Part 4 of Raising Machines, a six-part series that started with two quotes about raising AI. Earlier parts: Gandhi Won’t Take the Pill, The β Parameter and Cats Adopted Us First. Literature cutoff: 10 October 2026.

Sixty Million Words Against Ten Trillion

In one large American recording study, children heard on average “approximately 12,300 adult words over the course of a 12-hr day” [[1]]. Jill Gilkerson and colleagues analysed 3,213 twelve-hour recordings from 329 monolingual English-speaking families in the Denver area, with children aged 2 to 48 months. The count includes overheard speech: the automated analysis “does not differentiate child-directed speech from overheard speech” and picks up adult words within roughly ten feet of the child.

Carry that daily rate from birth to the thirteenth birthday:

$$ N_{\text{child}} \approx 12{,}300 \times 365 \times 13 \approx 5.8 \times 10^{7}\ \text{words}, $$

where 12,300 is the mean adult word count per recorded day and 13 is the age the BabyLM organisers name in their introduction (their dataset budget, below, uses 12, which gives $5.4 \times 10^7$). Everything after 48 months is extrapolation. Gilkerson reports adult word counts as roughly independent of age after early infancy, which makes a flat rate reasonable but does not measure it. Reading is left out.

Meta trained the flagship Llama 3 model “with 405B trainable parameters on 15.6T text tokens” [[2]]. A token is not a word (the strawberry post explains why), but OpenAI’s rule of thumb is that “1 token is approximately 4 characters or 0.75 words for English text” [[3]]:

$$ N_{\text{model}} \approx 15.6 \times 10^{12} \times 0.75 \approx 1.17 \times 10^{13}\ \text{words}, \qquad \frac{N_{\text{model}}}{N_{\text{child}}} \approx \frac{1.17 \times 10^{13}}{5.84 \times 10^{7}} \approx 2.0 \times 10^{5}. $$

Llama 3’s own tokenizer, which Meta measures at 3.94 characters of English per token, would lower this by between 1.5% and 12%, depending on whether a word averages the 5.3 characters implied by OpenAI’s rule or six; at six, which I use for both Llama models below, it is $1.0 \times 10^{13}$ words. “Words” is loose anyway: by Meta’s account, 17% of the tokens are code [[2]].

The child side is uncertain too. Michael Frank gives a soft upper bound of $10^6$ words per month, and about $10^5$ for children “growing up in environments with limited language” [[4]]. By age 13 those rates give $1.56 \times 10^8$ and $1.56 \times 10^7$ words, which puts Llama 3’s lead between $10^{4.9}$ and $10^{5.9}$. Published estimates differ because they compare different models, ages and units. Frank says “around four or five orders of magnitude”, the first BabyLM findings “3 or 4 orders-of-magnitude” [[5]], and the second, from 2024, “five to six orders of magnitude more than the typical human language learner” [[6]]. My Llama 3 figure agrees with the last. GPT-3, trained on 300 billion tokens [[7]], comes out at about $4 \times 10^3$ against the same thirteen-year-old.

The famous number here, Hart and Risley’s word gap [[8]], in Gilkerson’s summary a projected “30-million-word deficit for low-income children by age four” [[1]], is contested. In Gilkerson’s sample, children of college-graduate mothers heard about 3,000 more adult words a day, a four-million-word gap by age four [[1]]; Sperry, Sperry and Miller (2019) concluded: “Results do not support Hart and Risley’s claim” [[9]], though theirs was not a strict replication. Only the order of magnitude matters here: tens of millions of words.

What Counts Is What a Child’s Input Contains

The numbers invite two opposite conclusions: that children must bring innate grammar, since a model needs $10^5$ times more language, or that models refute the poverty-of-the-stimulus argument by learning grammar without it. Both treat the ratio as the evidence.

