Martin's Blog
As an engineering guy and a casual tech enthusiast, I’ve paid close attention to the AI wave of the past couple of years. Honestly, getting to witness the rise of artificial intelligence before growing old is genuinely exciting.
The biggest tech-world news lately is that Hangzhou-based DeepSeek’s open-source large model nearly upended a whole field of domestic and foreign LLMs, leaving US tech companies in a panic. The latest evidence: last night (January 27), US tech stocks crashed — NVDA dropped nearly 20% at one point. Unprecedented, and frankly terrifying.
I don’t actually think DeepSeek directly caused it, but DeepSeek gave a timely push that amplified the volatility of an already unstable system. NVIDIA’s stock will bounce back.
This Chinese company, under heavy constraints on advanced hardware, forced its way through with algorithmic optimizations and a whole series of clever moves — and open-sourced it to the world. Long-term, this shows that “alchemy” for large models can be done under far more economical conditions. The future, then, must be the popularization of compute. Mainframes were once only found at universities and research institutes, but Bill Gates’s vision of “a computer on every desk” actually came true — followed by a sequel: “a phone with surplus performance in every pocket.” Yet Microsoft and Apple didn’t go under; they rode the new wave and redefined their eras. NVIDIA is the same. In the future, GPU compute won’t belong only to the data centers of top tech giants. As long as AI needs GPUs, hardware companies like NVIDIA will keep selling chips.
The bigger question: can DeepSeek keep pulling out killer moves?
Read any tech history and you’ll know: software-layer stuff spreads and copies easily, but hardware — unless wholesale, directly “transmitted” — requires enormous talent + time + trial and error to succeed. In the end, it always comes back to hardware.
The biggest takeaway from DeepSeek’s success is that mainland China’s STEM education is still incredibly strong. Plenty of people will trot out the “small-town test-taker” meme to mock it, but that’s pointless. DS has a sharp eye for talent and genuinely recruited a group of top minds, mostly from Tsinghua and Peking University. From CEO Liang Wenfeng’s earlier interviews, you can see this company is relaxed about talent, career, and profit, but the mission is crystal clear: build China’s foundational AI, and they’re confident they can pull it off.
Here’s what I’d add: the mainland’s talent pool is far, far larger than what DS’s snapshot reveals. But for reasons that aren’t convenient to elaborate, they don’t have a free environment in which to fully unleash themselves. Most lose their potential spark chasing power and profit. I’m not criticizing “chasing power and profit” — that’s fair enough — it’s just that soil that could have grown oranges can only grow trifoliate oranges. So after “Attention is all you need” landed, notice how the author lists on top AI papers from North American companies are full of names that look like Hanyu Pinyin — yes, those are the mainland’s “small-town test-takers.”
DeepSeek’s success once again shows that, under externally imposed constraints, Chinese people can push a “1-to-100” task to its limit.
DeepSeek’s success is admirable but hard to replicate — unless there are always companies like DS, and environments like this, that find and accept researchers and give them enough freedom to focus only on doing one thing well. I hope one day the mainland will have more “0-to-1” breakthroughs.
On the back of DeepSeek’s hype, I also hope that different mainlands and different ideologies will cooperate rather than decouple — East-meets-West isn’t a bad thing. Trump, that country-bumpkin businessman (yes, he’s a successful businessman-president, but that doesn’t stop him from being a first-rate bumpkin), is back in office and will stir up plenty over the next four years — he’s already announced tariffs on chips exported from Taiwan to the US…
I hope that one day, in arts or sciences alike, mainland talent can unleash far more stunning power — on the premise that they don’t have to “neijuan” (involute/over-compete) to do it. Fighting the West with 30% efficiency and 300% internal grind won’t work. Full integration with the world — that’s the sweetest outcome.
I’d meant to post this on New Year’s Eve, but my kid was super sleepy and I put him to bed. When I woke up, NVDA was rebounding strongly.
On the morning of the first day of the Lunar New Year, the cold wave has receded and Dali’s blue skies are back. May the new year bring peace, health, and fortune!
