Note: This transcript preserves Liang Wenfeng's original words and sentence structures as much as possible, with three types of edits: 1. Removal of filler words, verbal tics, and situational interruptions (breaks, disconnections, etc.); 2. Correction of obvious transcription errors from the web version; 3. Addition of subheadings for readability.
I. Original Intent and Vision: Ordinary People Doing Good with Kindness
When my colleagues and I started this company, our original intent wasn't about how much money we would make, going to the capital markets, or getting listed. That wasn't the goal. The first few dozen people didn't think that way—if they had, they wouldn't have come.
Overall, we are doing this with a great deal of goodwill toward the world. We feel this is useful for humanity, something beyond money. Of course, later on, when we realized the stakes were huge, other temptations arose, but that's a different story. Our starting point, our vision, and the vision we maintain today is not based on maximizing commercial interests. I think this is key.
About twenty years ago, the manager I admired most was Jack Welch, the former CEO of GE. Looking back, most of what he said might be wrong now, but he got one thing right: The most important thing for a company is its vision.
Managing a large company doesn't rely on your rules and regulations; it relies on vision. What is vision? It's not a slogan on the wall; it's how you act, not what you say. It's how you actually operate.
How do we manage so many people and stay organized? Actually, we don't have a formal organization; we are vision-driven. This has pros and cons, and we will try to maximize the strengths in the future, but this is our characteristic. We don't operate by "what KPI I want to achieve"—there are no assessments, only vision.
This vision isn't even written down. It's in the way we do things and our attitude toward the world. Everyone in the company might understand it differently, but the general direction is consistent: doing things with great goodwill toward the world. This is what organizes us.
II. Why We Insist on Open Source
Why are we so persistent about open source? Because the vision itself demands it. Without this vision, you can't organize people. For example, Zhipu also open-sources, but their open-sourcing is different from ours. Theirs feels forced, as if it's not their original intent; for us, it is our original intent.
We were very clear about open source from the start. First, it's the vision; second, we believe that to succeed commercially in AI, open source is beneficial.
This sounds contradictory because historically, open source and commercialization have been in conflict. But I think AI is different. Historically, a software company might have a market of billions; if it open-sources, that might shrink to millions. But AI is big enough that it might eventually account for 10% of human GDP. You cannot monopolize this; you must share it with others, or you won't survive.
If we want to succeed with AI in our hands, we first need restraint. You can't think that a certain percentage of human or Chinese GDP belongs to you—the more you think that, the less likely you are to succeed. The more you restrain yourself, the more likely you are to succeed. This is a commercial consideration, a macro one.
We are just a group of ordinary people. If there's a narrative I like, it's: a group of ordinary people doing extraordinary things, rather than a group of geniuses doing extraordinary things.
III. Restraint: API Pricing and the "Ten-Month Payback"
Our restraint is shown not just in open source but in many aspects. We only seek a reasonable profit. Our API pricing logic is: a reasonable profit is roughly what it takes to buy a batch of equipment and recoup the cost in ten months. While we amortize servers over three to five years financially, commercially, we feel ten months is enough. That is the logic for our V3.2 Flash and other models.
This is not profit maximization. If we wanted to maximize profit, we would set prices higher because demand in this price range is inelastic. If I doubled the price, token consumption wouldn't change much, but revenue would nearly double.
When we lowered prices, many people in the company group chat cheered. They were happy because the purpose of our hard work was to make the model cheap and effective so everyone could use it. This is our consensus.
Some say a ten-month payback is still too high. There is indeed room for further price cuts as models optimize. But this cost—ten-month payback—is something we can achieve that others cannot. For companies like Alibaba or Tencent, their costs are likely several times higher because they lack our optimizations.
I view restraint as a strategy. Sometimes you give up something to gain more. Open source is a form of profit-sharing. It increases our probability of achieving AGI.
IV. Other Forms of Restraint: Not Chasing Users or Every "Sesame Seed"
Last Chinese New Year, when users surged, we didn't chase user retention or monetization. We didn't try to become the next "Super App" or compete with ByteDance or Tencent. We chose a very restrained approach because there are watermelons behind us, and what's in front are just sesame seeds.
We shouldn't grab every sesame seed. Compared to the future of AI, current opportunities are small. If we had spent a lot of money to fight ByteDance for users last year, what would we have gained? Nothing. The AGI opportunity is always much larger.
If AI revenue reaches $1 billion, our cash flow will break even and cover all R&D. This is possible, but it's not our priority. Most of us don't think B-side or C-side revenue is as important as AGI.
V. Open Source Again: No Conflict Under Six-Fold Profit
We will open-source our strongest models. I don't see the benefit of being closed-source. Even if you tell everyone everything, the barrier to entry is high. It's hard for others to deploy it at such a low cost.
As long as we only seek a "six-fold profit" (recouping costs in ten months), open source has no impact on our revenue. If you want a hundred-fold profit, then open source would hurt you because third parties would deploy it at a lower cost than your price.
There is no conflict between open source and commercial payment—provided you are satisfied with a six-fold profit. We aren't worried about others using our models to compete; we only worry they won't be able to deploy them successfully.
VI. Our Goal is AGI: The Technical Roadmap
AGI is our goal. The roadmap is quite clear. Current AI can exceed humans if given perfect context and instructions, but it lacks continuous learning.
AI development is a ladder. Last year was CoT (Chain of Thought). This year is Agents. After Agents, the next bottleneck is continuous learning: how to make a model learn like a human over time without heavy retraining.
