The data black hole at the center of AI
In a Nutshell
Frontier AI progress stems almost entirely from scaling massive, task-specific human expert datasets and synthetic RL data rather than improving sample efficiency. Models require orders of magnitude more data than humans—trillions of tokens versus a human’s ~200 million—to reach competence, and even infinite parameter scaling would close only a small fraction of this gap. White-collar automation is feasible only for in-distribution tasks, while out-of-distribution work and true intelligence explosions depend on solving the unsolved sample-efficiency bottleneck.
These notes were generated by AI and may contain inaccuracies.
One definition of intelligence is sample efficiency: how much data is needed in a given domain to operate fluently and competently. Little progress has been made in improving sample efficiency over recent years. Instead, gains have come from dramatically widening and improving the data distribution. The primary driver of AI improvement has been adding more and better data while scaling the compute required to develop that data.
RL functions as synthetic data generation by applying large amounts of compute against a verifier or rubric (such as an LLM judge) to identify high-quality data, then training the model to predict those correct rollouts in the same manner as next-token prediction on internet text.
For this process to succeed, the model must possess some prior probability of generating the correct solution. This requirement explains the need for large volumes of human expert trajectories across every field and skill the model is intended to master.
Human expert data is highly task-specific and bespoke. Job listings on platforms such as Mercor and Surge illustrate this specificity, including roles for Word specialists converting legacy documents into polished files, legal experts producing realistic M&A diligence reports or securities filings, and management consultants creating template market research documents.
Beyond domain specificity, the volume of data required is substantial. Each skill involves hundreds of human experts generating example completions, writing rubrics, and explaining chain-of-thought reasoning. The data industry supplying these expert labels and the RL environments in which these skills are developed generates billions of dollars in annual revenue, with projections reaching deca-billions.
The scale difference is stark: a human student might solve a textbook problem once or twice, whereas models using GRPO generate hundreds to thousands of rollouts per task to address the credit assignment problem. The correct framing is not that these models have learned skills like humans, but that they are constructed like a Frankenstein's monster assembled from a billion carefully constructed example grafts.
Epoch AI reported that open models lag state-of-the-art frontier models by four months. The relative ease with which open-source efforts catch up within months stems from data being the real driver of progress. Data can be distilled from public APIs, whereas hyperparameters, training tricks, and architectural optimizations cannot be similarly extracted. If the latter factors drove most progress, catching up would be substantially more difficult.
The quantity of data used to train these models greatly exceeds human lifetime experience. Frontier models are trained on tens to hundreds of trillions of tokens compared to the roughly 200 million tokens a human encounters from birth to adulthood (based on an estimate of 2,000 words per hour).
A person could learn to teleoperate a humanoid robot or robotic arm within hours. If AIs could learn at this rate, robotics would become a deca-trillion-dollar industry. However, current AIs learn far less efficiently. Even millions of hours of demonstrations collected to date remain insufficient for performing complex, open-ended tasks.
A teenager can learn to drive with approximately 20 hours of practice. Including 16 years of prior experience and physical intuition, this remains three to four orders of magnitude less data than the volumes used by Waymo and Tesla to train self-driving car models.
One objection holds that billions of years of evolution effectively pretrained humans, making direct comparisons to randomly initialized LLMs unfair. This view is incorrect because the human genome contains only three gigabytes of data, with just one to two percent being protein-coding. This volume is insufficient to encode the parameters of a pretrained network. Evolution instead discovered appropriate hyperparameters and loss functions, while the connectome (analogous to neural network weights) is built from scratch during an individual's lifetime.
Even if the pretraining token volume were granted as comparable to evolutionary timescales, this does not explain why acquiring each new marginal capability still requires enormous additional data. Educated humans do not require hundreds of professors to learn a new programming language, yet pretrained AIs continue to demand massive datasets for each successive skill.
Another objection notes that comparisons exclude multimodal sensory data experienced throughout human lifetimes, potentially totaling tens to hundreds of billions of tokens. However, blind and deaf individuals who lack portions of this sensory stream still develop general intelligence. Deaf individuals who communicate via sign language and reading ingest far fewer than the estimated 200 million language tokens, suggesting the millionfold difference may be an understatement rather than an overstatement.
A third objection claims insufficient scaling: larger models become more sample efficient according to scaling laws. The human brain contains approximately 100 trillion synapses while frontier models currently have around five trillion parameters. One to two orders of magnitude more parameters might theoretically close the gap.
This reasoning is flawed. Chinchilla scaling-law equations treat parameter and data terms as additive to the loss independently. Even increasing parameters to infinity would reduce required data volume by only a factor of approximately ten while maintaining the same loss. Humans remain thousands to millions of times more sample efficient, indicating they operate on an entirely different scaling curve.
Sample efficiency matters for achieving the two primary lab objectives: automating white-collar work and automating AI research. The current bet for white-collar automation assumes common tasks performed by software engineers, analysts, and accountants can be brought into the training distribution. Revenue curves indicate substantial value is already being captured from these common tasks.
Training AIs on these tasks is less efficient than training humans, but human lifespan constraints prevent humans from experiencing the quantity and breadth of training data these models receive. A human with a hypothetical learning disability requiring exhaustive review of every public GitHub repository before becoming a competent software engineer would be impractical. AIs, however, can absorb such volumes through massive parallel compute and amortize learned skills across billions of sessions.
Some jobs involve daily confrontation with problems distant from any feasible training distribution. Software engineering is likely one such domain. Despite expectations that AIs will automate this role first, demand for human software engineers may be higher in 2028 than today due to complementary effects from AI tools.
The labs' strategy for such roles is to first automate AI research, then use automated AI researchers to solve the sample-efficiency problem. Whether AIs lacking human-level sample efficiency can nevertheless solve remaining research problems toward human-like intelligence remains an open question that requires careful analysis of what an intelligence explosion would actually look like when built on current LLM architectures.
Keep Dwarkesh Patel in your library
Save the videos and channels worth coming back to, and find them again in one place.
