← wasaku.appInternal system

The self-scheduling learning system

Research goes in once, current and verified. A frontier model breaks it down and writes each page, fusing it with my own record as it goes: standing hypotheses, the watchlist, my dated takes, my opinions. What comes out is a course built around what I am actually doing, and every card then carries its own review dates.

research in · verifiedkarpathy · software 3.0anthropic · agent skillsowasp · llm top tenmem0 · agent memorye2b · sandboxingmy question, mid-buildone pass · opus + fablebreaks it down, writes the pagefused from my own recordmy hypothesesmy watchlistmy dated takesmy opinionscourses/director-of-agentsbyoaiconcepts/ · each card keeps its own datesenglish-is-the-new-code.mddoanext in 1dnext in 3dnext in 8dnext in 21dnext in 55d55d · graduatedsandbox-every-agent-write.mddoanext in 2dnext in 4dnext in 11dnext in 27d27d · graduatedmem0-outlives-the-context.mddoanext in 1dnext in 2dnext in 6dnext in 17dnext in 44d44d · graduatedbyoai-needs-a-firewall.mdbyoainext in 3dnext in 7dnext in 18dnext in 47d47d · graduatedday 0day 90
One pass turns research plus my own record into cards. After that each card runs on its own clock: a settled idea drifts months out, one I just got stuck on comes back in days.

Why it is built this way

Built around what I already know

The system is customized to me, not to a syllabus. Before writing a single card it reads my knowledge base: what I already know, which concepts I use daily, where my own record says I got stuck. Then it breaks the field down to the depth I actually need, and explains each new idea through examples I already own, my own deals, my own apps, my own notes. Two courses are live right now. Their tables of contents are below, titles only, because in this system the titles are the lessons.

Director of Agents

104 cards · about 10 hours · the judgment layer for directing AI that builds software

  • M0 · The director’s stance
  • M1 · How software is put together
  • M2 · HTTP and APIs
  • M3 · Data and SQL
  • M4 · Git and GitHub, the undo button
  • M5 · The machine room
  • M6 · Shipping and cloud
  • M7 · Security for shippers
  • M8 · How LLMs actually work
  • M9 · Directing agents (the flagship)
  • M10 · Product judgment

BYOAI Prerequisites

20 cards · about 2 hours · enterprise AI security, built on same-week verified research, feeding an investment thesis

  • BYOAI is already normal, and banning it only makes it invisible
  • The user-token problem gives an agent all your power with none of your judgment
  • Every wave of unowned devices got a boundary product, from VDI to MAM
  • The enclave and the enterprise browser are the laptop era’s boundary products
  • The BYOD money went to the safety layer, and the AI slot is still empty
  • Compliance is pre-approved budget, and audits are how it gets spent
  • HIPAA and FINRA in detail, who they bind and which markets they close
  • The AI rulebooks are forming, ISO 42001, the EU AI Act, and NIST AI RMF
  • OAuth replaced passwords with permission slips, and agents need the same moment
  • Agent identity arrived with Entra Agent ID, Cross App Access, and Auth for GenAI
  • Non-human identity consolidated in one quarter, and adoption still lags
  • DLP watches data leaving, and the prompt box is its newest exit
  • Zenity watches what agents do, and Cyera knows where the valuables are
  • The employer side is funded and consolidating while the worker side is one waitlist
  • RAG lets AI borrow knowledge at answer time, which makes access the product
  • A TEE is a locked room with a window onto proof
  • FHE computes on sealed data, and its real use is the small decision kernel
  • Federated learning ships the lesson and keeps the diary
  • The guards of one workday, the whole market in a single story
  • The two-sided thesis and the empty worker seat, said for an interview