m already very satisfied. From outside the company: we also feel that this price is something everyone is relatively happy about and more willing to see. It’s a win-win—our company wins, society wins, and everyone wins.
I think, OK—someone just commented on the screen that the profit from paying back in ten months is too high. It’s true that there’s still room to reduce prices, and there’s indeed room to cut further. There’s also room to optimize the model, so overall there’s still a fairly large space to lower prices.
But this cost—payback in ten months—we can achieve it, but others can’t. For example, maybe Alibaba or Tencent: their cost should be several times higher than that. There’s still a lot of optimization work here.
Why don’t we keep cutting the price? Because it has no elasticity. If I lower the price further, demand won’t increase more, or demand would increase very little. Since at this price everyone can afford it, and everyone feels the price is satisfactory, nobody would stop using it just because it’s too expensive.
So lowering prices first doesn’t bring more revenue to the company. For society, it also doesn’t create more value. Because at this price everyone is already satisfied. If you make it even cheaper, society’s happiness doesn’t increase much either. Yes—OK. But on that question, regarding pricing, we definitely aren’t starting from “maximizing company revenue” or “maximizing profit.” That’s part of our restraint.
Because in the short term, if your price is higher, you might make more revenue. But in the long term, it’s really not certain. Because I think restraint is a strategy.
Restraint is a strategy.
For me, restraint is a strategy. It’s about sometimes you can give up some things to get more other things. Not open-sourcing is the same way—it can also be seen as our pressure, or as our willingness to share the benefits.
First, internally within our company, we’re very happy. Everyone is very happy. Employees feel a strong sense of accomplishment, and we will have greater cohesion because of it. And this willingness to share benefits is also good for society—society is happy, and other peers or ordinary people are also happy.
So I understand this kind of restraint as: in the long run, it can increase the probability that we succeed in AGI.
When considering this matter, I have no doubt that AGI will have enormous commercial value. On that basis, my priority isn’t how to get more share, or how to take more share—I prioritize how to increase the probability that I can actually make it.
This restraint might also show up in many other aspects. For example, last year around the Spring Festival, user numbers suddenly surged. But we didn’t pursue “keeping these users,” or “turning these users into monetization,” or “grabbing those commercial benefits,” or cashing in with users. We didn’t go grab users—we didn’t go make money. We worked hard to find ways to serve users well.
We wouldn’t have the thought of, “I want to make the next super App, and compete with someone; I want to be the next ByteDance, the next Tencent”—we have no such ideas.
We could do it. But we didn’t. My understanding is that this is also part of restraint. Don’t think that you have to earn everything. You think once you have users, you can just become the next ByteDance, and then you go and eat up everything.
I think that’s commercially workable—it might be possible. If last year we had used a lot of money to go抢 users against ByteDance, that would also be one kind of strategy. But we chose a very restrained approach: we don’t compete with you for that. Because after the watermelon, the front is just sesame.
I shouldn’t go抢 every sesame. Of course, maybe some sesame is quite big. But I think the AI opportunities later will dwarf what’s in front. Judging from now, last year not going hard at the C-end may be right. Because you can see that there really is a bigger watermelon later, and the front is really just some small sesame.
If last year I had a lot of money and made this into something very big, what good would it bring? You wouldn’t get anything. Those are my real thoughts. Because I think the AGI opportunity later should be extremely big—and the AGI opportunity later will always be extremely big.
I don’t even need to consider whether at that time I’ll occupy a position inside it, or what my business model would be. We don’t even need to think about it. As long as there’s such a huge business opportunity, you’ll definitely find a way. So for the sesame at the front, we’ll pick it up too, but we’ll just pick it up casually. We won’t stop, and we won’t treat it as an important thing to do.
So those C-end metrics like daily active users from last year—maybe they were just a small matter. But we picked them up, and we maintained user usage with relatively low costs, because maybe later it will be useful. Even though we don’t know what those users will be useful for now, currently it’s just a pure cost. But later, it might be useful.
Since it’s something we can get casually, we’ll take it conveniently. Looking at this year too, it’s very possible that we also have opportunities in API or AI related ARR revenue. If the demand can continue to expand, and if we can buy more GPUs and more GPUs become available, then reaching several hundred million USD in ARR is very possible.
If AI can reach ten billion USD, then basically our company’s cash flow could turn positive—able to cover my R&D expenses and cover all our expenses. So this is also possible. But we haven’t treated it as a top priority.
