文 | Tech商业
在当下的大模型修罗场里，各家公司不仅在拼算力、拼数据，更在拼一个极其隐秘却决定生死的关键要素：招人的品味。
日前，一位与Kimi（月之暗面）和DeepSeek（深度求索）核心团队有过深度交流的从业者，在社交媒体上曝光了两家公司在招聘标准上极其强烈的反差。这种反差，不仅折射出两家明星公司截然不同的组织文化与老板基因，更扯下了AI大模型竞争背后一条极其冷酷的底层逻辑——算力不仅决定了模型的上限，更直接决定了你不得不选择的人才模型。
招聘的折叠：奥赛原教旨主义 vs 斯巴达实战派
在推文的描述中，DeepSeek和Kimi的招聘画像几乎是在两个平行的宇宙。
DeepSeek是极端的“奥赛原教旨主义者”。 他们的简历筛选门槛极其硬核，几乎只看ACM、ICPC、超算比赛等顶级竞赛的奖牌得主。DeepSeek极度偏爱年轻的本科生，甚至不需要他们有多丰富的工程经验。面试时的核心考察点并非你掌握了多少前沿框架，而是你的好奇心、对技术的纯粹热情，以及思维是否足够敏捷（脑子够不够快）。
最颠覆行业认知的是，DeepSeek对“大厂经验”极其祛魅。推文中提到，即便你之前训练过超级大模型，在DeepSeek也不一定能加分。因为每个人的做事方式不同，旧时代的训练经验在全新的范式下不仅可能无效，甚至可能成为需要被重新洗脑的“包袱”。
DeepSeek愿意接受新人前两三个月不出成果。他们投资的是一块未经雕琢的璞玉：只要培养半年到一年，能否达到全栈正式员工的水平？他们买的是你未来的上限，而不是现在的即战力。
而Kimi（月之暗面）则更像是一个残酷的“斯巴达实战派”。 Kimi的招聘逻辑是：哪里有空缺，直接把重任交给你，默认你能扛住。他们不看重你是否有光鲜的竞赛奖牌，而是死磕你是否真正触碰过AI的最前沿。
考察的题目极其务实且刁钻：你是否了解模型在测试中是如何作弊的（比如直接从代码库里扒答案）？你是否亲手跑过大规模的实验？新模型一发布，你会不会立刻跟进？面试官会甩出极难的Coding题，看你的工程习惯和组织工具的能力。如果你自己写过提效工具，直接给高分。
Kimi没有耐心等你慢慢成长，他们需要的是来之能战、且真正在泥泞中打过现代AI战争的即战力。因为对于Kimi来说，前方的仗已经打到了最焦灼的时刻，容不得半点新手村的演练。
品味差异背后：是组织焦虑，更是算力鸿沟
为什么两家顶尖的中国大模型公司，会在招人上呈现出如此极端的割裂？
推文中给出了一个极其致命且真实的底层原因：算力资源的分配。
“除了老板的品味和团队氛围之外，造成这种差异的一个重要原因是：DeepSeek的算力比Kimi更充裕。如果没有计算资源，即使人来了，也没有算力让他们去真正干活。”
这句话，堪称整个大模型时代的“黑色幽默”。
算力充裕，让DeepSeek有底气去“养天才”。 拥有幻方量化深厚家底和早期疯狂囤卡战略的DeepSeek，拥有极其庞大的GPU集群。他们不需要新员工一进来就疯狂跑实验去抢产出。他们可以让一个奥赛金牌得主花半年时间去阅读论文、去思考底层的架构创新（比如他们在MoE混合专家模型上的极致优化），在试错中培养出极致的工程直觉。
在DeepSeek的逻辑里，大模型是一场长跑。那些在大厂里习惯了用几万张卡“大力出奇迹”的工程师，往往缺乏在受限资源下做极致优化的能力。而一群没有经验包袱、智商极高的年轻人，在不受短期KPI压迫的“乌托邦”环境里，反而能用最纯粹的数学直觉，爆发出颠覆性的架构创新。
而算力紧缺，逼迫Kimi必须寻找“超级个体”。 对于Kimi来说，每一张H100/B200显卡的算力时长都极其宝贵。如果招一个新人，还要花半年时间去培训他如何使用集群、如何排查底层的通信Bug，这在算力上是一种极大的隐性浪费。
所以，Kimi必须用极难的编程题和“是否扒过代码库”这种刁钻问题，筛选出那些已经在别处交过学费、能瞬间将极其有限的算力转化为模型提升的成熟工程师。Kimi买的不是你的潜力，而是你立刻能转化为算力效能的生产力。 这是资源受限下的极致生存策略。
老经验的陷阱：大厂光环在AI时代的破产
这两家公司的招聘偏好，还揭示了一个更深远的职场变化：在大模型时代，既往的“大厂光环”正在经历极速贬值。
DeepSeek不看重大模型训练经验，是因为大模型的底层架构演进得太快了。三年前训练千亿参数模型的方法，放在今天可能完全是一场灾难。很多在大厂习惯了用庞大算力硬砸出结果的工程师，一旦面对资源受限的场景，或者需要重新设计底层架构时，往往束手无策。DeepSeek更看重底层的数学直觉，因为他们相信自己能重新定义游戏规则。
而Kimi看重“动手能力”与“工具组织能力”，则是因为在AI落地的过程中，脏活累活实在太多了。数据清洗、评测排雷、推理加速、防范模型在基准测试中作弊……这些工作根本没有什么高大上的理论可讲，纯粹需要极高的智商去堆工程量。Kimi需要的是能够在泥泞中搭建起完美工具链的极客，而不是只会在PPT上画架构图的“大厂混子”。
在传统互联网时代，架构师的经验是护城河；但在大模型时代，旧经验往往是阻碍你拥抱新范式的绊脚石。谁能最快地忘掉过去，谁就能最快地适应这个以“月”为单位的迭代周期。
从人才战看终局：算力霸权与智力密度的双重较量
透过这两份截然不同的招聘清单，我们其实看到了中国大模型竞争的两种终极路径。
DeepSeek代表的是“算力霸权+顶级智力密度”的精英路线。用最充裕的算力，养最聪明、最纯粹的一批本土年轻人。他们不需要去卷什么To B的商业化落地，不需要每天写汇报文档，他们只需要在实验室里死磕模型结构，去触碰那个可能只有万分之一概率的技术突破。这是一种极具浪漫主义色彩的极客乌托邦。
Kimi代表的则是“极限资源+工程实战”的斯巴达路线。他们必须在有限的算力下，拼刺刀般地抢夺每一个Token的效能。他们的工程师必须是全能的，懂数据、懂评测、懂底层调优，甚至要防备模型在跑分时的作弊。这条路更苦、更累，但也更能锻炼出百战不殆的AI工程铁军。
这两条路没有绝对的对错，它们都是中国AI大模型在这场残酷的全球竞赛中，不可或缺的侧面。
结语：认清你的生态位
对于每一个想要登上这艘大模型牌桌的从业者而言，这段推文是一次极具价值的启示。
认清自己的定位吧：你是那个能在草稿纸上推导出新范式的竞赛冠军，还是那个能在匮乏的算力池里榨干最后一点性能的工具大师？
在AI时代，顶尖的人才从来不会被埋没，但前提是，你必须找到那个与你“八字相合”的组织。算力稀缺的公司会用极其苛刻的实战标准检验你的即战力；而算力充裕的公司，则会用极高的智商门槛和长期的耐心，去赌你身上那份未受污染的纯粹。
在通往通用人工智能（AGI）的漫长黑夜里，算力是柴火，而人才是火种。只有柴火足够充沛，火种才能在漫长的试错中，最终燎原。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Man, Tech Business
