Why
The greed for learning is not a flaw to fix. It is the same signal that makes a good engineer read the spec twice. The problem is a mismatch of speeds: inputs arrive in minutes, understanding takes hours, and a day has a fixed number of them. Two wrong responses follow. One is to consume everything shallowly, which feels like learning and leaves nothing, the cognitive offloading Ng warns about. The other is to feel guilty and stop, which wastes the signal. The right response is a queue with a filter: everything you are curious about gets captured in seconds, only what serves the through line gets the hours, and the capture itself relieves the pressure, because the mind stops rehearsing what it might forget. Today's site already does the capture. The rest is a small amount of discipline placed where the environment does the work.
How it works
One through line decides depth
Greene's test is the filter. Write the through line in one sentence, the thing all the projects are versions of. For Jayverse it is close to: programmable money whose rules can be verified, and the tooling that lets a small team build it with agents. Then sort every wish into three bins:
- Learn deeply. It moves the through line. This week that is agent harnesses and boundaries (Tech #62, #104, #106), evaluation and tracing (#102), the money rails (#103), and the market basics (Invest 800 to 815). These get the hours and a Verified and unverified paragraph written by you, not pasted.
- Learn enough to converse. Adjacent and interview-relevant: robotics (#100, #113 to #116), world models, spatial intelligence. These get one item each and one sentence you could say in an interview, and no more until they touch a project.
- File and release. Interesting, unconnected. These get a card and a Key expressions table and nothing else. Filing is the point. Once written down they stop costing attention.
The sorting is itself a five-minute daily act, and it is where most of the relief comes from. Ng's context advantage argues for the same bias: depth comes from years in one domain, and the domain is the through line.
A fixed slot that ends in an artefact
Djokovic's environment rule and Fraza's curriculum rule agree: pick a time, not an amount of motivation. Two slots are enough.
| Slot | When | Length | Output |
|---|---|---|---|
| Daily | same clock time each day | 30 to 45 minutes | one item moved from New to Done, in your words |
| Weekly | one fixed weekend hour | 60 to 90 minutes | one deep-bin item worked by hand: rebuild the model, run the code, draw the chart |
The rule is that a slot ends with something written or built, because Fraza's point about output is the same as Ng's about offloading: retention follows production. Reading and highlighting do not count as the artefact. The Today and Yesterday buttons on the rail are the scoreboard for the daily slot. If Today is empty at night, the slot was skipped, and the number says so without an argument.
Three questions close every session
From Fraza, appended to the day's history file in three lines: what did I learn, what is still unclear, what is the next step. The second line is the one that matters for greed. The unclear thing is written down, so it is safe, and it becomes tomorrow's first candidate instead of a background worry. Over weeks those lines are the real curriculum, better than any list of videos, because they were generated by friction with actual material.
Replay the interval where it turned
Kobe's advice to Djokovic applies to learning as much as to tennis. When a project stalls or a concept refuses to land, the useful thing is not the summary but the exact interval: which commit range, which paragraph, which step in the derivation. Write that interval into the history entry. The three-line close asks what was unclear; this asks where. Together they make the next session start at the right place.
Let the system hold the index
Karpathy's pattern, as Tina Huang's video lays it out, has three layers: raw sources that are never edited, a wiki the model maintains, and a schema that tells the model how to read it, with an index and a log. Knowledge Notes has the wiki, the schema (the topics README), and the log (history). It lacks an honest raw layer and a machine-readable index. Both are small additions and both reduce the load on memory: the raw layer means a source can be reopened rather than re-remembered, and the index means the question "have I already learned this" is answered by a search rather than by anxiety. The greed shrinks when forgetting stops being a risk.
Accept the number
The Auditor's habit belongs here too. Today the site reads 61 done of 264 Tech items, 14 percent overall. That is not a failure rate, it is the honest ratio of curiosity to hours, and it is what the 1,000-done celebration is measuring. A learning plan that cannot show its ratio is a wish. This one can, every day, in the rail.
Where it lands in Jayverse
- Knowledge Notes: add the front filter and the back three lines. A one-word bin per item at capture time (deep, converse, file) and the three closing questions at the end of each day's history file. Both are conventions, not code, and both can start tomorrow. Started 2026-09-21:
--bin deep|converse|filein the kicker and the index; a Closing three block ends each day's history file. - Rail: the daily slot has a scoreboard already. Today, Yesterday and Done are the learning metrics; the weekly hour should end by pressing Done on one deep-bin item after working it by hand, not after reading it.
