The Long Way Around
Every industry has a favorite metaphor, and in artificial intelligence this year it was the flywheel. Models help build their successors, and products earn the revenue that buys the compute for the next model. Spend enough time in the field and you start to see flywheels everywhere, along with a steady supply of shortcuts that promise to spin them up faster.
The idea that stayed with me most this year was older and plainer. It is a Chinese word, 本分 (běnfèn), and it doesn't translate neatly. Roughly, it means do what is yours to do, and take only what is yours to take. It carries a hint of duty and a hint of modesty, and none of the glamour of a flywheel.
I learned it working close to Zhang Yiming, ByteDance's founder. I had expected his vision to be what I remembered. What I remember instead is how little distance there is between what he says and how he operates. He keeps his ego small, goes straight to the essence of a problem and thinks on a long horizon. None of that is unusual advice. What is unusual is seeing it practiced with no exceptions for when it is inconvenient.
The clearest test is distillation, which means training your model on the outputs of someone else's. Most leading labs forbid it in their terms of use. Doing it anyway is, at a minimum, not honest. The larger cost is strategic. Distillation can close a gap quickly, and that is precisely the problem. It lets a lab skip the things that are hard to build and impossible to buy, chief among them research judgment and an organization capable of producing it. The bill tends to arrive when the model you copied is no longer the frontier. Yiming has a word for the habit: 取巧, roughly "the clever shortcut." We don't take it.
Refusing the shortcut leaves the long way round. At Seed, ByteDance's AI research unit, the long way runs through a short list of things that rarely make a keynote slide. We watch the flywheels as closely as anyone, yet our discussions keep returning to the inputs. We need more compute. We need a culture where good research can happen. We have to balance the rhythm of model releases against bets that may take years to pay off. And there is data, which means both the infrastructure to handle it and the patient craft of curating it. None of this is novel. All of it is hard, and most of it is invisible from outside, which is partly why it is so easy to neglect.
Fundamentals also require focus, and focus is easiest to see in its absence. This year Google invested in outside model companies more than once and sold its own compute to others. There is a defensible logic to both, and a company of Google's size can afford to keep its options open. But options are not free. Every hedge quietly tells an organization that its leaders are not sure their own bet will win, and organizations listen. If building great models is the work, the 本分 answer is to stop hedging and put everything behind it.
The last ingredient is candor, and it starts at the top. Yiming talks about other companies without condescension. He studies those ahead, alongside and behind with the same attention, less to rank them than to find what applies to us. Confucius had a line for this: when you meet someone better, try to match them; when you meet someone worse, look for the same fault in yourself. Turned inward, it means saying plainly where you are behind. China's data industry is behind America's, and pretending otherwise would be comfortable. Admitting it is what lets you decide to build the techniques in-house rather than wait for a market to catch up.
The same honesty applies to evaluation. Benchmarks are easy to flatter and easy to game, and a lab that games its own evals is, in effect, lying to itself. An honest eval is one designed to deliver bad news early. It is less satisfying than a leaderboard, and far more useful.
I first learned a version of this in school, and not gracefully. For years I optimized for what would look impressive rather than what would make me good, and ended up with neither. Things turned only when I stopped performing and did the unglamorous work. At the time I took it as a lesson about studying. It took watching the same principle run through an entire company to see that it never was.
None of this guarantees winning. A rival can outspend you on compute or ship first, and the flywheels may spin faster elsewhere. 本分 makes no promises about outcomes. What it offers is narrower and more durable: whatever you achieve is actually yours. Results that run ahead of capability, whether they come from a borrowed model or a flattering benchmark, tend to be handed back. Results built on fundamentals tend to stay. It is also, I have found, a calmer way to work.