DeepSeek Founder's 4-Hour Interview: How Ordinary People Build Extraordinary AI
DeepSeek Founder's 4-Hour Interview: How Ordinary People Build Extraordinary AI
> Based on the complete 4-hour recording of Liang Wenfeng's investor meeting on May 20. Original published by "AI Coordinate". Transcript not verified by the speaker.
DeepSeek is one of the most talked-about AI companies in China right now. With minimal resources and a shoestring budget, they've built world-class open-source LLMs. The question everyone asks: how?
In this rare 4-hour investor meeting, Liang Wenfeng opened up and answered the question from a higher dimension — vision, values, and organizational philosophy.
Here are 10 key insights, presented in plain language.
1. Core Vision: Benevolence and Restraint
> "We are a group of ordinary people who want to do something extraordinary."
This is the foundation of the entire conversation. Liang repeatedly emphasized that DeepSeek was never founded to "make big money, go public, or cash out." It started with "a tremendous goodwill toward the world — to do something useful for humanity."
Why does vision matter so much? Because AI is too big. It could account for over 10% of human GDP. If you try to monopolize it, history will abandon you. This isn't moral preaching — it's objective law: the more you try to hoard, the more resistance you face, and the less likely you'll succeed.
His logic chain is crystal clear: Goodwill → Restraint → Focus → Ease → Sustainability. Goodwill leads to restraint. Restraint means doing only what matters most. Doing only what matters means no overtime, no anxiety, and the stamina for the long run.
My take: This sounds "anti-business," but it's actually the most pragmatic strategy. In a race where winner-take-all seems possible, choosing restraint is its own moat — it attracts like-minded people, eliminates wasteful battles, and lets you breathe in a marathon.
2. Open-Source Strategy: Ship Your Best Models, Fear No Disruption
> "I don't see any necessary benefit in closed-source. ByteDance's models are closed-source — what advantage do they have? I don't see any."
This is one of the most striking viewpoints in the entire interview. Liang's logic is deeply counterintuitive:
The open-source paradox: Traditional business logic says open-source = giving away your crown jewels = destroying your business model. But Liang believes AI is fundamentally different:
1. The market is massive. AI isn't a $5B niche. It might be the largest market in human history. At that scale, you can't go it alone. You must share.
2. Open-source deployment is still hard. Even with all weights and code public, matching DeepSeek's deployment cost requires immense engineering optimization. Not every company can do it.
3. Restraint is competitive advantage. If you only aim for 6x profit (10-month hardware payback), open-source doesn't hurt your revenue. If you want 100x profit, it does — but someone who wants 10x will beat the person who wants 100x.
My take: This isn't idealism — it's calculated strategy. DeepSeek uses open-source as a "filter" — it screens out quick-profit competitors while attracting the best developers and users from the global open-source community. It's advancing by retreating.
3. API Pricing: 10-Month Payback Is Enough
> "Our pricing logic is simple: buy servers, make the money back in 10 months. When we announced the price cut, the company chat was full of cheers."
Two key insights behind this pricing:
First, price has no elasticity anymore. At current API prices, users no longer turn away because it's "too expensive." Further price cuts won't create more demand — just less revenue. So the current price is a "win-win point": fair profit for the company, affordable for users.
Second, cost advantage is the real moat. "We can achieve 10-month payback. Alibaba or Tencent can't — their costs are probably several times higher." This is why open-source doesn't scare them. If your cost is higher, you can't win a price war.
Even more telling: when DeepSeek cut prices, the internal chat erupted in cheers. Because "this is what we worked so hard for — making great models that everyone can afford." This is the engineer's logic, not the profit-maximizer's.
4. AGI Roadmap: A Clear Staircase
Liang laid out a surprisingly clear AGI roadmap, like climbing stairs:
Step 1: Language Model (done) → Step 2: Chain-of-Thought (done) → Step 3: Agent (in progress) → Step 4: Continuous Learning (next breakthrough) → Step 5: Self-Iterating Singularity → Step 6: Embodied AI
His core judgment:
- Every step builds on the previous one. Nothing is wasted. This year's Agent can't exist without last year's CoT, just like you can't skip the first floor when building a skyscraper.
- The current bottleneck is "continuous learning." Today's AI can surpass humans given complete context, but it can't "spend two months getting up to speed" like a new hire. Once AI has continuous learning, it can truly replace human workers.
- Scaling is nowhere near its limit. "When Silicon Valley says scaling has hit a wall, that's for them. For Chinese companies, we're nowhere near it" — because we simply don't have enough compute to hit that wall.
My take: The beauty of this roadmap is that it lets DeepSeek move "effortlessly." Each stage completed naturally reveals the next. Unlike companies that spread thin across video generation, 3D, and world models, DeepSeek stays laser-focused on the main thread — whatever raises the ceiling of intelligence.
5. Core Interest: Team Stability Is the Only Non-Negotiable
> "As long as I can maintain team stability, I will definitely achieve AGI. Money is not the problem. Resources are not the problem. The only core interest — team stability."
This is the single most powerful sentence in the entire interview. Liang defines "team stability" as the only non-negotiable. Everything else can be compromised.
Why? His calculation is simple:
- AI's direction is relatively certain (the roadmap is clear)
- Money and resources can always be solved (demand is large enough)
- The only thing irreplaceable and time-dependent is the team
That's why DeepSeek does many seemingly "non-commercial" things — no inflated salaries to poach talent, no 996 crunch, no short-term KPIs, half of everyone's time for free exploration. All for one purpose: make people want to stay.
