---
id: ai-memory-palaces-and-superdidactic-ai/05-anki-for-what-you-ought-to-do
title: 'AI memory palaces and superdidactic AI: 5. Anki for what you ought to do,
  and the interface endgame'
report: https://ai.hiina.space/reports/ai-memory-palaces-and-superdidactic-ai.md
section: 5
version: 1
updated: '2026-10-27'
author: Hiina
---

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

# AI memory palaces and superdidactic AI: 5. Anki for what you ought to do, and the interface endgame

- Section 5 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/04-sleep-subliminal-and-targeted-memory-reactivation.md
- Next: https://ai.hiina.space/reports/ai-memory-palaces-and-superdidactic-ai/06-the-far-end-uploads-distillation-and-the.md

## 5. Anki for what you ought to do, and the interface endgame

Prompting people toward their own goals works, but after bias correction the effects are small and they decay. The value of a personal agent assistant is more likely in capturing intentions that would otherwise be lost and tying them to concrete cues than in clever timing. Nobody has framed intentions as a spaced-repetition problem; the gap is real, and the building blocks exist.

### The effect sizes, corrected

| Intervention | Naive effect | After bias correction or at scale |
| --- | --- | --- |
| Implementation intentions ([Sheeran, Listroma & Gollwitzer 2025](https://kops.uni-konstanz.de/server/api/core/bitstreams/d703c468-46e9-47fc-8900-d32d7d19c8d9/content), 642 tests) | d = .36 | d = .15 (RoBMA); field about half of lab |
| Nudges in general ([Maier et al. 2022](https://pmc.ncbi.nlm.nih.gov/articles/PMC9351501)) | d = .43 (Mertens 2022) | d ≈ .04 |
| Nudge units ([DellaVigna & Linos 2022](https://www.nber.org/papers/w27594), 126 RCTs, 23M people) | 8.7 pp in journals | 1.4 pp in deployed trials |
| Gym megastudy ([Milkman 2021](https://authors.library.caltech.edu/records/mys88-cc756), 61,293 people) | 45% of 54 interventions beat control | 8% still worked after the program ended |
| Flu-shot texts ([Milkman 2022](https://knowledge.uchicago.edu/record/9653/files/milkman-et-al-2022-a-680-000-person-megastudy-of-nudges-to-encourage-vaccination-in-pharmacies.pdf), N = 689,693) | best +2.7 pp | experts couldn't predict winners (r = 0.03) |
| HeartSteps just-in-time prompts ([Klasnja 2019](https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=EXT_ID:30192907%20AND%20SRC:MED&resultType=core&format=json), N = 44) | +107% steps early | +14% overall (p = .06), fading over weeks |
| Chatbot exercise coaching ([meta-analysis, 12 RCTs](https://pmc.ncbi.nlm.nih.gov/articles/PMC12254675)) | SMD .29 under 8 weeks | .06 past 8 weeks; no LLM trials included |
| Commitment contracts ([Giné, Karlan & Zinman](https://poverty-action.org/sites/default/files/publications/put_your_money_where_your_butt_is.pdf)) | +3.4–5.7 pp quit at 12 months | only 11% took the contract |

LLM-specific evidence on behavior is close to nil. The one randomized field trial found is a [2026 preprint](https://arxiv.org/abs/2604.03881) with 233 dorm residents: LLM-personalized messages cut electricity use relative to plain text feedback, with no significant effect on hot water. [GPTCoach](https://hci.stanford.edu/publications/2025/Joerke_GPTCoach.pdf) measured no behavior at all.

### What the literature says to build

- **Specific cues beat reminders.** Plans with a time and a place outperform time alone (d = .46 vs .25 in the 2025 meta-analysis). An assistant that turns "I should look at that design decision" into "after coffee Saturday, at the desk, open the queue" is doing the active ingredient.
- **Timing by receptivity is worth something.** ML-timed messages got up to 40% better receptivity than random timing ([Mishra 2021](https://arxiv.org/abs/2011.08302)), with predictors like phone use, activity and time of day ([Künzler 2019](https://cs.dartmouth.edu/dfk/research/kunzler-receptivity/index.html)).
- **Novelty decays, so rotate.** Duolingo's recovering bandit penalizes recently used message templates with a 15-day half-life ([Yancey & Settles 2020](https://research.Duolingo.com/papers/yancey.kdd20.pdf)). The online gain was +0.4–0.5%, a reminder that timing optimization is a small lever.
- **People under-offload.** Gilbert's lab finds people are underconfident in unaided memory yet set reminders by perceived rather than actual need, and reminders help most under high load ([Chiu & Gilbert 2023](https://discovery.ucl.ac.uk/10176285/1/chiu-gilbert-2023-express-influence-of-the-physical-effort-of-reminder-setting-on-strategic-offloading-of-delayed.pdf)). Capture is the bigger win.
- **Fade prompts to build a habit.** [Wood & Rünger 2016](https://dornsife.usc.edu/wendy-wood/wp-content/uploads/sites/183/2023/10/wood.runger.2016.pdf) argue habits bind to stable context cues; an external prompt that never fades substitutes for the internal cue. That is the same tension the companion report on AI enfeeblement covers from the other side.

