What a safety cohort learns, in plain words

What is the cohort actually learning — and can someone outside the program understand it in one sitting?

The owner of this site is in TARA — Technical Alignment Research Accelerator, cohort 2 (Sep–Dec 2026), a part-time program built on the ARENA 3.0 curriculum. This project is the companion to it: each week, the one idea the cohort's session introduced, rebuilt as a standalone interactive page. No code, no math, no program-internal references — every page passes the same test: could a curious non-technical reader finish it? The curriculum guide with the exercises lives in the sibling project, the ARENA guide.

The weeks, in order

WEEK 1 · SEP 13 · DRAFT

A computer has never seen a cat — only numbers

pictures are grids of numbers · the walking window · edges into parts into objects · the shortcut that made depth trainable

WEEK 2 · SEP 20 · DRAFT

A word becomes a place, and attention does the reading

the space of meaning · the running notebook · the attention explorer · the no-peeking rule

WEEK 3 · SEP 27 · DRAFT

The model doesn't know the next word — it bets on it

scores become shares of 100% · the temperature dial · a training loop small enough to watch think

WEEK 4 · OCT 4 · DRAFT

You can read a model's internals with one straight line

the probe sandbox — draw the line yourself · where the difference lives · from truth to deception

The first page, whole

The week-2 page, embedded as it ships — everything runs offline in one file.

Artifacts

• charts/*.html — four single-file interactive pages: attention explorer, probe sandbox, feature-hierarchy explorer, next-word picker. Hand-set, seeded data only; two loads identical.
• posts/*.md — the thin frames: claim title, lede, the embed, sources.
• data/raw/README.md — provenance for everything the pages quote, pinned to vault commits.

Weeks land as the program's sessions happen. The personal record — hours, audits, receipts — stays in the program vault; these pages carry only the ideas.