We will do it, but I think this is an important matter. It’s not the first priority we consider today, or the thing we truly care about right now. The bigger opportunity should be still ahead. The opportunities in front—like last year’s C-end and this year’s B-end—I think those things should be done and done well, but they aren’t our goal.
Or, most people in our company don’t think it’s a very important matter, and they don’t think it’s as equally important as AGI.
We can talk a bit more about open-sourcing, because previously many questions asked were about open-sourcing. First, I think we will open-source. And our strongest model will likely also be open-sourced.
Because I can’t see any benefit to being closed-source, and I can’t see a necessary benefit. ByteDance’s model is closed-source—what benefit does that have? I can’t see any benefit.
Even if the model is open-sourced, if you tell everyone everything, that barrier is very high. For others to use it, the barrier is also very high. They’ll find it hard to use. Also, to use it, they must reduce their costs to very low levels—that’s also very difficult and not easy.
It’s not that when I open-source, it would be easy for someone to deploy with the same cost as me. There’s still a lot of work to do. Even though the principles behind these works are understood, not every company is willing—or has the intention and capability—to organize manpower to achieve this goal.
I’m also used to this. They might just not be good at doing it because the resistance is too high. It’s hard for them to control the costs. There are many constraints, both managerial and physical. This is an advantage of startups. If a startup is too small, it doesn’t have the strength to do this. If you’re a big company, it’s hard to organize people.
[00:38:36]
So there are challenges at every scale level, which means for companies of our size it’s a “sweet spot.” If we get bigger, we might not have those other problems; if we get smaller, we might not have enough power from “friends.”
So regarding open-sourcing, I think we should set pricing appropriately. We should recognize that it won’t pressure people into it. For the pricing model, I also won’t charge a very high fee. I might also charge at the payback-in-ten-months level. With payback in ten months, it would already make independent-deploying third parties non-profitable…
With payback in ten months, it would allow independent-deploying third parties to have no profit. They can’t do it—they can’t achieve those costs.
So open-sourcing won’t affect my revenue. Of course, if I want to earn one hundred times the profit, then open-sourcing would be…
( I can hear your voice, but the video seems to have dropped, boss. )
Just now, maybe the call came through.
I mean: open-sourcing, I think, has no impact on our business model. The premise is that we only earn six times the profit. A payback in ten months roughly corresponds to six times profit. If we only earn six times profit, then open-sourcing won’t have much impact.
But if you want to earn one hundred times the profit, then open-sourcing would indeed affect your ability to earn one hundred times profit. Because third parties would deploy; their costs might be twenty times, and thus they can undercut you.
Can this model be sustained in the long term? I think it can. Within our vision, I think open-sourcing can be sustained long term, or at least that’s what we intend to do. You can think of it as restraint—also a willingness to share, leading to longer-term advantage.
This strategy gives us more opportunities ahead of technology, and it increases the probability that we can make AGI. We can be more relaxed.
You think, we don’t even need overtime, because it’s not that hard. But for other companies, it might be very hard—because you’re thinking too much.
Actually, it’s not difficult. It’s really not difficult. At the beginning it’s like… from outside, maybe it looks like we chose a very hard mode: we’re doing research, we’re doing the hardest thing, it seems like a hard mode. But in reality, we sacrificed many things outside, which makes us still very capable—we’re still doing it very lightly.
So regarding open-sourcing, my judgment is that it’s sustainable. There’s no conflict between open-sourcing and charging commercially, as long as there’s no conflict at the six-times-profit level.
Six-times profit looks high, but it’s not that high. Because AI efficiency is already so high now. What counts as a reasonable profit may be about that much. In the future it may drop to, say, four times, three times. I think it’s already… and it can’t drop any further than that. But it will still remain a large profit.
Just looking at selling APIs alone, I don’t think selling APIs is that attractive. But for this situation, you can say: I haven’t seen any conflict.
And I’m also not worried that others will deploy our model and compete with us. Not at all. We actually hope they can deploy it.
We’ll do our best to provide help to the open-source community, and assist everyone in deploying our model. I’m not worried that they’ll steal this business from me, because the market is big enough. What I’m worried about is that they might not be able to deploy it. If they get some details wrong, the performance could be worse, or their costs could end up higher.
Yes—there’s no conflict here.
