“Liang Wenfeng’s Talk Transcripts” from the group (Part 1)

As well as other colleagues in our company—when we first started this company, our original intention was not to think about how much money I would ultimately make, about going to the capital markets, about listing, or how we would do that. So we didn’t have that goal in mind.
At the very beginning, when we were just a few dozen people, none of us thought that way. If he had thought that way, he wouldn’t have come. So overall, we did this with a very big goodwill toward the world. We believed this was useful to humankind—something beyond money.
Of course, later on, once we found that the benefits were huge, there were other temptations—that’s another matter. But our starting intention, our vision, and the way we have maintained that vision until now are not done in a way that maximizes commercial interests. I think that’s the most important point.

About twenty years ago, in terms of management, the person I admired most was Jack Welch, the former CEO of GE. Looking back now, most of the things he said may no longer be correct, but one of his most important points was right: the most important thing for a company is its vision.
Managing a big company doesn’t rely on your rules and regulations—it relies on vision. What is vision? Vision is not a slogan posted on the wall. Vision is how you do things, not how you talk. It’s how you actually operate. Anyway, I’ve forgotten Welch’s exact original wording, but it’s roughly this meaning.

So how do we manage so many people and organize them? Actually, we’re not organized. It’s vision-driven—organized by vision. We’re not organized in the conventional sense.
This has its pros and cons. In the future, we’ll figure out how to capitalize on strengths and avoid weaknesses, but that’s our characteristic. We don’t do it in a way like “I want to achieve these KPIs and there’s no assessment.” It’s only vision.
Even this vision isn’t written down; it isn’t something we wrote. We haven’t written anything out. This vision exists in the way we do things and the way we treat the world. Perhaps each person in our company understands this vision differently, and perhaps everyone’s vision differs too—but in terms of the big direction, it’s consistent.

I think it’s still with a very big goodwill toward the world, and wanting to do something. We used that as the way to organize ourselves.
Next, I’ll talk first, and after I finish, everyone can ask questions. I may base the rest of my talk around this vision. This vision is real—it’s not something we made up. We truly think this way and truly do it this way. Otherwise, you wouldn’t be able to explain many of the things we’ve done.

Why do we insist on open-sourcing so much? Because this vision itself requires open-sourcing. If you don’t have that vision, you can’t organize people.
For example, Zhipu is also open-sourced, but Zhipu’s open-sourcing is different from ours. Zhipu’s open-sourcing feels somewhat forced—they feel this isn’t their original intention. But for us, this is exactly our intention.

And regarding open-sourcing, from the very beginning we thought it through very clearly. First is the vision itself. Second, we believe that to make AI succeed commercially, open-sourcing has benefits.
This sounds a bit contradictory, a bit against intuition, because historically open-sourcing has conflicted with commercialization. But I think AI is different from before. Because historically, in a software company, the market size in a year might only be tens of billions of dollars. If it’s open-sourced, it basically has no market left—it might end up with only a few tens of millions or a few hundreds of millions of dollars.
But the AI market is big enough. In the end, it might account for ten percent of the GDP of human society. That’s a very large number. If one person monopolizes it, you can’t monopolize it. You must share with others, otherwise you definitely won’t survive.

This is different from open-sourcing a piece of software in the past, because that software market wasn’t that large. But AI is too big. If we want to monopolize the benefits, we will surely be left behind by history. I think the most important part is an objective law—an outlook on history.
It’s not that if we don’t open-source, we can monopolize that market. In theory, 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 your 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 obtain are limited—then you may be able to pull it off. You need restraint. 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 thing—what percentage of human GDP belongs to me—or what percentage of China’s GDP belongs to me. The more you think that way, the less likely you can make it happen.
So from the very beginning, we felt we needed restraint. The more restrained you are, the more likely you can pull it off. This is a business consideration, of course—a macro consideration.
I think it’s intuitive—at least it matches my intuition, or at least it’s what I truly think. We don’t have many other advantages. We don’t have any special abilities. We’re not richer than others. We also don’t have better people than other companies. There really isn’t that.
Think about it: two years ago when we founded this company, we didn’t have much money, no special cards, no reputation, no influence. We were just a group of very ordinary people.
[00:11:49]
I’m really just a group of ordinary people. If you like the narrative “ordinary people achieved extraordinary things,” rather than “geniuses achieved extraordinary things,” that’s closely related to our restraint. It’s in line with our restraint, and with our vision.