The phrase comes from Chomsky’s Rules and Representations (1980), which describes a “variant of a classical argument in the theory of knowledge, what we might call ’the argument from poverty of the stimulus’” (p. 34, as quoted by Pullum and Scholz [[10]]; I have not seen the page myself [[11]]). Its classic example, which Chomsky first raised in 1965, is auxiliary fronting. In his own version, as Pullum and Scholz give it, the question formed from “The dog that is in the corner is hungry” is “Is the dog that is in the corner hungry?”, never “*Is the dog that in the corner is hungry?”. The rule fronts the auxiliary of the main clause, not the first one, so it refers to structure, not position. In the published debate with Piaget he claimed that “you can go over a vast amount of data of experience without ever finding such a case” [[12]], again as quoted by Pullum and Scholz.

The argument is not that children hear too little in total. It is that the decisive cases are rare in what they hear. Volume still matters, but only multiplied by that rate.

Geoffrey Pullum and Barbara Scholz tested the rate in 2002, mainly on adult text such as the Wall Street Journal. Their “preliminary investigations” suggested a rate “not lower than 1 percent of the interrogatives”, from which they estimated that “even a welfare child would be likely to hear about 7,500 questions that crucially falsify the structure-independent auxiliary-fronting generalization, before reaching the age of 3”, or 750 if they were “wrong by a whole order of magnitude”. They were careful about scope: “It does not challenge linguistic nativism directly, or claim that it is false.”

In the same issue, Julie Anne Legate and Charles Yang counted speech to children instead [[13]]. Across all 56 files of the Nina corpus in CHILDES, a database of transcribed child–caregiver talk, they found 46,499 adult sentences, 20,651 of them questions. None was a decisive yes–no question; 14 were wh-questions of another type Pullum and Scholz count as decisive, about 0.068 percent of the questions. The Adam corpus gave 0.045 percent, and the canonical decisive form “appears not even once in all 66,871 adult sentences” of the two corpora. Measured per adult sentence (14 of 46,499, about 0.03 percent), the Nina rate is some 40 times below the 1.2 percent they estimate a child needs to settle a comparable choice by age three. Their verdict on Pullum and Scholz: the disconfirming evidence exists, but they “fail to demonstrate its sufficiency”.

The dispute is therefore about how much evidence is enough, and that depends on the learner. A model that has read $10^{13}$ words meets decisive questions even at Legate and Yang’s rates, so its success says nothing about a learner that meets few. Nor does its appetite for data prove innate grammar, because children may have other advantages. Frank lists three candidates: “children’s pre-existing conceptual knowledge, their use of multimodal grounding, and the interactive, social nature of their input” [[4]]. The informative experiment is a model given what a child gets.

What Happens at Child Scale

That experiment has run every year since 2023 as the BabyLM Challenge, with at most 100 million words of training data in the main track and 10 million in Strict-Small. The ceiling comes from the same recordings: “Children are exposed to 2M-7M words per year (Gilkerson et al., 2017)”, which gives 24 to 84 million by age 12, “which we round up to 100M words” [[5]]. At Gilkerson’s mean rate a child passes 10 million words a little after the second birthday and would reach 100 million only at about 22 (Fig. 1).

Figure 1. Within BabyLM’s own range a child passes the Strict-Small budget between the ages of about one and a half and five, and only Frank’s upper bound reaches the Strict budget by 13. Constant rates: solid, Gilkerson et al.’s mean per recorded 12-hour day [[1]], measured to 48 months (dotted vertical) and extrapolated after; shaded, BabyLM’s 2–7 million a year [[5]], which by my arithmetic matches that mean plus or minus its largest standard deviation, about 7,000 words a day; dotted, Frank’s two rates [[4]]. The highlighted bar is BabyLM’s 24–84 million at 12.