An "SRS for intentions" would treat each commitment as a card: dense prompts right after a plan is formed, lengthening intervals when the person acts, shorter ones and a changed cue after a miss, a contextual bandit with a dose penalty picking time and channel, and rotated wording. HeartSteps and Oralytics already encode prompt "dosage" as a decaying count with a burden cost ([Oralytics design](https://arxiv.org/html/2406.13127v1)). The difference from flashcards is the outcome signal: "did it" is sparse and late, where recall is graded instantly. The ought-driven tab feed in the companion report on personal web-surfing models is the reading-side instance of the same scheduler.

### Model-rendered computers

Pixel-rendering world models went from DOOM at \~20 fps ([GameNGen](https://arxiv.org/abs/2408.14837), 2024) to 720p at 24 fps with about a minute of visual memory ([Genie 3](https://deepmind.google/discover/blog/genie-3-a-new-frontier-for-world-models/), 2025). Memory horizons are still minutes. The only real "neural desktop", [NeuralOS](https://arxiv.org/abs/2507.08800) (ICLR 2026), runs 512×384 at 18 fps on an H100 after \~23k GPU-hours. It tracks the cursor to 1.6 px but predicts the right state transition only 37.7% of the time, recalls a created folder about half the time after 64 frames, and can't reliably render typed characters. Its authors call inference too costly for practical use.

Production generative UI is LLM-written code, not pixels: [Google's Generative UI](https://research.google/blog/generative-ui-a-rich-custom-visual-interactive-user-experience-for-any-prompt/) (Gemini 3, Nov 2025) and Anthropic's five-day Imagine with Claude preview. The evidence is rater preference over chat, not task performance. The only measured performance gain for generated personalized UIs is old and narrow: [SUPPLE++](https://www.eecs.harvard.edu/~kgajos/papers/2008/gajos08improving.shtml) made motor-impaired users 26.4% faster with 73% fewer errors. Adaptive-UI research also warns that interfaces that keep changing break motor memory ([Findlater 2009](https://www.cs.ubc.ca/labs/edapt/papers/findlater2009.pdf)). For the "computer optimized for what one ought to do" endpoint, that argues for model-generated but stable surfaces with state kept in ordinary program memory, not a model streaming pixels.

### BCI as the endpoint

| System | Throughput | Caveat |
| --- | --- | --- |
| [Willett 2023](https://www.nature.com/articles/s41586-023-06377-x) speech implant | 62 wpm, 23.8% WER at 125k words | 1 participant |
| Metzger 2023 ECoG ([summary](https://physicsworld.com/a/high-performance-brain-implants-restore-communication-to-those-who-cannot-speak/)) | 78 wpm, \~25% WER | 1 participant |
| [Card 2024](https://profiles.wustl.edu/en/publications/an-accurate-and-rapidly-calibrating-speech-neuroprosthesis/) speech implant | \~32 wpm, 97.5% over 8.4 months | 1 participant |
| Neuralink cursor ([Decrypt](https://decrypt.co/230021/neuralink-implant-patient-gaming-sets-bci-record)) | 8 bits/s (2024); "above 10" claimed 2026 | company-reported only |
| Meta sEMG wristband ([The Register](https://www.theregister.com/2025/07/23/meta_wristband_ui_control/)) | 20.9 wpm handwriting, no calibration | phone keyboard is \~36 wpm |
| [Brain2Qwerty](https://arxiv.org/abs/2502.17480) non-invasive | 32% character error (MEG), 67% (EEG) | 35 volunteers |

Writing into the brain is much further behind. Tactile microstimulation halved a grasp task's time in one participant ([Flesher 2021](https://pubmed.ncbi.nlm.nih.gov/34016775/)), visual-cortex arrays give crude phosphenes and a few letters ([Fernández 2021](https://research.vu.nl/en/publications/visual-percepts-evoked-with-an-intracortical-96-channel-microelec/)), and nothing credible writes semantic content.

The deeper constraint is [Zheng & Meister](https://arxiv.org/abs/2408.10234): human behavior runs at about 10 bits/s against \~10⁹ bits/s of sensory input. Motor BCIs have reached roughly mouse parity, which is about where that ceiling sits. If the bottleneck is central, a better pipe doesn't help much. The "computer optimized for what you ought to do" is then the one that spends a person's 10 bits/s only on decisions and taste calls and expands each into rich output. That is a decision queue, which brings this back to section 2.

---

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