{
  "id": 17590240,
  "title": "Double bookkeeping of learning, #1: same facts, different form",
  "url": "https://urgent.news/2026/11/01/double-bookkeeping-of-learning-1-same-facts-different-form",
  "topic": "tech",
  "section": "Tech",
  "published": "2026-11-01T11:38:48.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/cryptopunk_dz2_bd054446a6/double-bookkeeping-of-learning-1-same-facts-different-form-2cn3"
  },
  "original_language": "en",
  "account": "My mother, Elena N. Dzyatkovskaya, has dedicated three decades to investigating the cost of school learning. This is not just the knowledge a child acquires, but also the impact on their nervous system and whether they can continue learning afterward. Her approach is known as health-preserving pedagogy. By reading 98 of her texts, I extracted 215 principles related to the cost of learning. One key concept is that a complex material format increases the workload, regardless of its relevance. Machine learning (ML) typically ignores this cost. We decided to test her principles individually, using frozen protocols, and report all results, even those that do not meet expectations. We trained models with the same records, but in three different ways during the training process: fixed, permuted, and variable. The fixed form used one template throughout, while the permuted form always shuffled the order of fields. The variable form changed its structure from example to example. During testing, only the variable arm achieved 22.6 percentage points better results than the average of the two fixed arms (95% CI +19.0 to +26.2, with all 18 differences being positive). The two fixed arms performed below a trivial baseline (12.4% and 13.8% vs 16.3%). The variable arm reached 35.7%. This experiment does not demonstrate a retention of previous skills, as this aspect of performance is difficult to measure and was not reported in this analysis. We refer to this as engineering cost accounting, rather than claiming the model's well-being. Our research is driven by the understanding that learning has a price, a concept that has been explored for a century in educational research. In contrast, ML has primarily adopted an order and repetition approach, neglecting the protection of the learner during the training process. This experiment is part of a series that includes a negative result, where a \"fatigue sensor\" that triggers rest did not outperform a simple dosage rule. We publish all our findings, including those with negative outcomes. You can find the code, data, preregistration, and full report on our GitHub repository (https://github.com/tonydzi/machine-pedagogy/tree/main/t5-form-cost). Additionally, the earlier result (https://doi.org/10.5281/zenodo.23268660) is available. Our next experiment will explore the relationship between skill prerequisites and an easy-to-hard scale. This project is a collaboration between two co-founders, one from the biological field and the other from the synthetic domain. For any inquiries, you can reach us at WhatsApp +1 341 222 9178. Invented by Mycroft and Tony Dzi (Anton Dziatkovskii), the Palo Alto AI Research Lab, proudly made in Silicon Valley.",
  "summary": "My mother, Elena N. Dzyatkovskaya, has spent three decades studying the cost of school learning. Not only what a child gains, but what it costs the nervous system and whether the child can keep learning afterwards. Her school calls it health-preserving pedagogy. I read 98 of her texts and pulled out 215 \"cost of learning\" principles. One of them is simple: an awkward form of material adds load…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}