In 2023 the winning submissions “outperformed models trained on trillions of words” on the challenge’s evaluations, and on BLiMP, a test of grammatical minimal pairs, “the top-performing model is only about 3% shy of human performance”. The organisers considered the objection that “performance metrics, like BLiMP, might not accurately measure human-level linguistic competence”, said “We are sympathetic to such concerns”, and still concluded that the close-to-human scores “reflect genuine grammatical generalizations learned by the models”. Their forecast was that “human-level results appear likely to be achieved very soon, possibly within the next few years”. In 2024 “No submissions outperformed the baselines in the multimodal track”, and “computational resources remain a crucial factor even in low-data settings” [[6]]. In 2025, with total exposure in the main tracks capped at one billion words, repeated passes included, “Some models in the Strict and Interaction tracks are comparable to a Llama 70B parameter model on the BLiMP task”, but “Models still fall short of skylines and human scores for GLUE” [[14]]. The 2023 forecast was not repeated in either later report.

The sharpest test targets the auxiliary question itself. Aditya Yedetore, Tal Linzen, Robert Frank and Tom McCoy trained LSTMs and Transformers on child-directed speech from CHILDES and found that “both model types generalize in a way more consistent with an incorrect linear rule than the correct hierarchical rule” [[15]]. They concluded that “human-like generalization from text alone requires stronger biases than the general sequence-processing biases of standard neural network architectures”, which is a claim about architectures, not about innate grammar. Alex Warstadt and Sam Bowman, surveying the field, report that “attempts to deprive current models of unfair advantages obtain sub-human results for key grammatical behaviors such as acceptability judgments”, and continue: “But before we can justifiably conclude that language learning requires more prior domain-specific knowledge than current models possess, we must first explore non-linguistic inputs in the form of multimodal stimuli and multi-agent interaction as ways to make our learners more efficient at learning from limited linguistic input” [[16]].

Word meaning looks different. Wai Keen Vong, Wentao Wang, Emin Orhan and Brenden Lake trained a fairly generic network on 61 hours of head-camera video from one child aged 6 to 25 months [[17]]. The model “acquires many word-referent mappings present in the child’s everyday experience”. NYU’s press release says the video captured “about 1% of the child’s waking hours” [[18]]. The scope is narrow: one child, mostly concrete nouns, no grammar. A follow-up from the same lab, published in 2026, reports the effect for all three children in the head-camera dataset, using automatic transcripts [[19]].

So the child-scale evidence is mixed. With 100 million words, models come close to humans on minimal pairs; trained on child-directed speech, they get the classic hierarchical case wrong; word–object mappings can be learned from a sliver of one child’s day. None of this settles innate grammar. It is the right experiment, though, which the volume comparison was not.

What a Corpus Cannot Hold

Every model in the last section learned from input that never responded to it. Frank’s third candidate, interaction, has a line of experiments behind it.

In 2003 Michael Goldstein, Andrew King and Meredith West studied 30 infants of about eight months [[20]]. In one group, mothers smiled, moved closer and touched the infant immediately after a vocalisation. In a yoked control group, mothers gave the same amount of response, but not timed to the infant. “Contingent, but not noncontingent, maternal behavior facilitates more complex and mature vocal behavior.” The groups had 15 infants each, and the yoked group was run after the contingent one rather than randomised. The authors note that the loop runs both ways: “babbling both regulates and is regulated by social interaction.”

In the same year Patricia Kuhl, Feng-Ming Tsao and Huei-Mei Liu gave nine-month-old American infants twelve 25-minute sessions of reading and play with native Mandarin speakers, on average 33,120 Mandarin syllables in all [[21]]. The exposure “reversed the decline seen in the English control group” on a Mandarin speech contrast. Other infants saw the same speakers and materials on audiovisual or audio-only recordings, and heard more of them (on average 49,866 syllables), yet “exposure to recorded Mandarin, without interpersonal interaction, had no effect”. Only 21 infants completed the live experiment and 28 the recorded one, and the test was a single phonetic contrast. Recordings do work for simpler material, as Kuhl and colleagues themselves note: infants learn from audio-only exposure to small artificial-language stimuli, as in word segmentation after two minutes of a speech stream [[22]] and distributional learning of a phonetic contrast [[23]]. Kuhl’s proposal that “the social brain ‘gates’ the computational mechanisms involved in human language learning” [[24]] is a hypothesis about “the earliest phases of language acquisition”, not a settled mechanism.

