Tencent's Yao Shunyu and the AI "overtime theory": How can big tech respond to DeepSeek?
Tencent Chief AI Scientist Yao Shunyu was said to have stated internally that the team cannot avoid overtime the way DeepSeek does, but that through competition it can "force DeepSeek into overtime." The specific context of this claim has not been independently verified. The article notes by comparison that DeepSeek-R1's core training used only about 2,000 H800 GPUs at a cost of about $5.57 million, while reinforcement learning incremental training used 512 H800s over 80 hours at a cost of about $294,000. The controversy reflects big tech's anxiety over organizational efficiency and geek culture in the face of AI upstarts.
Body · AI translation
According to a report by LatePost, a member of the Hunyuan team recalled that Tencent's chief AI scientist Yao Shunyu once expressed something similar internally: the team could not avoid overtime the way DeepSeek does, but through competition it could "grind DeepSeek into working overtime." The specific context of this remark has not been independently verified. But later circulating versions further amplified its controversy.
Regardless of whether the wording was taken out of context, the violent shock it triggered in the public opinion arena precisely hit the deep-seated pain point of China's current large-model competition: when an internet giant accustomed to large-scale, coordinated campaigns faces a startup that has broken through with extreme engineering efficiency and geek culture, the first weapon it thinks of is still "grinding overtime"?
This is not merely a war over model benchmark scores, but an ultimate collision over organizational DNA, cultural conflict, and technological faith.
The Giant's Anxiety: Why Is Tencent Eager to "Defeat DeepSeek"?
Over the past year, although Tencent's Hunyuan large model has released multiple versions in an orderly manner and completed a large number of application integrations internally, in terms of external buzz and developer mindshare, it has consistently been suppressed by two extremes: on one side is ByteDance's Doubao, which is flush with cash and has used traffic and paid promotion to build hundreds of millions of users; on the other side is DeepSeek, which is sweeping the world with extremely low compute costs, hardcore technical strength, and an open-source ecosystem.
For a company like Tencent, accustomed to occupying absolute dominance in social and gaming, this "early start, late arrival" situation is extremely uncomfortable. Especially from the perspective of the capital market, if Tencent cannot prove its absolute leadership in the underlying infrastructure of AGI, then its vast ecosystem may be reduced to a "pipe" for others.
Therefore, poaching a top young scientist from OpenAI at a high price at the end of 2025, giving him the extremely high title of chief AI scientist, and letting him take charge of a core business line is itself a risky move by Tencent to break its internal rigid system and restart its geek engine. This 28-year-old young man carries not only the heavy responsibility of technology R&D, but also Tencent's KPI for defending its glory in the AI era.
The "Lazy" DeepSeek and the Overtime-Grinding Giant
However, the most jarring words in that internal speech were none other than saying DeepSeek "slacks off and doesn't work overtime," and that "overtime" should be used to grind DeepSeek down.
This is not only a misunderstanding of DeepSeek, but also exposes a certain deep-rooted path dependence of traditional internet giants when facing AI upstarts.
Is DeepSeek really "lazy"? The facts are exactly the opposite. If you understand how DeepSeek operates, you will find that not only are they not lazy, but they have invested an extremely terrifying density of intelligence in researching underlying architectures. According to public disclosures, the core training of DeepSeek-R1 used only about 2,000 H800 GPUs, with a training cost of about $5.57 million; subsequent incremental reinforcement learning training even used only 512 H800s for 80 hours, at a cost of about $294,000. This cost is only one-tenth or even less of comparable closed-source models. What they rely on is not piling up hours, but algorithmic innovation, systems engineering, and extreme code efficiency.
DeepSeek's open-source spirit, their bold breakthroughs in the MoE (Mixture of Experts) architecture, and their extreme optimization of inference frameworks prove one truth: in the AI era, what breaks through technical bottlenecks is often not "sweat," but "inspiration" and "extreme engineering capability."
By contrast, big companies are often filled with cumbersome reporting processes, lengthy alignment meetings, and "ineffective involution" for the sake of KPIs. If a team needs to "grind overtime" to compete with an opponent, it often shows that they have already lost the initiative in the correctness of the technical route and organizational efficiency.
"We can't be like DeepSeek and slack off without working overtime"—behind this sentence is actually a kind of extremely powerless tragic heroism. Its subtext is: we know others are more efficient, but apart from keeping employees up late, we seem to have no better way. But this is destined to be a war that cannot be won by physical stamina.
The Dilemma of the 28-Year-Old "Paratrooper": How to Break the Giant's Gravity?
The real challenge facing this 28-year-old genius scientist is not DeepSeek far away in Hangzhou, but the massive organization behind him.
Large-model R&D is an extremely complex systems engineering effort that requires efficient collaboration among teams of hundreds. For a young leader parachuted in, even with a top-tier academic vision and Silicon Valley experience, pushing a true technological revolution inside a giant like Tencent is as difficult as changing the engine on a high-speed train.
He needs to face:
Coordination of compute resources: how to coordinate extremely expensive and scarce GPU resources among various BGs (business groups)?
Resistance from veteran employees: big companies are often divided into factions, and will deeply senior algorithm veterans fully obey the radical instructions of a 28-year-old?
KPI pressure: in order to see results in a short time, will he be forced to do some face-saving projects such as "benchmark gaming" or "heavy packaging," thereby deviating from true underlying innovation?
Big companies have the top salaries and the largest resources, but also the heaviest inertia. If the "overtime theory" truly represents the inner thoughts of this young scientist, then he may already have fallen into the trap woven by big-company culture—using tactical diligence to cover up strategic anxiety.
Setting aside the feud between Tencent and DeepSeek, this controversy has also triggered deep reflection across the industry on AI R&D models.
At the frontier of algorithms, the proposal of a key architecture (such as Transformer, MoE, RLHF) often emerges only from a few exceptionally brilliant minds. One hundred mediocre researchers working overtime often cannot match the flash of insight from ten genius-level researchers in an extremely free, non-involution environment.
DeepSeek's success proves that the "geek utopia" can still defeat "capital-driven big companies" today. And Silicon Valley giants such as Anthropic and OpenAI also rely on extremely high "talent density," not simply "working-hour density."
If Tencent wants to defeat DeepSeek, it must not only surpass its opponent in benchmark scores, but also carry out a profound self-revolution in organizational form. It needs to build an efficient R&D culture that truly respects the geek spirit, allows trial and error, and even allows "no overtime"; it needs to let data speak, not let reporting documents speak.
Conclusion: A "War" Without Smoke
"Defeating DeepSeek" is not only a technical indicator, but has also become a haunting demon in the minds of China's big AI companies.
That 28-year-old young scientist carries extremely high expectations and also bears enormous controversy. His "internal speech," whether or not it was taken out of context, reflects the stress response of a giant in the face of a technological wave.
We are glad to see Chinese companies stand on the top podium in the world, and we also hope to see the awakening of the giants. But a true AI breakthrough is by no means tying employees to their desks, still less shouting a few harsh words at competitors. It requires a higher-dimensional technological foresight, a flatter and more efficient organizational structure, and awe for and reshaping of the underlying principle of "using compute to exchange for intelligence."
On the journey toward AGI, only extreme technological faith and pure innovation can laugh last. Grinding time will never grind out a real future.