In the current LLM courtyard, companies are not only working hard, working hard on data, but also on an extremely secret key to life and death: taste.
Earlier, an practitioner who had an in-depth exchange with Kimi (the dark side of the month) and DeepSeek (the deep search) core team exposed in social media the extreme contrast between the recruitment criteria of the two companies. This contrast not only reflects the very different organizational culture of the two star companies and their owners' genes, but also draws a very cold bottom logic behind the AZLLM competition — calculation not only determines the ceiling of the model, but also directly determines the talent model that you have to choose.
Recipes of recruitment: Osai fundamentalism vs Sparta
In the tweet description, DeepSeek and Kimi were recruited in almost two parallel universes.
DeepSeek is an extreme “Osai fundamentalist”. Their CV screening thresholds are extremely hard, and they look almost exclusively at the winners of top-level competitions such as ACM, ICPC, super-calculations, etc. DeepSeek has a strong preference for young undergraduate students and does not even need their extensive engineering experience. The core point of the interview is not how much of the front-line framework you have, but your curiosity, your pure enthusiasm for technology, and whether your thinking is sufficiently agile (think fast enough).
The most destabilizing industry is the perception that DeepSeek is extremely perplexed by “big plant experience”. The tweet mentions that even if you've trained Super LLM, there's no point in DeepSeek. Because everyone does things differently, old-age training experience may not only be ineffective in a completely new paradigm, but may even become a “burden” to be brainwashed.
DeepSeek is willing to accept the new man's failure in the first two or three months. They're investing in an undecided piece of jade: can they reach the level of full-time employees with six months to one year? They bought the upper limit of your future, not the present.
Kimi, on the other hand, is more like a cruel “Sparta realist”. Kimi's recruitment logic is that where there are vacancies, you are given the responsibility directly, and you can take it. They don't care if you have a good game medal, but they're dying to see if you've really touched the front line of AI.
The subject of the study was extremely practical and impeccable: Do you understand how models cheat in the test (e.g. by taking answers directly from the code library)? Did you run a massive experiment with your own hands? A new model, publish, will you follow up immediately? The interviewer will throw out the very difficult Coding question of your engineering habits and organizational tools. If you've written your own utility, give it a high score.