After continuous learning, we reach a singularity: the model can develop its own next version and iterate itself. This is a continuous process, not a sudden jump. After that comes Embodied AI, where it enters the physical world to do housework or elderly care.
Our roadmap is: solve "learning to learn" first, then the self-iteration singularity, then Embodied AI. This is the "easiest" path because you can use previous technology to help develop the next. If we did Embodied AI first, it would be a very grueling, manual task.
VII. Dimensionality Reduction: C-side and B-side are Byproducts of AGI
While others fight for C-side traffic or B-side revenue, we view these as byproducts. By aiming for AGI, we occupy a higher technical ground, which allows us to perform "dimensionality reduction" on lower-level applications.
Last year, we didn't put much effort into the C-side, yet users wouldn't leave. This year, B-side revenue is growing well without a dedicated sales team. We are just doing AI, and these are intermediate outputs. This vision attracts better talent and creates a stronger organization.
VIII. The Core Interest: Team Stability
We can be restrained in everything except one thing: Our greatest and only core interest is maintaining team stability.
If the team stays together, we will definitely achieve AGI. Money and resources are not the problem. Our challenge is keeping the core, veteran employees. They aren't here just for money; they want to be in an environment that can achieve AGI. Our talent turnover is much lower than our peers.
We only do the "main line" of AGI: language models, CoT, Agents. We don't do 3D or video generation because we don't think they are on the main line of intelligence. We won't do something just because it's a good business; we only do it if it's on the intelligence roadmap.
IX. Compute: The Gap with the US is Only Resources
The only gap between us and the US is resources. We don't have enough cards. We have about 20,000 H-equivalent compute units, most of which just arrived. We are aggressively expanding. If I could spend 20 billion RMB on cards this year, that would be a huge success.
Talent is not the bottleneck; resources are. Because we have less compute, we have fewer experimental opportunities. The largest US models have ~800B active parameters; in China, we are still at the tens of billions scale. We are an order of magnitude behind.
This is currently unsolvable. We must do the best we can at the scale we can afford (tens of billions of active parameters) before moving to 150B or 250B.
X. The Final Gap: Cost, Time, and Experience
In the end, the gap between models will be in three areas: Cost, Time, and User Experience. Cost is the biggest barrier. Can you provide the same quality at a lower price? Time is second—when can you achieve it?
XI. Commercialization: Always Doing It, Never the Goal
We have been commercializing, but it's never the goal. Focusing on product lines now is a waste of time because things change too fast. We want to share commercial opportunities with partners rather than monopolizing them. We don't have the energy to do everything.
XII. Global Division of Labor
Chinese companies will likely play the role of the largest producer. We will make AI the cheapest. There might not be a huge difference in effect, but Chinese AI will be systematically cheaper.
XIII. Domestic Chips and TileLang: The Waning of NVIDIA's Moat
NVIDIA's CUDA moat is being rapidly dismantled for three reasons:
- AI can now help write code to build equivalent ecosystems.
- New technologies like our TileLang (a high-level language for CUDA kernels) make it easy to rewrite the ecosystem.
- The shift from gaming-derived cards to dedicated AI chips decouples the hardware from the old CUDA ecosystem.
Within a year, the perception that domestic chip ecosystems are "unusable" will be overturned. The only problem is capacity. Our V3 was trained on NVIDIA cards but barely used the NVIDIA ecosystem; we used TileLang. We are working with Huawei to do the same on their chips. I am optimistic about domestic compute.
XIV. Multimodal, Scaling, and Low Cost
Multimodal is a component, like search, not intelligence itself. We will support it in V4. We believe in Scaling—larger scale equals better results—but we are limited by compute. We haven't hit the "Scaling wall" yet; we are far from it.
We care about efficiency because we are ordinary people who understand that using these tools costs money. Also, higher efficiency allows us to train larger models with limited compute.
XV. Those Who Take Less Defeat Those Who Take More
If I try to take 5% of the global GDP through AI, I will be defeated by someone who is willing to take only 1%. The one who wants more will be defeated by the one who wants less. This is a macro law. We only seek a reasonable return.
XVI. Organization: Two Lines, No Overtime
We have two lines: Bottom-up (people do what they want, no KPIs) and Top-down (formal projects like launching V4). We try to ensure "formal" work doesn't exceed 50% of a person's time.
We don't really do overtime. Research needs a relaxed environment. Also, because we are so focused and restrained, we have fewer things to do. Many of our products are imperfect, and we don't rush to fix every detail. That is our culture.
Q&A Highlights
- On Talent: AI talent shortage is temporary. China has too many base model companies; it will eventually converge to 3-4 major players.
- On Data: We can't afford the massive manual labeling costs that US companies pay. Half of our core researchers are effectively "labeling data" because data is the core of AI right now.
- On the US Gap: We are 6-12 months behind the US, but we achieve our results with 1/20th of their compute. Our goal is to close that time gap while maintaining our efficiency advantage.
- On Coding Agents: This is our most certain and important near-term goal. A model that can help us develop the next model is the fastest way to AGI.
- On the Future: AGI is a non-linear process. AI will accelerate AI research. Eventually, we must move into Embodied AI because human needs are physical, not just digital.
- On Being a Company: We aren't Bell Labs. We receive no government funding. We must be a company that can survive. B-side revenue is our insurance for the future.