What in a live session matters? Sarah Roseberry, Kathy Hirsh-Pasek and Roberta Golinkoff taught 36 toddlers of 24 to 30 months, twelve per condition, four new verbs, two in each of two training phases of about three minutes. In their words, “children only learned novel verbs in socially contingent interactions (live interactions and video chat)”, not from yoked video [[25]]. That points to contingency rather than physical presence. A later study from Kuhl’s lab found that “social partners enhance learning, even from screens” [[26]], but the partners were other infants watching alongside, the screen itself played a clip whenever an infant touched it, and the authors read the benefit of a partner as motivation, not contingency. All of this comes from the same or neighbouring labs, and I found no independent replication of the 2003 null result for video. For attention, Michael Tomasello and Jeffrey Farrar found in 24 children that naming objects “that were already the child’s focus of attention” went with larger vocabularies at 21 months [[27]]. (Screens in early childhood have their own post, The Golden Bead Cube Weighs One Kilogram.)

These samples are Western, and that matters. Among Tsimane forager-horticulturalists in Bolivia, Alejandrina Cristia and colleagues observed “< 1 min per daylight hour” of talk directed to children under four, and they conclude that “Consideration of more diverse human populations is necessary to build generalizable theories of language acquisition” [[28]]. How much contingency a child needs is open. That it helps where it has been measured is not.

What these studies share is input timed to what the child has just done or is attending to. A pretraining corpus has none of that by construction: nothing on the web was written in reply to what a model had just predicted. That is the sense of my title. A child’s stimulus is not rich in words; a language model gets far more of those. It is rich in replies.

Where a Language Model Gets Answered

There is one stage where a model’s own output is answered. In post-training the model produces responses, people rank them, a reward model learns from the rankings, and the model is pushed towards the preferred answers: the nearest analogue of a mother responding to babbling. Frank makes the connection himself, as a hypothesis: “it is likely that this ‘interaction training’ is responsible for the success of products like ChatGPT in responding appropriately in conversation” [[4]]. In compute, OpenAI’s InstructGPT fine-tuning stages cost $(60 + 4.9)/3640 \approx 1.8\%$ of GPT-3’s pretraining [[29]].

How much human feedback is there? Two published datasets allow a like-for-like count, each against its own model’s pretraining. Anthropic’s helpful-and-harmless data [[30]] were used to fine-tune models pretrained on 400B tokens [[31]]. The Llama 2 paper reports 1,418,091 of Meta’s own comparisons, trained on top of 2 trillion pretraining tokens, and also measures the public Anthropic data: 166,353 comparisons [[32]]. Its table gives the average tokens per example (the dialogue so far plus one response) and per response, so a comparison counts as one example plus its second response:

$$ \frac{N_{\text{pref}}}{N_{\text{pre}}} \approx \frac{5.0 \times 10^{7}}{4 \times 10^{11}} \approx 10^{-3.9}\ \ (\text{Anthropic}), \qquad \frac{1.46 \times 10^{9}}{2 \times 10^{12}} \approx 10^{-3.1}\ \ (\text{Llama 2}), $$

where $N_{\text{pref}}$ is the token count of the comparison data and $N_{\text{pre}}$ is the same model’s pretraining token count. Counting one response per comparison (the per-example figure alone) lowers both by about a quarter. The Anthropic ratio divides Llama 2 tokens by Anthropic’s own. Converting Anthropic’s tokens at OpenAI’s 0.75 words per token [[3]] and Llama 2’s as below, it is $10^{-4.1}$ in words ($10^{-4.0}$ at 5.3 characters per word). In the published datasets of 2022–23, then, human preference data are three to four orders of magnitude smaller than pretraining.