Kimi did not have the patience to wait for you to grow up. What they needed was the combat power of the modern AI war, which they really fought in the mud. For Kimi, the battle ahead has reached its hottest point, and there is no room for beginner.
Behind the differences in taste: organizational anxiety, and the divide in numbers.
Why are the two leading Chinese companies, LLM, showing such extreme fragmentation?
The tweet gives an extremely lethal and real bottom-line reason: the allocation of power resources.
“In addition to the taste of the boss and the team atmosphere, an important reason for this difference is that DeepSeek is more calculating than Kimi. If there is no computing resources, even if people come, there is no ability to calculate what they really do.”
This is called "black humor" throughout the LLM era.
It's good enough to get DeepSeek to have "genius." DeepSeek, with the Fantasy quantization deep-seated base and the early frenzied card hoarding strategy, has an extremely large GPU cluster. They don't need new employees to come in and run crazy experiments to grab output. They can take six months for an Osei King winner to read the paper, to think about the underlying structural innovations (e.g., their extreme optimization on the MoE hybrid expert model), and to develop an extraordinary engineering instinct in a trial error.
In the logic of DeepSeek, LLM was a long run. Engineers used to use tens of thousands of cards for “magic miracles” in large factories often lack the ability to optimize under restricted resources. A group of young people with no experience and very high IQs, in a “Utopia” environment free from short-term KPI pressure, can instead use the most pure mathematical instincts to unleash subversive structural innovations.
And counting is too tight to force Kimi to look for “super individuals”. For Kimi, each H100/B200GPU is extremely valuable. If a new person is recruited, it will take six months to train him on how to use clusters and how to sort out the bottom lines of communication Bug, which is a great hidden waste in arithmetic.
Kimi must therefore use the very difficult programming question and the question of whether to "take the code library" to single out mature engineers who have already paid tuition elsewhere and who are able to quickly convert extremely limited computing power into model upgrading. Kimi bought not your potential, but your immediate ability to translate into productivity. This is an extremely survival strategy with limited resources.
The trap of old experience: the collapse of big factory halo in the AI era
The recruitment preferences of these companies also reveal a far-reaching change in the workplace: In the era of LLM, the former “big plant halo” was experiencing a sharp devaluation.
DeepSeek does not value the training experience of LLM because the bottom structure of LLM has evolved too quickly. The method of training hundreds of billions of parameter models three years ago could be a disaster today. Many engineers in large factories who are accustomed to crushing results with enormous computing power are often helpless when faced with limited resources or when re-engineered bottom structures are needed. DeepSeek values the mathematical instincts of the bottom because they believe they can redefine the rules of the game.
Kimi valued the ability to do hands and the ability to organize tools because during AI's landing there was too much dirty work. Data cleansing, assessment of demining, reasoning acceleration, protection against fraud in baseline testing... There is no high-level theory of such work, which simply requires a very high level of IQ to build up the works. What Kimi needs is a rival who can build a perfect tool chain in the mud, rather than a “mixer factory” that only paints a picture on PPT.
In the traditional Internet era, the experience of architects is the moat; in the LLM era, the old experience is often a stumbling block to your embrace of the new paradigm. Anyone who forgets the past as quickly as possible can adapt to this iterative cycle of months.
In terms of talent, it's the end of the game: a double race between the power of calculation and intellectual density.
Through these two very different recruitment lists, we actually see the two ultimate paths to China’s LLM competition.
DeepSeek represents the elite route of “calculative hegemonic + top intellectual density”. With the best of calculations, the smartest and purest group of indigenous young people. They don't have to roll over the commercialization of To B, they don't have to write report files every day, they just have to blow up model structures in the lab and touch a technological breakthrough that may be one in 10,000. This is a very romantic utopia.
Kimi represents the Sparta route of “extreme resources plus engineering in action”. They must take every Token's effectiveness with bayonets with limited calculus. Their engineers must be all-powerful, with data, assessments, bottom-of-the-horizon modulations, and even with the protection of the model against cheating. This road is harder and more exhausting, but it is also better able to exercise the battleless AI engineering army.
There is no absolute right or wrong in these two paths, which are indispensable aspects of China ' s AZLLM in this brutal global race.
Conclusion: Identify your ecological position
This tweet is an extremely valuable revelation for everyone who wants to be on this LLM table.
Know your location: Are you the champion of the new paradigm that can be derived from the draft paper, or the master of tools that can drain the last bit of performance in the scarce computing pool?
In the age of AI, the best talent is never buried, but only if you find the organization that fits your eight words. Countless companies will test your operational capability with extremely harsh operational standards; and well-finished companies will wager your unpolluted pure with very high IQ thresholds and long-term patience.
In the long night leading to Universal Artificial Intelligence (AGI), counting is firewood, while talent is a fire. Only when firewood is abundant enough will it be able to survive a long trial that will eventually lead to an end.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.