Markdown version: https://ai.hiina.space/reports/ai-memory-palaces-and-superdidactic-ai/03-superdidactic-ai.md · index

Index of everything on this site: https://ai.hiina.space/index.md

AI memory palaces and superdidactic AI: 3. Superdidactic AI

• Section 3 of 9 of the report "AI memory palaces and superdidactic AI" by Hiina (v1, updated 2026-10-27)
• Report overview and contents: https://ai.hiina.space/reports/ai-memory-palaces-and-superdidactic-ai.md
• Previous: https://ai.hiina.space/reports/ai-memory-palaces-and-superdidactic-ai/02-context-reinstatement-prompt-engineering-the-human.md
• Next: https://ai.hiina.space/reports/ai-memory-palaces-and-superdidactic-ai/04-sleep-subliminal-and-targeted-memory-reactivation.md

3. Superdidactic AI

AI tutors are good but not magic: about 0.1–0.3 SD at scale and around 0.6 SD in a tight lab trial, against a 2-sigma benchmark that was never real. Persuasion and teaching share one active ingredient, dense relevant information, and personalization adds little to either. The binding constraints are on the learner's side: engagement, retrieval effort and an encoding rate near 2 bits/s.

Tutoring effect sizes

One thing to drop: the Wang & Fan ChatGPT meta-analysis (g = 0.867) was retracted in 2026.

Persuasion as the mirror

The largest persuasion study, Hackenburg et al. in Science 2025 (76,977 people, 19 models), maps the levers cleanly. Information density correlated r = 0.76 with persuasion and an information-focused prompt added +2.29 pp. Personalization added +0.43 pp, and each 10x of model scale about +1.6 pp. 36–42% of the effect remained a month later. The warning sign: GPT-4o's claims were 62% accurate under the information prompt, against 78% under another prompt. Costello, Pennycook & Rand found ~20% durable reductions in conspiracy belief, again from tailored evidence.

So "superpersuasive" mostly means "superdense with relevant claims", and teaching gets that part for free. What persuasion doesn't need and learning does is the learner's own retrieval: belief can update from one conversation, but a skill needs effortful practice spread over time. Khanmigo shows the failure mode, a well-built tutor students barely used. Bastani shows the other: an AI that does the work for the learner raises practice scores and lowers learning. I found no study comparing one model's persuasion effect with its teaching effect.

The ceiling is the human. Landauer 1986 estimated long-term memory takes in about 2 bits/s, reaching ~10⁹ bits by midlife. Deliberate practice explains only ~12% of performance variance overall and under 1% in professions (Macnamara 2014). A superdidactic AI can remove dead time and choose what to practice; it can't raise the encoding rate.

ML spaced repetition: scheduler, cards, deck

The scheduler is mostly solved, and better models buy little. On the srs-benchmark (~10k Anki collections, ~350M reviews, no same-day reviews):

The neural models predict recall noticeably better, but a better recall predictor mostly saves review time at a fixed target retention rather than raising retention. FSRS's 21–35 interpretable parameters and its desired-retention knob are a plausible reason it ships and RWKV doesn't; the README gives no reason. I found no RCT of FSRS vs SM-2 on learning outcomes.

Content-aware scheduling looks better offline than online. KAR3L (EMNLP 2024) used card text to reach AUC 0.864 vs FSRS's 0.752 offline, but in its 27-user online test accuracy was no better (0.86 vs 0.88). LECTOR's gains come from simulated learners only.

LLM-written cards need a human to pick the targets. In a Brown medical school pipeline, 4.8–6.4% of generated cards were hallucinated and ~21.5% needed edits for information density, too much or too little. Matuschak's notes found GPT-4 writes usable declarative prompts when given pre-selected targets and an angle hint, and fails on conceptual material. An anime-subtitle vocabulary premine pipeline, described in a companion report on Japanese learning and Anki tooling, already works this way: the pipeline picked targets, and the hard part was filtering, not generation.

Replacing the deck has one good result behind it. Varied retrieval cues beat a fixed cue across seven experiments, and learners wrongly judged the fixed cue better (Butowska-Buczynska et al., PNAS 2024). That supports an LLM generating fresh questions per review from a learner's notes. Nobody has tested conversational review against fixed cards over months, or LLM grading of free recall inside the loop; both gaps look real. The base effects are solid: retrieval practice g = 0.50 (Rowland 2014, 0.73 with feedback), and the optimal gap shrinks from 20–40% of a one-week retention interval to 5–10% of a one-year one (Cepeda 2008).

The realistic "I know kung fu"

The closest real thing to a download is perceptual learning. Kellman-style modules in histopathology gave effect sizes of 0.8–1.6 on a speed-plus-accuracy score after 11–18 minutes of training, with speed gains lasting 6–7 weeks (Krasne et al. 2013). They work by many short trials over varied examples, adaptive sequencing, and retiring items after three fast correct answers. That is spaced repetition applied to pattern recognition.

My inference for the near-term stack: an LLM picks targets from the learner's own material, generates varied retrieval prompts, a neural or FSRS scheduler spaces them, perceptual modules cover recognition-heavy parts, and a coach keeps engagement up. That might compress the declarative and perceptual parts of a skill several-fold. It does nothing for motor skill or the judgment that comes from consequences, and I found no sourced figure for what share of a skill is declarative.

Text licensed CC BY 4.0 (attribution: Hiina, ai.hiina.space). Source: https://github.com/hiinaspace/ai-hiina-space