The Data Black Hole: Understanding the Sample Efficiency Gap in AI

AI Progress is Driven by Data Volume, Not Sample Efficiency

Modern AI progress is primarily the result of widening and improving data distributions and scaling compute, rather than fundamental improvements in how efficiently models learn from data. Intelligence can be defined as "sample efficiency"—the amount of data required to operate fluently in a given domain. While AI capabilities have expanded, the underlying efficiency of the learning process has not significantly improved.

Reinforcement Learning (RL) serves as a mechanism for synthetic data generation. By applying compute against a verifier or a rubric (often an LLM acting as a judge), models identify high-quality data and are trained to predict correct rollouts. However, this process requires a baseline prior probability of the model anticipating the correct solution, necessitating massive amounts of bespoke human expert data across every target skill.

The Role of Human Expert Data

To achieve competence in specific fields, AI labs employ hundreds of experts to generate completions, write rubrics, and explain chains of thought. This has created a multi-billion dollar data industry specializing in highly specific tasks, such as:

• Converting legacy documents into polished Word files.
• Writing realistic M&A diligence reports or securities filings.
• Creating template market research.

The Sample Efficiency Gap: Humans vs. AI

There is a massive discrepancy between the amount of data a human requires to learn a skill and the amount required by a frontier AI model. This gap is characterized as a "black hole of data" that supports the visible capabilities of the AI.

Quantitative Comparisons

• Language Acquisition: A human adult may have encountered roughly 200 million tokens by the time they reach adulthood (assuming 2,000 words per hour). In contrast, frontier models are trained on tens to hundreds of trillions of tokens—a millionfold difference.
• Robotics: Humans can learn to teleoperate a robot arm within hours. AI models require millions of hours of demonstrations and still struggle with complex, open-ended tasks.
• Driving: A teenager can learn to drive in approximately 20 hours of practice. Self-driving models from companies like Waymo and Tesla use data that is three to four orders of magnitude greater than what a human uses.

Addressing Common Counterarguments

• Evolutionary Pretraining: Some argue that billions of years of evolution "pretrained" humans. However, the human genome is only three gigabytes, with only 1-2% protein-coding, which is insufficient to store the parameters of a pretrained network. Evolution likely optimized hyperparameters and loss functions, but the connectome (the weights and parameters) is still built from scratch during a lifetime.
• Multimodal Data: The argument that humans ingest more data via sight and sound is countered by the fact that blind or deaf individuals still possess general intelligence, suggesting that massive sensory token streams are not the primary driver of human intelligence.
• Model Scaling: Scaling laws suggest that larger models are more sample efficient, but the effect is marginal. According to Chinchilla scaling laws, even increasing parameters to infinity would only reduce the required data by a factor of ten to maintain the same loss, which does not bridge the millionfold gap.

Implications for Automation and AI Research

Despite the lack of sample efficiency, AI remains economically viable for automating white-collar work because the cost of "firehosing" massive amounts of data into a model can be amortized across billions of user sessions.

White-Collar Automation

For common tasks performed by software engineers, accountants, or analysts, the data is readily available to be brought into the training distribution. While AI is less efficient than humans at learning these tasks, the ability to scale the output across millions of instances makes the inefficiency irrelevant to the bottom line.

The Limit of Distribution-Based Learning

Some roles require "out-of-distribution" thinking—dealing with problems that are distant from any existing training data. Software engineering is cited as a primary example of a job requiring this capability. Because of this, there may be a higher demand for human software engineers in 2028 than there is today, as AI acts as a complementary tool rather than a total replacement.

The Path to Human-Like Intelligence

AI labs aim to automate AI research first, with the goal of having automated AI researchers solve the sample-efficiency problem. This would allow models to move beyond simply being a "Frankenstein's monster" of sewn-together examples and toward a more human-like ability to learn new marginal skills with minimal data.