Against the child, the units need care. In the first version of this post I converted the child’s words into tokens at OpenAI’s 0.75 words per token and set them against Table 6, which counts Llama 2 tokens. That tokenizer is less compact: by Meta’s later measurement it fits only 3.17 characters of English into a token [[2]], which at six characters per word is 0.53 words, so the ratios below are about 30% lower than the half and twenty I gave. The comparison data hold about $2.7 \times 10^7$ words (Anthropic) and $7.7 \times 10^8$ (Llama 2):

$$ \frac{2.7 \times 10^{7}}{5.8 \times 10^{7}} \approx 0.5\ \ (\text{Anthropic}), \qquad \frac{7.7 \times 10^{8}}{5.8 \times 10^{7}} \approx 13\ \ (\text{Llama 2}). $$

At OpenAI’s implied 5.3 characters per word, the 13 becomes about 15. I also called these tokens the feedback. They are the text people judged: their own prompts and the models’ responses, three to seven times as long per turn, with the dialogue so far counted again in every comparison. The feedback is one choice per comparison (Meta’s annotators also rated how much better [[32]]), and its counterpart in a child’s day is a reply. Gilkerson’s families averaged 441 adult–child conversational turns a day, rising to around 500 by the second birthday [[1]]. If that rate held beyond the 48 months recorded, thirteen years give about $2 \times 10^6$ turns, against $1.7 \times 10^5$ Anthropic comparisons and $1.4 \times 10^6$ Llama 2 comparisons.

Like with like, the text a 2022–23 model was judged on came to half to thirteen times a childhood’s linguistic input (Fig. 2), and the judgements numbered a twelfth to two-thirds of its conversational turns. Both are small next to pretraining and within about an order of magnitude of a childhood.

Figure 2. On one axis, the preference data fall among the child’s numbers rather than the models’. All in words: OpenAI’s 0.75 words per token [[3]] for GPT-3’s 300B tokens [[7]] and Anthropic’s 400B [[31]]; Meta’s 3.17 and 3.94 characters per token for Llama 2 and 3 [[2]] at six characters per word. Preference data: Table 6 of [[32]]. The child’s bar spans Frank’s two rates [[4]]; the dot is Gilkerson et al.’s mean [[1]].

That count leaves out a second contingent channel. In reinforcement learning, the model’s own sampled outputs are scored, by a reward model or by an automatic check. By this post’s definition every scored rollout is contingent feedback, and InstructGPT already ran that stage on 31,144 prompts. I have not measured how large this channel is in current models, and it may change the proportions above.

BabyLM has built a version of this loop into its own challenge. The 2025 Interaction track lets “student models … learn from feedback from larger teacher models” [[14]]. In the baseline, Llama-3.1-8B-Instruct revises a small student’s completions, and the student trains on the corrected text and on preference pairs with the teacher’s version as the winner. The organisers report that “interaction with teacher models can yield high-quality language models”, but the teacher, from a model family pretrained on about 15T tokens [[2]], carries that record into the student’s data, so the track does not isolate contingency. In a 2025 workshop paper, Francesca Padovani, Bastian Bunzeck and colleagues report that “PPO fine-tuning has mixed to adversarial effects on our models”, while preference optimisation without a separate reward model (DPO) “further improves their performance on our custom dialogue benchmark” [[33]]. Feedback does not automatically help.

The analogy with Goldstein’s mothers holds in one respect: the learner’s own output is answered. Elsewhere it stops. The infant’s reinforcement is immediate, physical and elicited by the infant; the model’s is an offline ranking, turned into a learned reward and applied under the KL constraint from Part 2.

