If you want one question that captures today’s most hidden structural shift in AI, ask: what does the most valuable training data look like now?
Two years ago the answer was “high-quality human annotation.” A year ago, “expert chain-of-thought.” Today the answer is turning into — enterprises’ real production workflows.
This is not an incremental shift. It touches the foundations of the training paradigm and will decide which companies rise and which get washed out over the next three years.
Data Is a Surfing Business
The industry has an easily missed regularity: the winner of the data market keeps changing. Every time a model capability bottleneck breaks, the most valuable data changes form. The record so far: Scale AI caught autonomous driving and early-LLM annotation; Mercor and Surge AI rode the expert-data wave; and now, as training pivots to long-horizon tasks, the compass points at real-world workflow data.
Reasoning and long-horizon are the mainlines of model progress. The unit of work models handle is climbing — from a single code change to a file, to a codebase, to a complete project. The longer the task, the harder it is to simulate artificially, because real projects teem with historical state, tool dependencies, organizational rules, and unexpected branches — details no expert can invent sitting in a sandbox. So the frontier labs have started buying real-world data.
Two Textures of Data
Type 1: data captured from real work. Close to a recording of the work process — session replays, screen actions, internal collaboration records. The best Type 1 data preserves not just actions but the intent behind them. GitHub is the canonical example: commit messages, issues, and ticket resolutions chained together almost completely preserve a problem’s journey from appearance to resolution — what someone wanted to fix, what they tried, why, and whether it was accepted. A natural Type 1 treasure house.
Type 2: artificially constructed data. What most human-data companies do today: recruit domain experts → design tasks by hand → have experts complete them in a sandbox → collect results → process into trainable formats. Type 2 suits pretraining and early capability ramping, but its ceiling is obvious — simulated environments never fully reproduce real-world complexity. When a model needs to grasp an internal approval flow, a cross-team collaboration rhythm, or an industry’s unwritten rules, Type 1 is irreplaceable.
The most viable real-world route is Type 1.5: make Type 2 behave more like Type 1 — long-term binding of a small set of high-quality experts who understand the training logic, redesigned incentives that prevent reward hacking, multi-layer QA systems progressively engineered.
Three Rulers for Data Quality
Whether a data vendor earns model labs’ trust rests on three capabilities. 1. Data taste — genuinely knowing what good data looks like: telling “data that looks right” apart from “data the model actually needs.” 2. Research taste — knowing where models are stuck now and what kind of data addresses it; especially scarce, because most data vendors don’t understand training and most researchers don’t understand data production. 3. Scalability — maintaining quality as volume grows; plenty of companies staff their best experts at demo stage with founders watching and heavy manual QA, and look beautiful at small scale. That proves nothing about 10× or 100× replication.
To judge whether a data company is “selling labor” or “building a factory,” read the hiring structure: frequent SPL (special project lead) hiring means taking orders, expanding teams, custom delivery — selling labor. Hiring only members of technical staff means the emphasis is data engineering, environment engineering, and QA automation — building a factory. If data value keeps migrating toward Type 1, the window for a pure Type 2 model may be two years.
Hillclimbability is the metric that matters
Many teams design training tasks on the logic “if I build something the model can’t do, that proves value.” But designing a task a frontier model fails is easy. The hard part is placing the task exactly on the model’s capability boundary. Picture two scenarios:
-
Worth training: Pass@1 = 0%, Pass@32 = 30%. The model fails outright once but succeeds 30% of the time within 32 attempts — it occasionally touches the right path and just needs stabilization.
-
Not worth training: Pass@1 = 0%, Pass@256 = 0%. Even 256 attempts never succeed — the task is beyond the model or too obscure to matter.
This frame — hillclimbability — measures training value far better than “can the model do it.”
Four Directions RL Environments Are Evolving
As models shift from answering questions to completing tasks, training necessarily moves from supervised fine-tuning toward reinforcement learning — and RL environments are evolving hard in four directions. 1. Space expands: an agent no longer completes everything in one sandbox but moves across many — a sales-ops agent may need the CRM, email, the customer database, and the approval system simultaneously. 2. Tools go live: agents will call tools with uncertain outputs — the environment’s “search tool” may itself be another SOTA model, and the agent must learn parameter control, result judgment, and cross-validation. 3. Organizations grow complex: hard tasks get decomposed to sub-agents and re-aggregated, and reward must weigh cost and latency alongside completion — finishing in 40 minutes versus 3 carries entirely different business value. 4. Task boundaries move up: as models strengthen, twenty-step tasks collapse into one, and RL environment goals climb — from executing an action, to completing a sub-process, to planning and orchestrating whole workflows.
Why coding is a natural RL fit
Jason Wei proposed Verifier’s Law: the ease of training AI on a task is proportional to how verifiable it is. Coding suits RL not merely because the reward signal is clean, but because the whole environment copies, parallelizes, and resets easily — one repository clones into ten thousand containers, and the model edits, tests, fails, and retries at will.
The real world offers no such save points. Entrepreneurship, management, investing, law, sales — all have outcomes, but none will spawn ten thousand parallel worlds for practice. That is why AI entering the real world cannot rely forever on brute-force drilling; it must raise sample efficiency — extracting transferable patterns from small amounts of unrepeatable, ambiguous-feedback experience. This remains one of humans’ clearest advantages over models.
If You Are a Founder or a CTO
1. Should your enterprise post-train its own model? Post-training trades Capex for Opex — a one-time training cost replacing perpetual inference premiums. The math works only if four multipliers multiply out: data uniqueness × eval clarity × task frequency × value per improvement. Any factor near zero kills the investment.
2. Are you a data company? If the Type 1 migration holds, pure-Type-2 players without a Type 1 roadmap have under two years. If you are a VC, do not fund a data company with no Type 1 plan.
3. Are you an AI user? When real workflows become training data, the way you work daily is becoming model capability. Whoever deposits more structured, high-quality operation traces — code, documents, decision records, retrospectives — is quietly building their own data moat.
The data market’s logic is turning from “find more people to annotate” to “extract from real work.” The driver is not technological idealism but economic rationality: when training moves to long-horizon tasks, artificially built environments get exponentially expensive, while real-world data is already there — it only lacks an effective collection and conversion pipeline. That pipeline is the most interesting AI infrastructure opportunity of the next two to three years.
