Synthetic Data's Hidden Ceiling: Why $400 Per Sample Signals the End of Easy AI Scaling

Real Data Peaks: AI Begins Feeding Itself

The AI community acknowledges an open secret: high-quality public human data — Wikipedia, open-source code, academic papers, Reddit, Zhihu, Tieba — has been largely exhausted. Research institutions predict the stock of high-quality natural language text could be fully depleted within years. To avoid the Scaling Law hitting a "data wall," the industry has turned to synthetic data .

From rule-based generation and self-play to model distillation, chain-of-thought (CoT) generation, and automated annotation, AI now enters a "self-generated data" phase. OpenAI's o1 series and Anthropic's Claude rely heavily on engineered, automated synthetic reasoning trajectories as a core moat.

A Single Sample Costs Thousands: Synthetic Data Becomes a Luxury Good

Contrary to the belief that synthetic data is near-zero cost (just prompting an API), the article details why hardcore synthetic data that triggers intelligence leaps is extraordinarily expensive . In domains like mathematical reasoning, frontier science, formal verification, and complex long-horizon logic, producing one "truly usable, fully correct, long-horizon reasoning" sample involves:

Multiple Monte Carlo Tree Search (MCTS) and Reinforcement Learning (RL) path sampling : Exploring optimal or counter-intuitive solution paths may consume tens of thousands of large-model token generations per attempt.

Rigorous formal verification and multi-agent review systems : Code execution environments, compilers (Lean 4, Isabelle), or cross-review by multiple high-level models must filter out hallucinations and logical traps. The article states the valid conversion rate is alarmingly low .

Tiny proportion of elite human expert verification : Top-tier reasoning chains still require algorithm experts and mathematicians for spot-checks and anchoring.

Compute cost + API consumption + verification rejection rate + expert involvement layer up so that one cutting-edge synthetic sample for top-tier reasoning model reinforcement can cost thousands of RMB . This is no longer a game for ordinary startups but an arms race among giants comparing capital depth and engineering foundations.

Why Synthetic Data Is Also Hitting a Ceiling

More alarming than cost is the emerging ceiling. Developers observe clear diminishing marginal returns and three dilemmas:

1. Model Collapse: The Autophagic Curse

Academia has long warned: training successive generations on model-generated data is like photocopying an image 100 times or inbreeding. The model loses diversity, converges to bland statistical averages, and eventually degrades or collapses. AI can distill the essence of human knowledge but struggles to create genuine physical laws and serendipitous miracles from nothing.

2. Ineffective Involution Within "Known Boundaries"

In math and code where ground truth exists, compilers can verify synthetic data. But in open domains — real-world business, social games, humanistic creativity — "what is correct" has no formal standard. Model outputs merely circle within its existing prior knowledge. You cannot weave a new continent with an existing logic net.

3. Compute Consumption and Return Curve Diverge

When compute per valid synthetic sample grows exponentially while benchmark gains shrink from 5% to 0.2%, the commercial logic falters. Expensive synthetic data replays the early autonomous-driving annotation trap: "billions invested, last 1% nearly impossible."

Where Is the Breakthrough in the Second Half?

The narrowing marginal effect of synthetic data does not signal AI's end but forces a shift from "brute-force fitting" to "mechanism reconstruction":

From "Talking to Itself" to "Physical Interaction (Embodied Intelligence)" : Pure text synthesis nears its limit. Real-world physical interaction data — robot force feedback, sensor streams, environmental collisions — becomes the most irreplaceable asset.

Self-Play in Formalized Environments : Like AlphaZero, place AI in a rule-consistent sandbox (Lean math proof environment, code sandbox, financial simulation) where agents discover paths humans never took through trial and error, rather than memorizing expensive generated QA pairs.

Small-Entry, High-Density Vertical Private Loops : When general synthetic data becomes costly and capped, deep vertical loops in industrial, medical, legal scenarios with "real verification feedback chains" form the deepest moats.

The pattern of technology development holds: no single-point technical dividend extends infinitely. Two years ago pre-training corpus was hyped as endless; last year synthetic data was bet on as a silver bullet. Today, a multi-thousand-yuan bill and flattening performance curves drag the industry back to reality: Large models never lack data; they lack reverence for the real world and cognitive frameworks that break mental boundaries. The ebb of synthetic data is not winter arriving, but the signal that hype has faded and hardcore engineering and scientific exploration truly begin.