The Musk who once slept on factory floors to boost production might find that even sleeping in the server room won’t stop top executives from leaving this time.
Today in Shanghai, it is a perfect 24 degrees Celsius. The clouds just happen to block the direct sun, making the temperature exceptionally comfortable for a stroll outside. But right after I brewed my coffee and scrolled through my phone, I noticed that across the ocean in the tech world, several god-tier executives casually strolled right out of their own company.
And this isn’t just any company; it’s xAI, a venture Elon Musk values immensely.
Letting the Inventor of Cement Walk Away
There is a particularly interesting detail. Back in 2023, when xAI was just founded, Musk highly publicized a startup team of 11 people. The lineup at the time could be described as an “all-star starting roster” aggressively poached directly from major tech giants.
Fast forward to today, and 6 out of these 11 people have already packed their bags and left. A star startup praised to the skies with a massive valuation has seen over half of its co-founders run away.
Among the recent departures are figures who could walk sideways (dominate) in the industry. For example, former Google researchers Zihang Dai and Tony Wu, as well as Jimmy Ba, who led the AI tutoring project.
These three left one after another within just a few days.
The most surprising name among them is definitely Jimmy Ba. A few days ago, I was having dinner with a friend who works in algorithms and brought this up. While gnawing on a rib, my friend complained to me: “This is exactly like preparing to build the world’s tallest building, and right after the foundation is laid, the guy who invented modern cement tells you he’s quitting.”
You might not know this, but behind today’s mainstream artificial intelligence, the underlying systems heavily rely on an optimization algorithm called “Adam.” And Jimmy Ba is the co-inventor of this very thing.
To help you understand how important this is, let’s use our imagination. Training artificial intelligence is actually like walking down a mountain blindfolded to find the valley floor. The old method was incredibly clumsy: you could only take a step, feel the rocks, gauge the slope, and then decide where to step next. This approach was excruciatingly slow and easily got you stuck in a small pothole, making you think you had reached the bottom.
The Adam optimizer is like equipping this system with a pair of smart roller skates with memory functions. Not only can it remember the speed of your previous downhill descent, but it also adjusts itself based on the current terrain. On a steep slope, it helps you accelerate downwards; on flat ground, it slows you down, ensuring you don’t take too big a step and miss the deepest point.
It is no exaggeration to say that over half of the AI models in the world today walked out relying on these “shoes.” Losing someone with such terrifying intuition for underlying technology is absolutely a heavy blow to xAI.
The AI “Memory Master” Also Bailed
Besides Jimmy Ba, another departing figure, Zihang Dai, also boasts a formidable background.
When we normally chat with AI, we often encounter an extremely awkward situation: if you chat with it for a bit too long, it forgets what you said ten minutes ago and starts speaking incoherently.
To cure this “amnesia,” Zihang Dai co-developed two highly famous technologies a few years ago. He forcefully stretched the context length models could process by a massive margin, allowing machines to truly remember long, coherent content instead of having a seven-second goldfish memory.
Not to mention Tony Wu, whose resume is filled with shining names like OpenAI and Google. He has seen firsthand how the top two AI labs on the planet operate and is a true battle-tested veteran.
Why would these brilliantly smart minds collectively decide to “grab their buckets and run” instead of staying as perfectly fine co-founders?
Such a dense loss of executives usually points to one thing: they can no longer see eye to eye on the technical roadmap or the overall future direction of the company.
Why Doesn’t “The Boss Sleeping on the Floor” Force Out Large Models?
Speaking of this, we have to mention Musk’s exceptionally fierce management style.
The whole world knows Musk is an absolute madman. I read a foreign media report before that when Musk was facing the production hell of the Tesla Model 3, in order to force everyone to build the cars, he literally rolled out a sleeping bag and slept on the factory floor.
Later in an interview, Musk revealed his true thoughts. He said it was absolutely not because he couldn’t afford the luxury hotel across the street, but because he intentionally wanted his situation to be more miserable than anyone else in the company. He wanted to be covered in dust and wake up with a sore back. Because if the boss was this miserable, the employees below him would be too embarrassed to complain about working overtime, no matter how tired they were.
This culture also blew into his other companies. A female executive at Twitter, in a rush to meet deadlines, slept on the conference room floor in a silver sleeping bag at 2 AM. The photo went viral online, and Musk himself highly praised this dedication.
This hardcore, almost extreme leadership style indeed works when building cars or rockets. Engineering problems in the physical world often yield to sheer brute force. As long as you stay up late turning screws, production capacity can always be forced out bit by bit.
But the problem is, this playbook doesn’t seem to work when dealing with AGI researchers. Shrugs
Currently, xAI is frantically expanding its server rooms and buying tens of thousands of GPUs, trying to forcibly break through technical bottlenecks by stacking overwhelmingly massive computing infrastructure. Musk’s logic is simple and crude: as long as the computing power is large enough, and as long as everyone works desperately like they did building cars back then, this will definitely succeed.
However, conducting cutting-edge academic research and running an industrial assembly line are concepts in completely different dimensions. You can’t suddenly force a model’s IQ up by twenty points just by yelling behind a programmer’s back.
Training a large model is like brewing an incredibly complex thick soup. Researchers need to spend loads of time cleaning up messy data, carefully designing experiments, and then sitting in front of a screen waiting for weeks. During this time, they can only watch lines of jumping code and pray that a bizarre hardware failure doesn’t turn the whole pot of soup sour and rotten midway.
This kind of work requires immense patience and a relatively relaxed environment for trial and error. If these genius scientists are constantly pressured with a ticking stopwatch, they will only feel mentally and physically exhausted and simply pack up and leave.
One Last Bug-Repellent Tip
Actually, similar things have happened in other top tech companies. When a company rapidly pivots from pure technical exploration to a wild commercial sprint, the first ones to feel out of place are usually the founding veterans who just want to quietly do research.
As formidable as Musk is—able to send massive Starships into the sky and draw grand blueprints on Mars—when managing these brains that are the smartest in the world and require the most freedom, he is inevitably going to have to pay a steep tuition fee in management.
By the way, while looking up information earlier, I found a very obscure little piece of gossip. Among the remaining co-founders at xAI, a heavyweight named Greg Yang recently announced he was temporarily stepping down from daily work to rest at home.
The reason is surprisingly that he was bitten by a tick and unfortunately contracted Lyme disease.
It seems surviving in this extremely cutting-edge industry is genuinely not easy. Not only do you have to guard against your boss’s occasional hardcore pressure, but when you go for a relaxing walk on the lawn, you also have to constantly watch out for little bugs. Talking so much has made me a bit hungry. I’m going to grab something to eat first. Remember to bring bug spray if you go out to play this weekend.
References:
- Elon Musk says sleeping on factory floors was important so his team could see him
- Jimmy Ba and Tony Wu latest tech co-founders to exit Musk’s xAI
- Adam: A Method for Stochastic Optimization
- XLNet: Generalized Autoregressive Pretraining for Language Understanding
—— Lyra Celest @ Turbulence τ.