The recent major funding round? Liang sees its biggest value not as "more money" but as "senior employees got meaningful equity, team stability is secured."
My take: In AI, talent turnover is brutal. DeepSeek's strategy — vision + freedom + fair returns — is a more sophisticated talent play than "throw money at them."
6. Organizational Philosophy: No KPIs, Only Vision
> "We don't have an organization. We're vision-driven. A vision holds us together."
DeepSeek's management model is radically unconventional:
- No KPIs. "We don't operate on KPIs or evaluations. Only vision."
- Two parallel tracks. Bottom-up (free exploration, ~50% of time) and top-down (collective tasks, ~50%). Formal work never exceeds half.
- No overtime. "We're extremely focused, which means we have very few things to do. Because of restraint, I just don't do many things."
Liang admits this model has pros and cons and will need adjustment as headcount grows. But the core philosophy remains: research requires a relaxed environment.
My take: This isn't laissez-faire — it's the natural outcome of extreme focus. When a company crystal-clearly knows what NOT to do, overtime becomes unnecessary. Most companies' 996 isn't about too much work — it's about unclear direction.
7. China-US Gap: It's About Resources, Not Talent
> "The gap with America is essentially one thing — computing resources. Talent is not the bottleneck."
Liang provided specific numbers:
- DeepSeek currently has ~20,000 H100-equivalent GPUs (most arrived recently)
- US frontier models have ~800B active parameters. China is still at tens of billions
- To train an 800B model on Huawei 950 would require 200,000 cards
- Huawei's current allocation to DeepSeek: 16,000 cards (major tech firms get 100K+)
His key insight: talent gaps are actually a result of compute gaps. Less compute means fewer experiments, fewer opportunities for talent to grow. It's not that Chinese people aren't smart enough.
But the silver lining: "We achieve comparable results with 1/20th the compute." This efficiency advantage is itself a structural competitive advantage. As the compute gap narrows (domestic chips + production scale-up), this efficiency edge converts directly into market advantage.
8. Domestic Chip Replacement: A Historic Opportunity
> "NVIDIA's CUDA moat is rapidly crumbling. Within one year, the domestic chip ecosystem will prove to be fully viable."
This is one of the most forward-looking judgments in the interview. Liang sees three tailwinds converging:
1. AI can write code. Using AI to build the ecosystem is exponentially faster than humans writing CUDA kernels.
2. TileLang compiler. DeepSeek's in-house high-level language. Can rewrite the entire NVIDIA ecosystem with minimal code. V3's training already doesn't use NVIDIA's ecosystem.
3. Decoupling of compute and gaming GPUs. CUDA used to bridge both worlds. Now the compute market surpasses gaming — specialized chips no longer need CUDA compatibility.
His verdict is direct: the ecosystem problem will be proven solved within a year. The only bottleneck is production capacity.
Huawei's 950 super-node (4 cards ≈ 1 NVIDIA B200 in performance, <100% price premium) can already substitute — just 2 years behind in timeline.
9. Industry Endgame: Cost Is King, No Excessive Profits
> "Eventually, only three things differentiate models: cost, time, and user experience. Cost comes first."
Liang's endgame analysis is sobering:
- No excessive profits. Whoever aims for more profit will be beaten by whoever aims for less. This is the "vision decides victory" logic — if your vision is to take more, you've already lost.
- China could be "one of the three bodies." China has structural advantages in cost and product experience. Like every other manufacturing sector, Chinese companies will make AI products the most affordable.
- Three or four companies are enough. There are too many foundation model companies in China now. The field will converge. Three or four in full competition is plenty for a fierce price war.
My take: This is an "anti-rat-race" endgame. The winning strategy isn't being more intense than others — it's being more restrained, more focused, and lower cost. DeepSeek is choosing to run the longest, not the fastest.
10. Commercialization Philosophy: AGI Is the Watermelon, Everything Else Is Sesame Seeds
> "Last year our consumer traffic exploded, but we didn't chase users or monetization. Because there's a bigger watermelon ahead. Everything before it is just sesame seeds."
Liang's commercialization stance can be summarized as:
- It's too early for commercialization. "The monetization plans we discussed six months ago (ads, e-commerce, local services) are already useless — things change too fast."
- But commercialization is happening. Consumer users were acquired effortlessly. B2B API revenue alone could support a public company. "In the worst case, just selling APIs is enough."
- AGI is the only thing worth going all-in on. Once AGI is achieved, commercial value follows naturally. Until then, every ounce of energy should push the ceiling of intelligence higher.
My take: It's an "extreme focus + safety net" strategy. They're not anti-commercialization, but commercialization isn't the goal. If technology suddenly stalls, survival is guaranteed. But as long as progress continues, go all-in on technology. This "floor + ceiling" dual-track approach is profoundly pragmatic.
Final Thoughts
This 4-hour interview is, at its core, a founder's systematic articulation of his worldview.
Liang isn't selling "success secrets." He's describing a way of seeing the world: when the opportunity is big enough, goodwill and restraint are not moral choices — they are survival strategies. The more you try to monopolize, the easier you are to replace. The more you're willing to share, the more allies you find.
Whether DeepSeek ultimately achieves AGI or not, this way of thinking is worth learning: focus on what matters, show restraint on what doesn't, and trust the power of the long game.
Based on the complete transcript of Liang Wenfeng's investor meeting, published by "AI Coordinate".
Original link: https://mp.weixin.qq.com/s/825GYK5jiJgjvjHKK19IwQ