Last year, when there was a To B business, people asked more: since our C-end is open-sourced, will it conflict with my C-end? Because I don’t have traffic advantage. Or Tencent itself has lots of traffic—if it deploys our open-sourced model, won’t it connect all C-end users and take away my C-end users?
But actually, that won’t happen for many reasons.
And the question is: is the open-sourced model we provide the same as the model we deploy ourselves? It’s the same. We wouldn’t open-source a worse model and then use a better one ourselves when deploying. No—it's the same.
This also shows there isn’t actually a conflict. Last year throughout the year, on the C-end, basically I open-sourced. And we didn’t see any conflicts in the C-end services. Really, we didn’t see any conflict. So that’s this part about open-sourcing.
Our goal should be AGI.
Then there’s also the company’s long-term vision. I think our goal should be AGI. Everyone’s definition of AI might not be exactly the same, but that doesn’t prevent us from using AGI as our goal.
From a technical roadmap perspective, the AGI route map is actually fairly clear. Compared with the current generation of AI technology: if you can describe a problem very clearly to it, and provide it with complete context and instructions, it has already surpassed humans. But there’s a definition and a prerequisite here: you give it complete context, and you give it complete instructions.
And that definition is hard to meet. For example, today in meetings, there’s actually a very long and strong context in front of us. Everyone might have decades of context. That’s something AI doesn’t have. What AI can do now is within a limited context it can perform better than humans.
But it still can’t replace humans. There’s one more missing piece: continuous learning. Because humans can continuously learn. If you hire an employee, it might take them two months to familiarize themselves with the company’s environment and their job. After chatting for two months, they can get up to speed. Then they can do many things. They understand what you say—if you say, “Call that guy Wang,” they know who Wang is.
But for AI, because the context doesn’t include those two months of learning, if you tell it to call Wang, you have to tell it who Wang is, his job title, where he is, how to find him, and what to pay attention to when looking for him. You’d have to give AI all that context. In that situation, AI can do it, but you can’t give it all the context, and it’s unrealistic.
[00:51:02]
So AI can’t replace your employees. But if AI had continuous learning capability—if it could learn at the company for two months the same way employees do—then it could replace all people. So we’re still one step away from the next thing: “learning to learn.”
We can understand AI development as moving up a ladder. Last year’s step was CoT, or chain-of-thought. Because we discovered that using chain-of-thought can bring intelligence to a higher level. Through thinking itself, it can raise the upper bound, enabling the AI to do more things.
Then we crossed another ladder this year: the ladder of Agents. Because we found that using Agents, even when there are many tasks, it can do them. Its capability range is larger, and its intelligence upper bound is higher.
Why call it ladders? Because every later step is built on the foundation of earlier ones. Agents need to use CoT, and CoT also uses earlier ladders. The earlier ladder is the language model. So it’s not that any step is wasted.
So there’s a traceable path in where AI intelligence is heading. This year’s ladder is Agents—but the Agents ladder will also eventually be completed. It will solve all possible kinds of problems. But it still won’t be able to replace your employees; it just reaches the upper limit of what it can do.
Like CoT: once CoT reaches its upper limit, it’s already beyond the most top-tier humans in math olympiad problems and programming. But it still stops there. The technology hasn’t reached AGI.
So you see, AI intelligence has a traceable direction. After Agents, the problem we think should be solved is continuous learning—how to make the model continuously learn rather than requiring you to give it strong training so it can learn over a long time like people do.
This problem is related to completing tasks, etc. They’re connected, and they solve the same underlying issue. Standing at the Agents stage, what we can see is the next bottleneck: continuous learning. The next problem to solve—how to enable continuous learning—is visible and relatively clear.
It’s stuck behind the obstacle in front of it. You must cross it, and there’s definitely a way to cross it, but it takes time. After continuous learning, we might reach a singularity point. That singularity would be: once the model can continuously learn, it would already be able to do all the things humans can do.
It would be able to develop its own version, research again, and develop its next version—develop more advanced AI models. So it would reach a singularity where it can iterate itself.
But this “singularity” isn’t really a singular moment—it’s a gradual process. This process might also be a relatively long gradual change, not a sudden mutation. But by habit, we all think it might be a singularity.
Because long ago, those prophets thought there would be a singularity here. But in reality, it isn’t a singularity. It’s a continuous process. And only after this step is done, I think that’s when embodied intelligence arrives.
This is our speculation—this is what we think the timeline should be: first solve learning to learn, then reach the intelligent singularity—an iterative singularity—and only then embodied intelligence.
