Outside the window, it is the typical wet and gloomy March weather of Shanghai; you could practically wring water out of the air.
(Taking a bite of a freshly bought sakura-flavored donut, sugar frosting dusting my lips)
Although today is March 4th—which, according to the calendar, might not be an earth-shattering date—in our slightly crowded tech circle, we’ve been absolutely stuffed with gossip these past few days.
You’ve likely seen the news: The AI core team of the “Eastern Tech Giant” in Hangzhou that we had such high hopes for—the Tongyi Qianwen (Qwen) group—has seen the resignation of four core technical heavyweights.
What makes this most intriguing is that one of them didn’t choose to leave to start a “Chinese version of OpenAI,” but instead bought a plane ticket, headed straight across the ocean, and joined Meta.
Now, this is interesting.
If they went to build a startup, we’d wish them “grand prospects”; if they went to a competitor, we’d joke about “good birds choosing the best trees.” But jumping from “China’s strongest open-source model team” to “the world’s strongest open-source model team” isn’t just a job hop—it’s practically a pilgrimage regarding “open-source faith.”
Today, let’s not talk about stock prices or gossip. Let’s talk about the “gravitational waves” behind these lines of code.
01. When “Hardcore” Hits the “Ceiling”
To be honest, I wasn’t too surprised when I saw the news.
Qwen’s performance over the past two years has been as fierce as a geek who just downed a triple espresso. On the HuggingFace leaderboards, it was at one point the only Eastern face capable of wrestling with Llama. If you peel back the layers of those so-called “Domestic Lights” wrapper applications you use, eight out of ten are running on fine-tuned versions of Qwen.
But that is exactly where the problem lies.
(Wiping breadcrumbs off the keyboard)
In the system of the “Hangzhou Tech Giant,” no matter how awesome the technology is, it ultimately has to answer one question: “What is the business model?” Or more bluntly: “How does this empower Double 11? How does it drive Alibaba Cloud’s DAU?”
These four departing colleagues were perhaps tired. They didn’t just have to train models; they probably had to accompany them with PPTs explaining why a certain parameter could lead to conversion rates. And over at Meta? Although Zuckerberg is a businessman, when it comes to open-source AI, he acts like a philanthropist spending money like water.
The big shot who chose to join Meta didn’t so much choose Silicon Valley as he chose “purer computational freedom.” There, he doesn’t need to explain to operations why the model needs more GPUs; he only needs to explain to LeCun why the architecture isn’t elegant enough.
02. Two “Dialects” of Open Source
There is an interesting blind spot here that many people haven’t noticed.
Although everyone is doing Open Source, Qwen and Llama actually speak two completely different “dialects.”
The flashing Logos in this image are not just code repositories, but the glory exchanged for the hair of engineers over countless late nights.
Eastern open source is, to some extent, an “Ecological Defense War.” Big tech companies open-source models to occupy a niche, to make developers dependent on their cloud services. This is a generosity with the temperament of an actuary.
Meta’s open source, on the other hand, is more like a “Scorched Earth Policy.” Since Google and OpenAI are running faster than me in closed source, I’ll destroy the roadbed—giving away the strongest models for free, turning your closed-source moats into a joke.
For top-tier engineers, which environment is sexier?
Even someone who loves West Lake Vinegar Fish would probably have their heart race when facing this kind of “I want to destroy the world” grand narrative. That heavy hitter who jumped ship was perhaps drawn away by this nearly insane “destructive innovation.”
03. The “Two-Way Rush” and “Dimensional Strike” of Top Masters
Let’s widen our view a bit.
We used to worry about “brain drain,” thinking that people leaving meant asset depreciation. But if you look at it from another angle, this is actually the “visa-free export” of Chinese technical talent.
The fact that these individuals could be poached by a company of Meta’s caliber proves one thing: the technical water level of the Qwen team is already on par with the world’s best, and perhaps even leading in certain tricks for processing Chinese context and long-text logic.
Looking at these flowing lines, you realize that so-called national borders are sometimes as thin as paper in the face of top-tier intelligence.
I’m wondering, will that colleague who went to Meta bring over that Eastern engineering capability of “doing fine embroidery within large parameters”?
You have to realize, while Llama is strong, it has been criticized for being “broad but coarse” in multi-language processing and some fine-grained alignment. If blood from Qwen is infused into it, the future Llama might understand “human sophistication” better.
This also serves as a reminder to other domestic giants: You can’t lock people in with just high salaries. Top minds need top challenges and a group of smart people who can understand each other’s jokes. If your computing cluster is full of reporting and wrangling, don’t blame them for embracing Zuckerberg’s “Metaverse.”
04. What If the Future Is a “Revolving Door”
I’ve also heard some worried voices: “It’s over, the core is gone, will Qwen rot halfway?”
Shh, don’t be so pessimistic.
(Turning the coffee cup in hand)
Technology has inertia; a mature team won’t collapse just because a few people leave. On the contrary, I prefer to view this as a “Graduation Ceremony” from the Whampoa Military Academy.
What if one day, this big shot who has been gilded at Meta, seen the world, and figured out the next-generation architecture, pulls a “carbine shot” and comes back to start a business? What they bring back then won’t just be a few lines of code, but the latest organizational paradigms and technical visions from Silicon Valley.
Isn’t Silicon Valley strong precisely because it is a “revolving door” for talent? Today at Google, tomorrow at OpenAI, the day after at Anthropic. This flow is the pump of innovation.
Perhaps we should learn to get used to this flow. As long as the “roots” are still here—massive application scenarios, massive data, and developers hungry for technology—no matter how far the birds fly, they will eventually land.
05. Final Thoughts
The rain seems to have stopped.
Looking at those familiar yet strange English names on the screen, I suddenly feel that code is a wondrous thing. It has no borders, yet it has character; it has no body temperature, yet it can perceive the warmth and cold of human hearts.
Four warriors left that familiar campus, changed their badges, and changed the flavor of their coffee.
For the “Hangzhou Tech Giant,” this might be a growing pain; but for the starry sea of AI as a whole, this is nothing more than a necessary thermal convection.
Wishing them a happy git push on their new battlefields.
After all, no matter which server it runs on, excellent logic always shines.
References：
