Organizational Strength Behind Kimi K3: Management Ledger of Two Generations of Chinese AI

I noticed an interesting detail: Ye Qiyi mentioned in the conversation that while Silicon Valley was still reviewing the shockwaves of the "DeepSeek Moment," K3 was already approaching or even surpassing closed-source SOTA on certain benchmarks. This isn't just a technical "Kimi Moment"; it's a watershed moment for organizational capability.

As a manager who has led AI platform teams for years, I tend to break down model performance into two variables: technical ceiling and engineering efficiency. The brilliance of K3 looks like an algorithmic breakthrough on the surface, but at a deeper level, it reflects the organizational dividend brought by two generations of talent migration in Chinese AI. Starting from the SenseTime cohort in 2015 to today's startups like Moonshot AI and DeepSeek, this decade of talent flow is essentially completing a "generational shift in technical management paradigms."

The first generation of AI companies, like us back at SenseTime, faced a core contradiction: "scale vs. efficiency." Our hiring strategy was headcount stacking—half of the 100 engineers were algorithm researchers, and the other half handled infrastructure and business. But the problem was severe lack of management bandwidth. Technical VPs had to simultaneously oversee model training, data platforms, online services, and customer customization requests. The team structure was "matrix + project-based," with each project having a tech lead, but resource allocation relied entirely on pulling strings, and ROI could only be calculated after post-mortems.

We often argued back then: Should a high-potential researcher spend six months on model distillation, or fix bugs in the data pipeline? No one could give a definitive answer because no one could calculate whether "a 10% accuracy improvement from distillation" or "99.9% stability in the data pipeline" was more critical to the business.

The second generation of AI companies, like Moonshot AI and DeepSeek, have a completely different management logic. From the start, they realized that belief in AGI isn't just a slogan but a fundamental constraint in organizational design. Because they believe in AGI, the team must be extremely flat to reduce decision-making friction; because they believe in AGI, resources must be concentrated on frontier breakthroughs rather than churning out mediocre implementation projects.

Ye Qiyi mentioned Yang Zhilin's process of "searching," which is actually about finding an organizational form where every engineer can directly feel the connection between their contribution and the AGI goal. This sense of connection is a leverage point for management efficiency. If you can make a training engineer feel like they are "driving the agent paradigm," their motivation and self-drive to work overtime tuning parameters will far exceed doing feature engineering for some random project.

I've seen a comparison: A large company's AI team, with 50 people maintaining a model service, dealing with daily pipeline exceptions and optimizing inference latency, with quarterly OKRs being "launch three new features." Meanwhile, K3's team, also 50 people, has an organizational structure integrating "algorithm + engineering + infrastructure," where everyone is responsible for the final model effect. In this structure, an engineer proactively researches compute scheduling strategies because they know avoiding one OOM error means running one more experiment cycle.

This is backed by ten years of accumulated talent migration. The first generation of AI trained many technical backbone staff, but these people only truly shed the "institutionalized" baggage when they joined second-generation companies. They no longer need to waste energy on process reviews and cross-departmental alignment, instead using all their time on the cutting edge. This is why K3 could iterate rapidly after the "DeepSeek Moment"—they had no historical burden, and the organization itself acts as a "high-performance model."

However, this organizational model comes at a cost. When team size exceeds 100 people, relying purely on faith and flatness won't hold up. I did the math: In a 100-person AGI team, if each member spends an average of 20% of their day "aligning information," that's 20 person-days wasted. Second-generation companies can currently maintain this low loss because they only hire "the most suitable people"—not just for technical ability, but for their identification with the AGI belief.

Ye Qiyi said the difference between China's two generations of AI lies in their understanding of "engineeringization." The first generation viewed engineeringization as "implementing research results," while the second views it as "accelerating research itself." This is precisely the core difference in organizational design. The former requires strict project management, milestones, and risk control; the latter requires flexible experimentation platforms, rapid trial-and-error, and automated compute scheduling.

I noticed that after K3, Moonshot AI didn't rush to expand the team but continued "searching" for a better organizational architecture. This is actually a form of restraint. Many companies blindly expand after a "moment," diluting engineer culture and causing a cliff-like drop in organizational efficiency. From a management perspective, keeping a small, elite team until you're confident the next stage's organizational model can support larger scale is the more rational choice.

Because once the thing called "AGI faith" gets diluted, it's very hard to recover.

Original link: https://www.tmtpost.com/8083639.html