After embodied intelligence, it enters the real world: it can do household chores for you, and it can care for you as you age.
We think this is an ideal roadmap, but everyone’s view is different, and there’s no right or wrong. We just feel this roadmap is the most relaxed.
It’s relaxed because in each step you only need to do very few new things. This roadmap might not require overtime. But if the roadmap were reversed—if it first needed to achieve embodied intelligence—then you’d be doing something very exhausting. That would be a hard job. We don’t want a roadmap like that. We want something that’s lighter.
If we first solve continuous learning, then solve the self-iterative singularity point, and then solve embodied intelligence, that road will be pretty light. Because later on, you can use earlier technologies to help develop later technologies. After the singularity, to reach embodied intelligence, you wouldn’t need people to do it—you wouldn’t need us to do it. The model would come out by itself.
So this answers what our long-term goal is. I told him: that is our long-term goal—AGI.
Now back to reality. The most important reality last year was that everyone had to do chatbots and compete for C-end traffic. So what about this year? The reality is that everyone has to compete for To B revenue and participate in it—because if you don’t participate, you’re not on the playing field, right?
But we don’t think it’s an important thing. Or within our company, what we truly care about is actually the AGI roadmap I just talked about, and how the next step in technology will break through.
But there’s something a bit strange: the thing you want the most, you actually can’t get. The thing you don’t care that much about is actually easier to get.
There’s a strategic advantage here. In our minds, we’re thinking about AGI, and we’re doing AGI. When we then build those applications and do the C-end and B-end, we basically don’t need to spend much mental effort to do it. In fact, you can do it with very little effort.
I think it’s like: if you stand at a higher technical level and do something at a relatively lower technical level, you’ll have a dimensionality-reduction “attack.” At least in last year’s C-end, we saw that clearly. We didn’t spend much effort on the C-end. Even at one point, we didn’t want to maintain those users. But users couldn’t be chased away. Really, they couldn’t be chased away, so they stayed.
Then a bit… and this year’s B-end revenue, looking at it now, the growth looks still fairly optimistic. I think that number, compared with peers, should also be relatively good. I estimate it is. But we didn’t spend a lot of effort on this. We didn’t do a whole deliberate task—it was incidental.
When you go online for internet intelligence, the AGI step is one I must take—it’s a ladder. On the road to AGI, I need to pass through this step. Then I provide those technologies to everyone through API. I didn’t do anything extra.
We’re still working on AI—that’s a byproduct. I just need a few people to maintain the API. Even customer service might not be necessary. No sales either—no need for anything. Users come themselves.
Or, the so-called C-end and B-end users—both are byproducts on the road to AGI. They’re intermediate outputs, with no conflict with doing AGI. I’m not making C-end or B-end for the sake of doing C-end or B-end. Our goal is AGI, and coincidentally it can produce those outputs, so we monetize them.
That’s different from other companies. Other companies do it in order to serve C-end users, or to serve B-end users. For us, the original intention isn’t like that. Our original intention is still to pursue AGI.
I think to a certain extent this is a form of dimensionality reduction. AGI is a bigger vision. This vision can rally more excellent people; it has stronger cohesion. So in terms of organization, we have an advantage. And I leverage that advantage…
It’s dimensionality reduction.
But if you’re a commercial company whose vision is to serve C-end users well, then it’s a different story. They have other advantages: in product, user service, and traffic. But they don’t have technical advantages.
What’s currently favorable is that model technology is the most important. If you make the model well, then everything else…
And this can explain our earlier development path. We really chose AGI. I never thought about making lots of users.
When we suddenly became popular around last Spring Festival, that wasn’t in our script. We never thought of that at all. We just wanted to make the technology good.
But I found that our organization has additional advantages in talent and organization compared to a fully commercial organization targeting products.
That’s also very strange. At that time, everyone on the C-end was fighting fiercely for users, and yet we were “pulled along” by an organization that didn’t go抢. This indeed shows that the logic I mentioned earlier makes sense—there really are advantages in talent and organization.
This talent advantage doesn’t mean my people are smarter than theirs. It’s how we organize them, how we motivate them, and how we get them to collaborate. That’s where the advantage is. Because simply gathering smart people together doesn’t mean they will naturally collaborate, and it doesn’t mean they will naturally have passion to race toward a shared goal and achieve it. You need a vision.
The lesson from our prior experience is that the vision of AGI is very powerful.