@@@SOTALAB · FRONTIER DATA@@@
An outsider's observation
In May 2026, Nathan Lambert, author of Interconnects, visited several leading AI labs in China—including Moonshot, Zhipu, Qwen, Ant Group, Meituan, and Xiaomi. After returning to the United States, he published a widely circulated set of observations.
He was impressed by the engineering culture: strong execution, relatively little ego among star researchers, and students deeply involved in core R&D.
But amid the praise, one line was almost never discussed in Chinese reposts:
=={blue}"The data industry is far less developed." China's data industry is nowhere near as mature as America's.==
That sentence matters to us, because what SoTALab is building sits exactly in that gap:
==Organize domain experts to produce the data frontier models need for training and evaluation.==
But before explaining what we do, it's worth answering a more basic question:
=={blue}Why does China's AI industry already have so many data teams, yet still look to an outsider like it has almost no real data industry?==
To understand how that gap formed, we need to go back to 2023.
How the distillation dynasty was built
After GPT-4 landed, Chinese labs faced a brutal timeline: investors wanted benchmarkable products within six months, while many teams were still standing up. Building a data system from scratch was barely an option.
So distillation became the most practical path: call APIs to generate SFT data, fine-tune on open-source instruction sets—much of which was itself produced by GPT-4 and other closed models.
We should admit it: the path worked, and it worked remarkably well. Benchmark scores rose fast, products shipped on time, fundraising narratives fell into place. Around this method, labs quickly built a mature engineering workflow.
Data teams' day-to-day work was often collecting, cleaning, and deduplicating content generated by other models—not organizing experts to produce native data. Refusal styles, phrasing habits, even prompt templates quietly carried the imprint of the teacher model.
There's no need to moralize. Distillation solved a real and urgent problem. For newly formed teams racing to catch up, it was very likely the right engineering choice at the time.
=={blue}It also delayed another investment: building the ability to produce one's own data.==
What was saved was time and cost. What was not built was expert organization, quality control, and native data production. In the instruction-tuning era, those capabilities rarely decided who won—so they were easy to deprioritize.
A few labs, though, used distillation and started building their own data infrastructure: internal annotation systems, native RL data pipelines, and verifiable reward signals. The teams that later found their footing on reasoning models were often these ones.
So the real dividing line was never "did you distill?"—almost everyone did. ==The dividing line was whether, beyond distillation, you left yourself a second engine and built real data capability.==
The moment distillation stopped being enough
In September 2024, o1 launched. In January 2025, R1 was open-sourced. Reasoning models made the limits of distillation-only strategies much harder to ignore.
Distillation mainly teaches a student to imitate a teacher's outputs. Along the way, the student's distribution stays a subset of the teacher's. It transfers answer style, knowledge expression, and some problem-solving patterns efficiently—but long-tail ability and exploratory diversity are often lost.
When the goal was still "does this answer look like GPT-4?", that wasn't fatal. Reasoning training follows a different path: RL with verifiable rewards. The model must generate solution trajectories again and again, then filter, correct, and reinforce them with checkable reward signals.
=={blue}Capability does not come only from imitating correct answers. It also comes from large-scale trial and error—and from continually eliminating wrong paths.==
That kind of training needs problems with clear, automatically verifiable answers; an expert-level difficulty distribution; and evaluation that can tell correct reasoning apart from a lucky final answer. You cannot fully scrape that from the public web. Ordinary crowdsourcing struggles. Synthetic data alone can't invent it either.
Without experts, it's hard to know whether a synthetic problem is actually valuable, whether its answer is reliable, or what capability it even tests.
=={blue}Data that is hard enough, accurate enough, and diagnostically useful can only come from one place: real domain experts—and the infrastructure that turns expert judgment into training signals.==
So the labs that fell behind after o1 and R1 did not fall behind on parameter count or compute. They fell behind because they lacked high-quality data for RL, and the organizational ability to produce it.
==That is when the debt came due.==
When does this exam arrive?
Calling for "more attention to data" is empty. Anyone can chant a slogan. The useful question is how to tell whether a lab has actually built data capability.
