The 150,000-to-1 gap: scaling is a workaround for a learning mechanism nobody has found

The multiple is worth doing by hand. Technology Review puts the difference at roughly a hundred thousand times more words for a model than for a person mastering a first language ^[11]; the published figures tighten that. Llama 3.1's 15 trillion pretraining tokens ^[2] against a well-read preteen's 100 million words ^[4] is about 150,000 to one ^[14]. If Wilcox is right that frontier runs use ten times more data ^[3], the ratio is nearer 1.5 million to one ^[15]. Tokens are word-like chunks rather than words ^[13], so read these as orders of magnitude, not measurements. They do not get smaller at any level of precision.

What makes the ratio a problem rather than a curiosity is where the low end sits. A toddler is producing grammatically correct sentences after something like 10 million words, 30 million at the outside ^[5]. That is roughly six orders of magnitude below the last publicly stated frontier-scale pretraining volume ^[16]. The comparison is not that machines are behind children; it is that the two systems are not on the same curve, and only one of them has a known method.

For the past decade, language models have mostly improved by getting bigger ^[17]. The alternative, learning more from less, has no owner and no schedule, because exactly how babies do it remains unknown ^[8]. It is not even settled whether a child arrives with a language instinct or could in principle learn a language from experience alone ^[9]. That second uncertainty is the one operators should weigh. If part of the child's efficiency is biological rather than architectural, then "a more data-efficient learner" is not an engineering task waiting for headcount. It is a bet on a result that cognitive science and machine learning have both failed to produce so far.

The field's own name for this, the data efficiency gap ^[1], flatters the situation. A gap implies a measured distance and a crossing. What the researchers describe is closer to an absent component: the only learner that reaches fluency on a household's worth of language cannot be inspected well enough to copy, while the learner that can be inspected needs a city's worth of text and then some ^[8].

Anyone treating data volume as the dependable input to capability is relying on a supply with an estimated end date ^[7] and no demonstrated successor. That is not a reason to discount current models, which work. It is a reason to stop describing the next order of magnitude as progress toward efficiency, when it is the opposite of it.