So will open-sourcing conflict with commercialization? I think for AI, if you don’t restrain yourself, you won’t succeed.
Open-sourcing is part of restraint. Our restraint isn’t only reflected in open-sourcing—we also show it in many other aspects. But overall, we don’t need to worry about the issue of open-sourcing, and we don’t need to worry about the issue of restraint.
The more restrained you are, the easier it might be to make it work—or at least so far, it has been validated. Up to now, it explains why we can pull it off. We don’t have any weapons; our starting point is very low; our resources are very limited. And our people are basically just a random group of ordinary people. Even I—I'm just a student who graduated from university; I didn’t graduate from the top schools.

This restraint is also part of the vision. AI is too big. The benefits are too big. We’re very restrained. As long as we can make it happen, the benefits will be huge. If you split just a little, the benefits are still huge. So right now you basically don’t need to consider which portion of those benefits to take or how to take them. You don’t even need to think about it, because the benefits are large enough.
If you just take a little bit, it’s already more than enough. So we used to say: we only earn a reasonable profit. We look at your willingness, not at 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 what it would take to buy a batch of equipment on the market back, and recoup the cost in ten months. We think that’s a reasonable profit.
Given the current situation, considering you have risks, the upfront investment, and so on—if, financially, a server is amortized over three years or five years, but commercially we think recouping costs in roughly ten months is enough, then we think it’s enough. So this is the logic behind our current API pricing. Our V3.2 Flash and others all recoup the equipment cost in ten months.
This is our standard. It’s actually not profit maximization. If it were profit maximization, we should set the price higher. Because in that price range, users’ demand is not elastic. If I double the price, or raise the price by one more time, there isn’t much difference in token consumption.
If I double the price, my total revenue would be close to doubling. Wait a second—let me check. Wow, that’s great.

Let me tell you a story: our DDCP model. At first, we worried about there being too much demand, so we set the price relatively high. The team wasn’t very happy. Later I lowered the price again—down to a quarter—and everyone was very happy.
I think that’s our true thinking. It matches the vision I said earlier: we still want this thing to be useful to people. Not that we earn the most money, but that everyone can afford it while we still achieve a reasonable profit.
I think this time, other people in our company also felt the same. When we lowered the price back then, in the company group, many people were cheering. Everyone thought it was great.
Because the purpose of all the time and effort we put into building this model so carefully was to make it very cheap, and the results very good—so that everyone can fully use it. We felt happy about that. This is our motivation. This is our vision. It’s also a consensus that allows our company to come together to do this—our internal consensus.

This part should be somewhat special. Because for our other competitors, lowering prices like this is definitely not a good thing. They definitely wouldn’t be cheering. Because your income, your ARR—if you cut it in half, your ARR would drop by half. That’s a difference from us.
We think this is enough. From inside the company: if we recoup costs in ten months, I’m already very satisfied commercially. From outside the company: we also think the price makes people more comfortable, more willing to see it, and it’s win-win—win-win for the company, for society, and for everyone.

I think, okay—someone earlier messaged on the screen that the profit from recouping in ten months is too high. That’s true; there is room to lower prices more. There is still room to lower prices. In terms of model optimization, there’s still room too, so overall, there’s fairly significant room to cut prices.
But with those costs, recouping in ten months—we can do it, others can’t. For example, maybe Alibaba or Tencent—if they optimize less, their costs should be several times higher than ours. There’s still a lot of optimization work here.

