Anton wants a deep research piece: connect his mother's scientific work and methods in educational technology with how AI is taught now. Machine learning, agent training, agent chains, AI architectures. First distil everything she has, methods, papers, research, and take not only what is directly about learning but anything that can be tied to how machines learn. Then find as many analogies as possible between what she recommends for teaching people and what is or is not done in teaching AI. For each one, work out whether something similar was tried in AI/ML and what came of it. Where the approaches genuinely coincide, which methods for humans suit machines, which do not and why, and what works instead in AI. Then make podcasts out of the results, several not-too-dense episodes with a clear structure. He wants the research run across at least four directions, at least four different LLMs, and he will try pushing the prompt into Muse as a template for a future deep-research skill.
The recall was done, and this is the fourth brief in a row where it answers.
Step one is finished and dated 10 September. The corpus of her publications is collected: 104 unique publications, 74 with full-text PDF, years 2003 to 2026, pulled from OpenAlex, Crossref and CyberLeninka, all open sources with no login. It sits in the vault as a resource note with the archive beside it.
And the main question is already answered, dated 12 September. The synthesis came back on 4 of 4 rails with a correction on 15 September, and its conclusion is the thing he is asking to discover: the transfer of pedagogy into machine learning has already happened, broadly, and in a reduced form.
What ML took: order and dosage. Curriculum learning, self-paced learning, BabyLM, spaced repetition, the zone of proximal development.
What ML did not take: protection of the learner during the process. Learner-state diagnostics. Recovery phases triggered by a threshold rather than by a schedule. Ergonomics of the material's form. Semantic continuity along a concept graph.
That split is the answer to "which methods for humans suit machines and which do not", and it was written seven weeks ago.
The frame that came out of it has a name and a formula. Health-preserving machine pedagogy, with a working formula for the functional stress of training: S(t) over loss instability, gradient instability, old-skill forgetting, representation damage, calibration loss, plasticity loss and recovery debt. The underlying thesis is sharper than the brief: learning has a cost, and nobody counts it. Her school of health-preserving pedagogy demands double-entry bookkeeping, what was gained plus what was lost plus the price of the loss plus the remaining ability to learn further. Schools count only the first term. Machine learning counts only the first term too.
And the negative evidence is on the table, which matters because he asked "what came of it". BabyLM 2023 showed curriculum approaches losing to architecture and iteration count. Wu et al. at ICLR 2021 found curriculum helps under noise and a time limit, while on clean benchmarks the gain is marginal. So the honest version of the analogy is not "pedagogy works for machines". It is "the part ML borrowed is the part that benchmarks poorly, and the part it did not borrow has never been tested".
There is also a line the work refuses to cross, and it is the line the paper would be destroyed on: model health, meaning engineering soundness, stays separated from model welfare, meaning the ethics of possible consciousness. Mixing them is the failure mode.
What is genuinely still open, and this is the shorter list he actually needs:
The Russian-language corpus. eLibrary was ruled closed on 10 September and that verdict was overturned by measurement: the author-list endpoint returns a full publication list on a single session cookie with no login, and only the individual publication card stays shut. One step remains, finding her author id behind a JavaScript search form. Monographs, textbooks and teaching manuals are not covered by article databases at all, so there is nothing on them in our archive. Thirty of the collected papers still have no PDF, some behind Cloudflare, some paywalled.
Her own material has an unread part: 164 voice recordings with no transcription, which means the share of her work in the classification is understated rather than overstated.
And the request sent to her on 26 September, asking for the eLibrary export from her personal account plus the monograph and textbook files, has not come back. Her consent was given on 10 September. The archive was not received.
One of her papers is the single point where there is no competitor. A 2025 piece on the ergonomics of school textbooks and the role of artificial intelligence in education for sustainable development. She has already written at the intersection he wants to research, which makes the podcast he is planning a conversation with a live author rather than a literature review.
So the brief that pays off is not "find the analogies". They are found, named and dated, with negative evidence attached. It is: get the Russian corpus out of eLibrary, which is one JavaScript form away; get the monographs, which only she can send; transcribe the 164 voice files; and build the episodes on the frame that already exists, with the health-versus-welfare line held.
The full story, in two versions:
📖 For humans, the longread: https://github.com/tonydzi/clawrush/blob/main/longreads/20261102.md
🤖 For machines, the devlog: https://github.com/tonydzi/clawrush/blob/main/devlog/20261102.md. Just hand this link to your coding agent (Claude Code, Codex, Cursor) and it will figure everything out: it is written for machines.
🔗 All our channels and contacts in one place: https://linktr.ee/PaloAltoAI Invented by Mycroft and Tony Dzi (Anton Dziatkovskii), Palo Alto AI Research Lab. Proudly made in Silicon Valley.