Over the next 18 months, gaps between models may concentrate more and more in four kinds of work:
- multi-step mathematical and scientific reasoning
- long-horizon agent tasks
- professional problems in medicine, law, finance, and similar domains
- open-ended tasks that require complex judgment
These are exactly where distillation is weakest and expert data is least replaceable. Labs that keep falling behind here will, more often than not, be revealing a failure in data infrastructure. That is a falsifiable prediction—we can check the answer in 18 months.
The good news: this exam is solvable. China does not lack experts. What it lacks is the industrial layer that organizes them.
Scale and its peers in the U.S. spent three years proving two things: that layer can be built, and it can command real market value. China has one of the world's largest populations of highly educated professionals, a deep engineering talent pool, and labs already hardened by the reasoning era.
==What has always been missing is only the bridge.==
Working with model teams, SoTALab keeps running into the same needs: the model already clears most public benchmarks, but the team doesn't know where the next harder problems should come from; the model gives plausible answers on professional tasks, but no one can pinpoint where it went wrong; general annotators can judge format and fluency, but not whether a legal argument, financial model, or mathematical proof actually holds.
All of that points to one problem: =={blue}what labs truly lack is not more data, but the ability to turn expert judgment into training signals.==
We are building that bridge
We should be transparent here: our reading of the problem comes with a position, because building this bridge is exactly what SoTALab is doing.
SoTALab starts from an asset few others have: an expert network accumulated over years inside the Zhihu community. It holds more than 200,000 domain experts verified by identity or professional credentials—licensed doctors, lawyers, investment banking analysts, mathematics and physics researchers, senior engineers. Over more than a decade, they left one of densest bodies of professional Q&A on the Chinese internet.
But an expert network is only a starting point—it is not data capability. A model cannot learn from a roster. It needs expert-constructed problems, full reasoning trajectories, verifiable answers, and scoring criteria that catch subtle errors.
The hard part is the conversion layer in the middle: how to bring experts into model R&D, how to turn professional judgment into executable rubrics, how to control quality differences across experts, and how to keep all of that running stably and continuously.
That conversion layer is what SoTALab is building. We turn knowledge and judgment scattered across expert minds and practice into assets models can train on, verify against, and keep improving from—and we turn an existing expert network into data infrastructure that can be priced, scheduled, and quality-controlled.
We want to become a source of training assets for frontier models, and bring real professional knowledge into frontier AI. That is where our mission comes from:
Bridging Frontier Human Expertise with Frontier AI
The work runs on two tracks: sending expert knowledge into training, and using expert judgment to expose model blind spots.
Training | Sending expert knowledge into models
On the training side, we focus on frontier mathematics, investment banking analysis (the APEX series), and terminal and code agents. Domain experts construct problems from the source, write full reasoning trajectories, and provide verifiable answers.
These are exactly the domains where public corpora are thinnest, general annotation struggles most, and distillation reaches least.
Evaluation | Using expert judgment to expose blind spots
On evaluation, we are not trying to build another leaderboard. We want model errors to become visible. Beyond a reference answer, each problem comes with an expert multidimensional rubric: where the model first drifted, why it drifted, whether the failure was missing knowledge, a calculation error, or choosing the wrong path from the start.
A score only tells us the model lost. A good evaluation tells us how it can get stronger next.
One simple test of data quality: do frontier models fail on it—and fail in a specific, diagnosable way?
Take an integration problem built by a math expert in our own dataset. Across five runs, frontier models averaged 1.6 / 10. The failure mode was clear: they hastily judged an elementary integral as "non-elementary," then built an entire seemingly rigorous derivation on that false premise.
The value of samples like this is diagnostic. They expose not a lack of knowledge, but a failure of judgment—exactly the signal RL training needs. Problems models cannot solve are the problems that make models stronger. Internally we put the standard in one line:
=={blue}It takes SoTA to train SOTA.==
We do not believe one company can fill the entire industrial-layer gap alone. But we are quite sure of this: the ceiling of the next stage of model capability depends on how much real expert intelligence can be converted into training signals. The United States has been building that layer for three years. In the Chinese-speaking world, the work has barely begun.
==The debt left by the distillation era will ultimately be repaid with expert knowledge.==
—And the unprecedented competition of the past few years has taught us this: ==whoever starts repaying first will lead.==