作为一个理工男，一个伪科技爱好者，对近两年的 AI 浪潮可是没少关心。说实话，有幸在尚未老去的时候见证人工智能的兴起，这是一件非常让人兴奋的事情。
最近科技圈最大的新闻就是杭州的DeepSeek 公司的开源大模型差点革了一众海内外大模型的命，搞的美帝科技公司人心惶惶，最新的例证就是昨晚（ 1 月 27 日）美股科技股暴跌，NVDA 一度跌去近20%，前所未有，见所未见的阔怕。
其实我觉得不是 deepseek 直接导致的，但 deepseek 正当时地推波助澜了一把，加剧了一个欠稳定系统的波动，英伟达的股价还是会涨回来的。
因为这家中国公司在先进硬件受限的条件下，通过优化算法等一系列操作，硬生生成功了，而且向全世界开源。这事情长期来看，说明大模型的炼丹可以在更经济的条件下进行，那么未来一定是算力的普及，就好比当年的计算机大型机只有高校和科研机构才能拥有，但后面比尔盖茨所描绘的"人人桌面都有一台计算机"当真实现了，而且还有续集"人人手里都有一台性能过剩的手机"。但微软和苹果公司并没有倒闭，反而在新的浪潮中革了时代的命。所以，英伟达也一样，未来的世界 GPU 的算力也不是只有头部的大科技公司的计算中心才能拥有，只要人工智能需要 GPU，英伟达之类硬件公司的芯片就一直可以卖下去。
反而是后面，DeepSeek 还能源源不断的拿出杀手锏来吗？
阅读科技史，可是知道，软件层面的东西，很容易传播和复制，但硬件的过程除非全盘得到直接"传授"，否则必须经过大量人才+时间+试错才可能成功。说白了，最后还是要回到硬件上。
DeepSeek 的成功，给人最大的感慨是，大陆的理工科教育还是非常牛叉的。虽然很多人又会拿出小镇做题家的梗来调侃，但没意思。DS 公司的人才储备非常有眼光，而且的的确确招到了一批清北为主的人中龙凤。从 CEO 梁文锋早几年的采访中就可以看出，这家公司在对待人才、事业和盈利方面，非常松弛，但目标非常明确，做中国的基础 AI，而且自信可以做出来。
到这里我想说，其实大陆的人才储备比 DS 公司展示出来的概况要多得多得多，但是，因为很多不便于细说的原因，他们并没有一个自由的环境充分释放而已，大部分都在逐权逐利的方向中失去了可能的火花。倒不是批判"逐权逐利"不好，这个无可厚非，只是土壤本可以长出橘，但无奈只能生出枳。所以，当Attention is all your need 横空出世之后，君不见，一众北美公司顶级AI 论文的作者署名里，总有一堆看起来像汉语拼音的名字，没错，那些，就是大陆的小镇做题家们。
DeepSeek 的成功，再次展示了中国人可以在外部受限的条件下，把一件 1-100 的事情做到极致。
DeepSeek 的成功难能可贵，但是难以复制。除非，一直有类似 DS这样的公司，这样的环境，发现和接纳并且给予研究人员充分的自由度，让他们脑子里想着做好一件事就够了。希望有一天，大陆也能有更多 0-1 的突破。
借由 DeepSeek 的爆火，也希望不同的大陆，不同的意识形态，宁合作，毋脱钩，中西合璧这不挺好的嘛。特朗普这个土鳖商人（嗯，他是一个成功的商人总统，但不妨碍也是一个登峰造极的土鳖）再次上台，未来的四年还会搞很多事情，比如已经宣布对 TW 出口到 US 的芯片增加关税……
希望有一天，大陆的无论文理人才，都能释放更多的惊艳力量，但前提是，他们不需要太"卷"就能做到，不然总是以 30%的效能和 300% 的内卷和西方干，是行不通的。当然，充分的融入世界，才是坠好滴！
上面写的东西本想除夕夜发出来，但崽子超级困，就带他睡觉了。起来一看，NVDA 又开始强势反弹了。
初一的早晨，寒潮远去，大理重新露出了蓝天。新的一年，安康，发财！
2025/1/29，大理