Why didn’t we lower prices further? Because it has no elasticity. If I lower the price again, demand won’t increase much—or even if it increases, it’s only a little. Because at this price, everyone can afford it, and everyone feels satisfied with it. They won’t stop using it just because the price is a bit higher.
So lowering prices first means the company won’t get more revenue. For society, it doesn’t create more value either, because everyone is already satisfied with the price. If you make the price even lower, society’s sense of happiness doesn’t increase that much.
Okay, but regarding that question—when it comes to pricing, we definitely are not using “highest company revenue” or “highest profit” as our starting point. That’s part of our restraint. In the short term, if your price is higher, you might get more revenue. But in the long term, it’s really not clear. Because I think restraint is a strategy.

Restraint is a strategy.

For me, restraint is a strategy. It’s about sometimes giving up a bit so you can get more of other things. Not opening-source—this is similar. It can also be seen as pressure, or as sharing benefits.
First, for our company internally, sharing benefits makes us very happy. Employees feel a strong sense of achievement, and that gives us cohesion. And for society, sharing benefits is also good. Society is also happy; other peers, or ordinary people, would also be happy. This kind of restraint, as I understand it, is that in the long run it increases the probability that we succeed in building AGI.
When considering something, I have no doubt that AGI will have very huge commercial value. On that basis, my priority is not how to add more share—how to get more share. My priority is how to increase the probability that we can make it.
This restraint may also show up in many other ways. For example, last year during the Spring Festival, user numbers suddenly increased a lot, but we didn’t pursue the goal of keeping those users, or monetizing those users, or grabbing those commercial benefits from the users and cashing it in.
We didn’t go after users. We didn’t try to make money. But we worked hard to serve users well. We also wouldn’t have thoughts like: I want to build the next super app, then compete with whoever; I want to build the next ByteDance, the next Tencent. We had no such ideas.
We could do that, but we didn’t. My understanding is that this is also part of restraint. Don’t think you have to earn everything. Once you have users, it feels like you can build the next ByteDance, and then you just “eat” that whole thing.
I think that’s commercially possible—it could work. If last year we had spent a lot of money and gone to抢 users with ByteDance, that would also be a kind of strategy. But we chose a very restrained approach: we don’t fight with you for that. Because after all, later there are watermelons; what’s in front might just be sesame.

I shouldn’t抢 every bit of sesame. Of course, the sesame in front might be bigger, but I think the AI opportunities later are bigger than what’s in front. Looking at it now: maybe last year not focusing on the C-end was correct. Because you can see that in the end there really are bigger watermelons behind, and the front is really just some small sesame.
If last year we had had a lot of money and made this happen at huge scale, what benefits would we have gotten? You wouldn’t gain anything meaningful. These are my real thoughts. Because I believe the AGI opportunities ahead are very huge. The AGI opportunities ahead will always be huge.
I don’t even need to consider whether I would occupy a position in it later, or what business model I would have. We don’t need to consider that. As long as there is such a huge business opportunity, you will definitely find a way to capture it.
So we will still pick up some of that front sesame, but just casually. We won’t stop and treat it as an important thing to do.

So last year’s C-end DAU might be just a small matter. But we did pick it up. We also used relatively low cost to sustain user usage because those users might be useful later—though we don’t know what they will be useful for now. For now it’s just pure cost, but in the future it might matter.
Since it’s something we can pick up easily, we’ll grab it when we can. This year as well: it’s very possible that our ARR income in APIs or AI has opportunities. If demand can continue to expand—if it expands enough so that GPUs can be purchased in more quantities—then ARR reaching a few hundred million USD is very possible.
If AI can reach $1 billion, then basically my company’s cash flow could turn positive—it could cover my R&D expenses and cover all our costs. So that’s also possible. But we haven’t treated that as a priority to do right now.

We will do that. But I think it’s an important thing. It’s not our first priority. Or it’s not something we genuinely care about today. The bigger opportunity should still be ahead. The front opportunities—including last year’s C-end and this year’s B-end—I think they need to be done and done well. But that’s not our goal.
Or most of the people in our company don’t think this is a very important matter, and they don’t think it’s as important as AGI.

We can talk a bit more about open-sourcing. Because before, many questions asked were about open-sourcing. First, I think we will open-source. And our strongest model will also be open-sourced. Because I can’t see any benefit to keeping it closed. I can’t see any inevitable advantages. ByteDance keeps its model closed—what advantages does it have? I don’t see any.

