Learn it. Then prove you
actually understand it.

A short talk about how learning works —
and a small open-source tool built around it.

I · how learning works

Rereading feels like learning.
It mostly isn’t.

Read a chapter twice and it feels familiar. But familiarity is recognition — the ability to nod along. Understanding shows up as recall: producing the idea yourself, with the book closed.

Psychologists call the gap the illusion of fluency: rereading and highlighting rate high on effort and low on effect.

I · how learning works

What works: retrieval.

Trying to pull something out of memory strengthens it more than putting it in again. Let’s do one:

I · how learning works

…and spacing.

Memory decays on a curve. Each successful recall, made just before you’d forget, flattens it — so the reviews get rarer while the memory gets stronger.

That schedule is mechanical — exactly what software is for. Spaced-repetition systems have run it since the 1970s.

I · how learning works

The new trap: AI answers feel like understanding.

A large randomized trial gave ~1,000 students GPT-4 for math practice. While the AI was there, practice went great. Then came an exam without it:

— worse than students who never had AI at all. A guardrailed tutor that wouldn’t just hand over solutions eliminated the harm. But it added no gain.

Bastani et al., “Generative AI Can Harm Learning”, PNAS 2025. Randomized controlled trial, ~1,000 high-school students.

Don’t use AI to get answers.
Use it to check you understand.

II · alix

alix — a plain-text trainer that takes this seriously.

Spaced repetition at the core, an AI layer woven through — a tutor on any card, generated decks, and an exam that gates progress on verified understanding. In your browser — an adult view, or a simpler one for kids. Local-first. Open source.

II · alix

Your cards are plain text you own.

## What does the CAP theorem state? <!-- reveal: line -->
A distributed store can guarantee only two of:
Consistency, Availability, Partition tolerance.
> The trade-off only bites while a partition is happening.

A deck is a file. Write it in any editor, version it with git, grep it, keep it for decades. No account, no database, no lock-in — and one comment-style directive can change how a card quizzes you.

II · alix

Review, the classical part — done properly.

• FSRS scheduling — the spacing curve from Act I, automated.
• Three session depths you pick per sitting — recognize, recall, reconstruct — over flip, cloze, and line-by-line cards, with both-direction and typed checks where they fit.
• New cards arrive as a low-stakes attempt first, never a cold quiz.

II · alix

Stuck on a card? Ask.

One keypress opens a tutor on the card you’re looking at — grounded in the card, its deck, and (if you allow it) the deck’s actual source. Good explanations can be saved back onto the card as a note.

II · alix

Drilling proves recall.
The exam checks understanding.

A deck can name its source — a chapter, a paper, real code. Once the cards are drilled, the AI writes fresh open questions from that source and grades your written answers against it — never against your cards. Pass, and the topics that build on this one unlock.

II · alix

Traces — drill the edges, not just the facts.

Cards hold the nodes of what you know. A trace walks a real path through a source — every step you predict what happens next, then the actual lines are revealed, and you judge your gap.

• predict → reveal → judge — the retrieval effect, applied to connections
• the frozen excerpt is the evidence; the live source stays the oracle when reachable
• the AI can discover the path for you (alix generate --trace)

II · alix

Start from your material.

• alix generate <url-or-path> — a facts deck from a lecture page, a chapter, a file, a repo — with citations back into the source.
• alix generate <dir> --goal "…" — a whole learning workspace: decks and traces, ordered by prerequisites, from one goal.
• Generated decks are a starting point — read them before you trust them.

III · honestly

What alix is not — on purpose.

III · honestly

What you should know before trying it.

• Early software, pre-1.0 — rough edges, and breaking changes still happen.
• Browser-based and local. A native Android app exists (sideload-only, no store listing yet); no sync service.
• The AI features need an AI CLI installed and signed in — Claude by default; Gemini, Codex, and Copilot work too. That usually means a paid account you already have.
• The flashcard core runs entirely without AI — offline, standalone.
• Privacy: everything stays on your machine, except the AI calls you explicitly make through your own account.

III · honestly

Where it sits.

Each neighbor does its own job well. The combination is the empty spot alix aims at — whether it fills it well is for you to judge.

III · honestly

Try it in five minutes.

Binaries for Linux, macOS, Windows — or cargo install alix. MIT/Apache-2.0, source on GitHub.

These slides are one file.
Make them yours.

alix.study · github.com/Alex6323/alix · the book: alix.study/book
Slides: alix.study/slides.html — MIT/Apache-2.0, adapt freely. Appendix follows for the curious.

A · appendix — the deck format on one slide

The full reference lives in the book.

A · appendix — the exam, precisely

• Questions are written fresh from the source:, favoring application and connections over card facts.
• Grading strictness: strict / balanced / lenient, an exam setting in your config.
• Failing produces remediation cards from your actual gaps — re-drill, re-sit.
• A deck turns exam due once every card graduates to FSRS’s review phase — no separate unlock knob to set.
• Drilling is never locked — prerequisites gate only the exam.

A · appendix — one checkpoint of a real trace

---
trace: how pressing the Good key becomes a saved grade
source: ..
---

## You press Good. What does the page send the server — and what not?
grade(g) POSTs /api/grade with { grade: g } and the session revision
echoed in a header — no card id; the server owns the seat.
<!-- at: web/alix/review/study.js:100 fingerprint: xxh64-9f21c07d3ab54e68 -->
> The page is a thin view; the server owns the session.

You answer the question before the cited lines are revealed — then judge your own gap. The next checkpoint picks up exactly where this one left off.

A · appendix — references

• Ebbinghaus (1885). Über das Gedächtnis — the forgetting curve.
• Roediger & Karpicke (2006). Test-enhanced learning. Psychological Science — the testing effect.
• Cepeda et al. (2006). Distributed practice in verbal recall tasks. Psychological Bulletin — spacing meta-analysis.
• Karpicke & Blunt (2011). Retrieval practice produces more learning than elaborative studying. Science.
• Dunlosky et al. (2013). Improving students’ learning with effective learning techniques. Psychological Science in the Public Interest — practice testing & spacing rated most effective; rereading & highlighting least.
• Bastani et al. (2025). Generative AI Can Harm Learning. PNAS — the −17 % result.

A · appendix — likely questions