- alice as the LLM wiki: raw layer and index. A
docs/topics/raw/folder for the pasted briefings and a generatedindex.mdwith one line per item are the two missing layers from Tech #109; they turn "did I already learn this" into a grep. Built on 2026-09-21:docs/topics/raw/,index.md+index.json(scripts/build-index.py),scripts/notes-search.py. - Eng and interviews: the converse bin is the interview bank. One sentence per adjacent topic, kept current, is exactly what a team-lead interview abroad draws on, and it is the cheapest form of learning the greed allows. Started 2026-09-21:
docs/topics/interview-bank.md, one row per converse-bin item (--sentence).
Verified and unverified
Verified on 2026-09-19: the sources this item combines are on the site as Tech #97, #104, #106, #109, #119, Life 1304 to 1309, and Invest 815, each with its own Verified and unverified paragraph; the site's counts quoted here are read from the rail today. The method itself is a synthesis, not a study result: the claims that production improves retention and that fixed cues beat motivation are consistent with mainstream learning research (retrieval practice, habit cueing), but the specific slot lengths and the three-bin sort are recommendations for jay, not measured findings. Their test is the ratio in the rail over the next months.
Sources: Tech #97 (an AI engineer builds the car), #104 (Pocock, fundamentals), #106 (harness engineering), #109 (Obsidian and the LLM wiki), Life 1304 (Ng: tasks not jobs), 1306 (research like a PhD), 1307 (Greene: the through line), 1308 (Greene: reading), 1309 (Djokovic: watch the interval), Invest 815 (MIT: markets and edge) · related: Tech #62 (agentic engineering writes the boundaries).
Key expressions
| Expression | 뜻 · 쓰이는 자리 |
|---|---|
| through line | 관통선(모든 프로젝트를 잇는 한 가지 주제) · 그린의 집중 개념. "choose one through line" |
| learning greed | 학습 욕심(시간보다 많은 배움의 욕구) · 이 항목의 이름. "Learning greed on a full calendar" |
| on a full calendar | 일정이 꽉 찬 상태에서 · 시간 부족을 말하는 관용구. "on a full calendar" |
| offload (v.) | 떠넘기다, 외주화하다(생각을 도구에 맡기다) · 응의 인지적 외주화. "what you offload to a model you do not retain" |
| cognitive offloading | 인지적 외주화(기억·사고를 외부 도구에 맡기는 것) · 학습 연구 용어. "the cognitive offloading Ng warns about" |
| mismatch of speeds | 속도의 불일치 · 입력과 이해의 속도 차이를 말할 때. "The problem is a mismatch of speeds" |
| rehearse (v.) | 되뇌다(잊지 않으려 마음속으로 반복하다) · 기억 부담을 설명할 때. "the mind stops rehearsing what it might forget" |
| bin (n.) | 통, 분류 칸 · 세 갈래 분류의 단위. "sort every wish into three bins" |
| file and release | 보관하고 놓는다 · 흥미롭지만 연결되지 않은 것의 처리. "File and release" |
| artefact | 산출물(쓰거나 만든 결과물; 영국식 철자) · 세션의 종료 조건. "a slot that ends in an artefact" |
| retention follows production | 기억은 생산을 따른다 · 만들어야 남는다는 원칙. "retention follows production" |
| scoreboard | 점수판 · 레일의 Today·Yesterday 버튼을 가리킬 때. "the scoreboard for the daily slot" |
| without an argument | 논쟁 없이(숫자가 스스로 말한다) · 정직한 지표를 설명할 때. "the number says so without an argument" |
| friction with actual material | 실제 재료와의 마찰 · 진짜 배움이 생기는 자리. "generated by friction with actual material" |
| the interval where it turned | 뒤집힌 구간(결과가 갈린 정확한 지점) · 코비→조코비치 조언. "Replay the interval where it turned" |
| raw layer | 원본 층(절대 수정하지 않는 소스 보관층) · 카파시 LLM 위키의 첫 층. "an honest raw layer" |
| LLM wiki | LLM Wiki, Large Language Model이 유지하는 위키 · 카파시의 패턴 이름. "Karpathy's LLM-wiki pattern" |
| a wish, not a plan | 계획이 아니라 소원 · 측정할 수 없는 계획을 비판할 때. "A learning plan that cannot show its ratio is a wish" |
| interview bank | 면접 은행(면접에서 꺼낼 한 문장들의 저장소) · Eng 연결. "the converse bin is the interview bank" |
| retrieval practice | 인출 연습(기억을 꺼내는 훈련이 학습을 강화한다는 연구) · 학습 과학 용어. "retrieval practice, habit cueing" |