Even if the model is open-sourced and you tell everyone everything, the barrier is still very high. For others to deploy it, the barrier is also very high. It’s hard. And even if they manage to deploy it, they also need to make the deployment cost very low—which is extremely hard. It’s not easy.
It’s not that once I open-source, it will be easy for others to achieve the same deployment cost as me. There’s still a lot of work to do. Although the underlying principles are understood, not every company is willing—or has the willingness and ability—to organize people to achieve that goal.

I’m also used to this. They probably aren’t good at doing this because the resistance is too high. It’s hard to control costs; there are many management and physical constraints. This is also an advantage of startups: if a startup is too small, you don’t have the strength to do this. If you’re a big company, it’s hard to organize.
[00:38:36]
So this has challenges at every stage. For companies of our scale, this is a sweet spot. If we get bigger, maybe we won’t have other problems. If we get smaller, then our friends’ strength won’t be enough.

So as for open-sourcing: I think we should set pricing. We should recognize that we won’t pressure anyone. Regarding the pricing model, I won’t charge a very high fee either. I may also charge at a rate that recoups in ten months. If you recoup in ten months, it can already make independent third parties unprofitable when deploying it independently...
If I recoup in ten months, then independent third parties wouldn’t have profit. They can’t do it. They can’t do that cost either—they definitely can’t.
So open-sourcing won’t affect my income. Of course, if I want to earn a hundred times the profit, then open-sourcing is—

(They can hear you, but the video seems to be dropping, boss.)
That might be the call coming in just now.

I’m saying: open-sourcing has no impact on our business model. The premise is that we only earn six times the profit—ten months to recoup roughly corresponds to six times the profit. If we only earn six times profit, then open-sourcing won’t have any impact.
But if you want to earn one hundred times profit, then open-sourcing would indeed affect your ability to earn one hundred times profit. Because third parties will deploy it, and their cost might be twenty times—so they can undercut you.

Can this model be sustained long-term? I think it can. Within our vision, I think open-sourcing can be sustained long-term. Or at least this is what we intend to do. You can think of it as restraint, and it also gives you a longer-lasting advantage.
This strategy gives us more opportunities ahead of the technology, and it increases the probability that we succeed in building AGI. We’re more at ease.

You know, 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 hard. It really isn’t hard. At the beginning, it might seem like we chose a difficult approach—like doing research, doing the hardest things, like a “hard mode.” But in fact, in the outside world we give up a lot, so we still remain very capable. We’re still doing it very lightly.

So for open-sourcing, my assessment is that it’s sustainable. There’s no conflict between open-sourcing and paid commercial use—provided there’s no conflict in the case of six times profit.
Six times profit sounds high, but it’s actually not that high. Because AI efficiency is so high right now. Under current conditions, a reasonable profit might be around this level. In the future, it might drop to four times or three times. I think that’s—at the very least, it won’t be able to drop much further, but it would still be very profitable.
Just looking at the business of selling APIs alone, I don’t think selling APIs is that attractive. But in any case, you can say: I haven’t seen any conflict.

Then I’m not worried that others will deploy our model and compete with us. Not at all. We even hope they can deploy it. We’ll do our best to support the open-source community so everyone can deploy our models. I’m not worried they’ll steal our business, because the market is big enough. I only worry that they can’t deploy it, or that they get some details wrong, and then the results become worse— or their costs become higher.
Yes, there’s no conflict here.

When we had To B business last year, the question asked more often was: if I open-source on the C-end, will it conflict with my C-end? Because I don’t have traffic advantages, and even if Tencent has a lot of its own traffic, if it deploys our open-source model, it will route all its C-end users over to that, and that would steal my C-end users.
But actually, that won’t happen. There are still 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 won’t do something like: open-source a slightly worse model, then when deploying ourselves we use a better model. We won’t do that. It’s the same model.
This also shows that there’s no conflict. Last year, throughout the whole year, on the C-end, I basically open-sourced. We didn’t see any conflict in C-end services. Really, we didn’t see any conflict. So that’s the open-sourcing part.