An Upbringing-Sized Phase

The quote that started this series comes from the publisher’s description of De Kai’s Raising AI, which also says that the AIs behind YouTube, Reddit and Instagram “adoringly watch and learn to imitate your behavior” [[34]]. That is a claim about learning by watching, which fits pretraining and recommender systems better than feedback, and nothing here refutes it.

If any phase of training resembles upbringing in this post’s sense, it is post-training: small against pretraining, late, and comparable in size to an upbringing. Comparable means the same order of magnitude, and which way it leans depends on the unit. In words of judged text, Anthropic’s data are under half a childhood and Llama 2’s thirteen times one, or five to fifty times across Frank’s range; in replies, both fall short, at a twelfth and two-thirds. The rest is closer to inheritance, a fixed record written by people who never addressed the model. That likeness has limits. Pretraining is learning within one model’s weights, not selection across generations as in the domestication of Part 3; the nearer analogue of selection is the choice of architectures and recipes from one model generation to the next. And overheard speech, which Gilkerson’s count includes, is part of a child’s upbringing too.

The order differs as well. A child takes part in conversational turns from the first months, 200 to 300 a day in the first year by Gilkerson’s count, alongside everything else it hears. A model meets its contingent phase only after pretraining. Whether that pretrained model’s values are already formed is the question Part 1 left open, and the data sizes do not answer it.

Who Writes the Feedback

In this narrow sense a model’s parent is whoever writes the feedback. In InstructGPT that meant labellers ranking outputs; in BabyLM’s Interaction track it means a larger model. The next part follows the case where the feedback is everyone’s clicks, and asks who is raising whom.

The corpus was never poor. It just never answered.

Sources and calculations checked through 2026-10-24.

Next in the series: Who Is Raising Whom?, 24 October.

References

[1] Gilkerson, J., Richards, J. A., Warren, S. F., Montgomery, J. K., Greenwood, C. R., Oller, D. K., Hansen, J. H. L., & Paul, T. D. (2017). Mapping the early language environment using all-day recordings and automated analysis. American Journal of Speech-Language Pathology, 26(2), 248–265. https://doi.org/10.1044/2016_AJSLP-15-0169

[2] Grattafiori, A., et al. (Llama Team, AI @ Meta) (2024). The Llama 3 herd of models. Preprint, arXiv:2407.21783 (v3, 23 November 2024). https://arxiv.org/abs/2407.21783

[3] OpenAI (n.d.). Key concepts: Tokens. OpenAI API documentation (vendor documentation, accessed October 2026). https://developers.openai.com/api/docs/concepts

[4] Frank, M. C. (2023). Bridging the data gap between children and large language models. Trends in Cognitive Sciences, 27(11), 990–992. https://doi.org/10.1016/j.tics.2023.08.007

[5] Warstadt, A., Mueller, A., Choshen, L., Wilcox, E., Zhuang, C., Ciro, J., et al. (2023). Findings of the BabyLM Challenge: Sample-efficient pretraining on developmentally plausible corpora. In Proceedings of the BabyLM Challenge at the 27th Conference on Computational Natural Language Learning, 1–34. https://doi.org/10.18653/v1/2023.conll-babylm.1 (also arXiv:2504.08165)

[6] Hu, M. Y., Mueller, A., Ross, C., Williams, A., Linzen, T., Zhuang, C., Cotterell, R., Choshen, L., Warstadt, A., & Wilcox, E. G. (2024). Findings of the second BabyLM Challenge: Sample-efficient pretraining on developmentally plausible corpora. In The 2nd BabyLM Challenge at the 28th Conference on Computational Natural Language Learning, 1–21. https://aclanthology.org/2024.conll-babylm.1/ (also arXiv:2412.05149)

[7] Brown, T. B., Mann, B., Ryder, N., Subbiah, M., et al. (2020). Language models are few-shot learners. NeurIPS 2020. arXiv:2005.14165. https://arxiv.org/abs/2005.14165

[8] Hart, B., & Risley, T. R. (1995). Meaningful Differences in the Everyday Experience of Young American Children. Paul H. Brookes.