Our goal should be AGI.

And next, for the company’s long-term vision: I think our goal should be AGI. Everyone’s definition of AI might not be the same, but that doesn’t prevent us from treating AGI as our goal.
From a technical roadmap perspective, the AGI path is relatively clear. With today’s generation of AI technology: if you can describe a problem very clearly, and give it complete context and instructions, it already exceeds humans. But there’s a definition and a prerequisite: you give it complete context, and you give it complete instructions.
And reaching that is very difficult. For example, when we have meetings today, we actually have very long and very strong context in front of us—maybe decades of context for everyone—and this 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. What’s missing is sustained learning. Humans can learn continuously. If you hire an employee, it might take them two months to familiarize themselves with the company environment, learn their job, listen to two months’ worth of things, and then they can start doing work. They can do many things. They can understand what you say—if you say, call that person Wang, they know who Wang is.
But for AI, because of the context limitation, it doesn’t have those two months of learning in advance. If you tell it “call Wang,” you have to tell it who Wang is, his title, where he is, how to find him, and what to pay attention to when finding him. You have to give AI all of that context. In this situation, AI can do it, but you can’t give it all the context—and it’s not realistic.

So AI can’t replace your employees. But if AI has the ability for sustained learning—if it can learn at the company for two months like your employees—then it can replace all of the people. So we’re still missing one step toward the next stage: “learning to learn.”

We can understand AI development as a ladder. Last year’s ladder was CoT, chain-of-thought. Because we found that through the chain-of-thought approach, intelligence can be brought to a higher level. By thinking by itself, it raises the ceiling and lets the AI do more things.
Then we crossed another ladder. This year’s ladder is agents, because we found that using agents, even if there’s a lot to do, its capability range becomes larger, and the intelligent ceiling becomes higher.
Why is it a ladder? Because each later step is based on the previous foundations. Agents use CoT, and CoT also uses previous ladders. The previous ladder is the language model—so none of the steps are taken in a vacuum.
So AI’s progress toward intelligence has a pattern. This year’s ladder is agents, but the agents ladder will also eventually be completed. Once it has solved all possible problems, it still won’t be able to replace your employees—but it will already reach the upper limit of its ability.
Just like CoT: after CoT reaches its upper limit, it already surpasses the top human level in areas like solving math olympiad problems and writing programs. But it stops there. Its technology can’t reach AGI.

So you can see: AI’s path toward intelligence is traceable. After agents, we think the problem that should be solved next is sustained learning—how to let the model keep learning continuously, instead of requiring you to train it heavily and expect it to do long-term continuous learning like humans.
This problem is the same as completing tasks, etc. They’re related and solve the same problem. Standing at the agents stage, what we can see is the next bottleneck: sustained learning. The next problem to solve is how to achieve sustained learning. This is visible and relatively clear.
It’s blocked by obstacles earlier; you have to cross them, and there is definitely a way to cross—but it needs time. After sustained learning, we might arrive at a singularity. This singularity is when the model can sustain learning; then it can already do everything that humans can do. It can develop its own versions, research again, and develop its next version—develop even more advanced AI models. Then it reaches a point where it can iterate on itself.

But this singularity isn’t a sudden singularity; it’s still a progressive process. It might be a relatively long gradual transformation, not a mutation. But by habit, we tend to think it’s a singularity.
Long ago, some futurists predicted there would be a singularity here. But in reality, it’s not a singularity—it’s a continuous process. After this step is completed, that’s when embodied intelligence happens.

This is our conjecture, and we think the timeline should be: first solve learning to learn, then reach the next stage of intelligent singularity—an autonomous self-iterating singularity— and then embodied intelligence. After embodied intelligence, it enters the real world, where it can do household chores, or take care of you in old age.
We think this is an ideal roadmap, but everyone’s view differs—there’s no right or wrong. We just feel this roadmap is the most effortless.
This roadmap is easier because in each step, you need to do very few new things. With this roadmap, we can avoid overtime. But if the roadmap is reversed—for example, if it has to achieve embodied intelligence first—that would be very exhausting, a very bitter job. We don’t want that roadmap. We want a lighter one.
If we solve sustained learning first, then solve the self-iterating singularity, and then solve embodied intelligence—along this path it would be quite light. Because later, you can use the earlier technologies to help develop later technologies. After the singularity, doing embodied intelligence no longer requires humans to do it, and we won’t need to do it either—the model can come out by itself.

So this is how I answer what our long-term goal is. I told him: this is our long-term goal—what we call AGI.

Now back to reality. The most important reality last year was that everyone needed to build chatbots and抢 C-end traffic. Then this year’s reality is that everyone needs to抢 To B revenue, and get involved in this—because if you don’t participate, you’re not on the chessboard, right?
But we don’t think it’s an important thing. Or within our company, what we truly care about is the AGI roadmap I just mentioned, and the technical breakthroughs in the next step.
But there is something strange: the thing you most want to get turns out you can’t get it. The thing you don’t care that much about, ironically, is easier to obtain.

There’s a strategic advantage here: in our hearts, we think about AGI, and we build AGI. Then when we build the applications, when we build C-end and B-end, we don’t need to put too much thought into doing that. It takes very little effort.
I think it’s because if you stand at a higher technical level and work on a relatively lower technical tier, there’s such a dimensionality-reduction attack. At least in last year’s C-end, we saw exactly that. We didn’t put much effort into the C-end. In fact, at one point, we didn’t even want to maintain those users. But we couldn’t drive them away.
Because we really couldn’t drive them away, so in the end they stayed anyway. And so... Looking at this year’s B-end revenue, the growth seems relatively optimistic now. I think the number should be relatively good compared with peers—I guess. But we didn’t put a lot of effort into doing it. We didn’t do anything special—it was just incidental.

When deploying internet intelligence, at some point on the way, the AGI step is a ladder I must climb. On the road to AGI, I have to pass this ladder. So I made all these technologies available to everyone via APIs—I didn’t do extra work.
We’re still building AI, and this is a byproduct. I only need a few people to maintain the API. Even customer service isn’t necessary, and we don’t need sales or anything. Users come by themselves.
Or rather, the users considered as C-end, whether C-end or B-end, are byproducts on our road to AGI. They’re intermediate outputs and don’t conflict with building AGI. We’re not doing C-end or B-end for the sake of doing C-end or B-end. We’re building AGI. And since this byproduct can be produced along the way, we take it and commercialize it.

That’s different from other companies. Other companies do it in order to serve C-end users or B-end users with that model. For us, our original intent isn’t that—we still aim for AGI.

I think to a certain extent, this is a kind of dimensionality-reduction attack. AGI is a bigger vision. That vision can attract more excellent people and has stronger cohesion. So I have an organizational advantage. And I use this advantage to—this is a kind of dimensionality-reduction attack.
But if you’re a commercial company whose vision is to serve C-end users, that’s another story. It will have advantages in product, user service, and traffic—but it won’t have the technical advantage.

The current relatively favorable situation is that model technology is the most important. If you get the model right, everything else...
That also explains the trajectory we took earlier. We truly chose AGI, and I never thought about building lots of users.
When we suddenly became popular during last year’s Spring Festival, that wasn’t in our script at all. We never thought about that. We just wanted to build the technology well.
But I later found that our organization, compared with organizations that are fully commercialized and product-driven, has an additional advantage in talent and organization.
That’s also quite amazing. At that time, everyone on the C-end was fighting desperately for growth—yet it got taken by someone who wasn’t even fighting. That also proves the logic I mentioned earlier is sound: it really is an advantage in talent and organization.

This talent advantage doesn’t mean my people are smarter than theirs. It’s about how I organize these talents, how I motivate them, and how we cooperate. Those are the advantages. Because gathering smart people doesn’t automatically make them cooperate, and doesn’t automatically make them passionately run toward a goal and complete it—so you need a vision.

The lesson from our past experience is that the AGI vision is powerful.