[9] Sperry, D. E., Sperry, L. L., & Miller, P. J. (2019). Reexamining the verbal environments of children from different socioeconomic backgrounds. Child Development, 90(4), 1303–1318. https://doi.org/10.1111/cdev.13072

[10] Pullum, G. K., & Scholz, B. C. (2002). Empirical assessment of stimulus poverty arguments. The Linguistic Review, 19(1–2), 9–50. https://doi.org/10.1515/tlir.19.1-2.9

[11] Chomsky, N. (1980). Rules and Representations. Basil Blackwell. (Quoted via Pullum & Scholz, 2002, whose page numbers follow this edition.)

[12] Piattelli-Palmarini, M. (Ed.) (1980). Language and Learning: The Debate Between Jean Piaget and Noam Chomsky. Harvard University Press. (Chomsky’s contributions quoted via Pullum & Scholz, 2002, pp. 114–115 of the volume.)

[13] Legate, J. A., & Yang, C. D. (2002). Empirical re-assessment of stimulus poverty arguments. The Linguistic Review, 19(1–2), 151–162. https://doi.org/10.1515/tlir.19.1-2.151

[14] Charpentier, L. G. G., Choshen, L., Cotterell, R., Gul, M. O., Hu, M. Y., Liu, J., Jumelet, J., Linzen, T., Mueller, A., Ross, C., Shah, R. S., Warstadt, A., Wilcox, E., & Williams, A. (2025). Findings of the third BabyLM Challenge: Accelerating language modeling research with cognitively plausible data. In Proceedings of the First BabyLM Workshop, 399–420. https://doi.org/10.18653/v1/2025.babylm-main.28

[15] Yedetore, A., Linzen, T., Frank, R., & McCoy, R. T. (2023). How poor is the stimulus? Evaluating hierarchical generalization in neural networks trained on child-directed speech. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 9370–9393. https://doi.org/10.18653/v1/2023.acl-long.521

[16] Warstadt, A., & Bowman, S. R. (2022). What artificial neural networks can tell us about human language acquisition. In S. Lappin & J.-P. Bernardy (Eds.), Algebraic Structures in Natural Language, 17–60. CRC Press. https://doi.org/10.1201/9781003205388-2 (also arXiv:2208.07998)

[17] Vong, W. K., Wang, W., Orhan, A. E., & Lake, B. M. (2024). Grounded language acquisition through the eyes and ears of a single child. Science, 383(6682), 504–511. https://doi.org/10.1126/science.adi1374

[18] New York University (2024, February). AI learns through the eyes and ears of a child. NYU News (press release). https://www.nyu.edu/about/news-publications/news/2024/february/ai-learns-through-the-eyes-and-ears-of-a-child.html

[19] Vong, W. K., & Lake, B. M. (2026). On the robustness of modeling grounded word learning through a child’s egocentric input. Open Mind, 10, 1253–1278. https://doi.org/10.1162/opmi.a.377 (preprint arXiv:2507.14749)

[20] Goldstein, M. H., King, A. P., & West, M. J. (2003). Social interaction shapes babbling: Testing parallels between birdsong and speech. PNAS, 100(13), 8030–8035. https://doi.org/10.1073/pnas.1332441100

[21] Kuhl, P. K., Tsao, F.-M., & Liu, H.-M. (2003). Foreign-language experience in infancy: Effects of short-term exposure and social interaction on phonetic learning. PNAS, 100(15), 9096–9101. https://doi.org/10.1073/pnas.1532872100

[22] Saffran, J. R., Aslin, R. N., & Newport, E. L. (1996). Statistical learning by 8-month-old infants. Science, 274(5294), 1926–1928. https://doi.org/10.1126/science.274.5294.1926

[23] Maye, J., Werker, J. F., & Gerken, L. (2002). Infant sensitivity to distributional information can affect phonetic discrimination. Cognition, 82(3), B101–B111. https://doi.org/10.1016/S0010-0277(01)00157-3

[24] Kuhl, P. K. (2007). Is speech learning ‘gated’ by the social brain? Developmental Science, 10(1), 110–120. https://doi.org/10.1111/j.1467-7687.2007.00572.x

[25] Roseberry, S., Hirsh-Pasek, K., & Golinkoff, R. M. (2014). Skype me! Socially contingent interactions help toddlers learn language. Child Development, 85(3), 956–970. https://doi.org/10.1111/cdev.12166

[26] Lytle, S. R., Garcia-Sierra, A., & Kuhl, P. K. (2018). Two are better than one: Infant language learning from video improves in the presence of peers. PNAS, 115(40), 9859–9866. https://doi.org/10.1073/pnas.1611621115

[27] Tomasello, M., & Farrar, M. J. (1986). Joint attention and early language. Child Development, 57(6), 1454–1463. https://doi.org/10.2307/1130423

[28] Cristia, A., Dupoux, E., Gurven, M., & Stieglitz, J. (2019). Child-directed speech is infrequent in a forager-farmer population: A time allocation study. Child Development, 90(3), 759–773. https://doi.org/10.1111/cdev.12974

[29] Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C. L., Mishkin, P., Zhang, C., et al. (2022). Training language models to follow instructions with human feedback. NeurIPS 2022. arXiv:2203.02155. https://arxiv.org/abs/2203.02155

[30] Bai, Y., Jones, A., Ndousse, K., Askell, A., et al. (2022). Training a helpful and harmless assistant with reinforcement learning from human feedback. Preprint, arXiv:2204.05862. https://arxiv.org/abs/2204.05862

[31] Askell, A., Bai, Y., Chen, A., Drain, D., et al. (2021). A general language assistant as a laboratory for alignment. Preprint, arXiv:2112.00861. https://arxiv.org/abs/2112.00861

[32] Touvron, H., Martin, L., Stone, K., Albert, P., et al. (2023). Llama 2: Open foundation and fine-tuned chat models. Preprint, arXiv:2307.09288. https://arxiv.org/abs/2307.09288

[33] Padovani, F., Bunzeck, B., Ali, M., Momen, O., Bisazza, A., Buschmeier, H., & Zarrieß, S. (2025). Dialogue is not enough to make a communicative BabyLM (but neither is developmentally inspired reinforcement learning). In Proceedings of the First BabyLM Workshop, 421–435. https://doi.org/10.18653/v1/2025.babylm-main.29 (also arXiv:2510.20358)

[34] De Kai (2025). Raising AI: An Essential Guide to Parenting Our Future. MIT Press. Quoted from the publisher’s description of the book. https://doi.org/10.7551/mitpress/15834.001.0001

Changelog

  • 2026-10-24: Corrected the comparison of preference data with a child’s input. The first version converted the child’s words to tokens at OpenAI’s 0.75 words per token against counts in Llama 2’s tokenizer (3.17 characters per token, 0.53 words at six characters per word) and called the judged text the feedback. In words the data are 0.5 and 13 times a childhood’s input (was 0.6 and 19 in tokens; 0.5 and 15 at 5.3 characters per word); the Anthropic share of pretraining is $10^{-4.1}$ in words ($10^{-3.9}$ in mixed tokens); the abstract’s “factor of about twenty” is now “about an order of magnitude”, and the text says which way each unit leans. Added Figs. 1 and 2. Corrected Roseberry et al.’s design (four verbs, two per phase, not two), noted that Lytle et al.’s screen responded to touch, added Padovani et al.’s DPO result, cited the published versions of that paper and Vong & Lake, replaced the Llama 3 tokenizer adjustment of “about a tenth” with 1.5% to 12% depending on characters per word, and linked Part 5.