alpha · invite-only
Practice that behaves like the work
Bench is a daily drill environment for people moving into AI safety
research engineering. Not flashcards, and not FAANG interview prep —
problems with the shape of real research work, a mentor that pushes
back, and a schedule that decides what you see next.
Bench is early. It is invite-only while the loop gets sharper on real
users, and access requests are read by a person.
The gap this closes
Most preparation optimises for recall on demand. Research engineering
does not work like that: you are handed an underspecified problem, you
have to find the constraints yourself, choose between approaches that
are all defensible, and then be honest about whether it worked. The
skills that decide the outcome — framing, trade-off reasoning,
debugging your own reasoning — are exactly the ones a problem set with
an answer key never exercises.
Bench structures every core problem around five phases, and the editor
stays locked until the plan has content in it. You cannot skip to
typing.
- Explore pin down the I/O, constraints, edge cases
- Brainstorm two or three approaches, with trade-offs
- Plan a paragraph-deep outline
- Implement write the code
- Verify properties, boundary cases, hand-traced inputs
For researchers
You are trying to do the work, not pass a quiz
Bench picks what you do each day from where you are actually weak,
across 13 skill axes — transformer internals, training
mechanics, eval design, numerical stability, debugging, and
interpretability and safety fundamentals alongside them.
-
A mentor that disagrees with you. Mira is short,
specific and honest, and pushes back on a weak experiment design
rather than congratulating you for having one.
-
Retrieval on a schedule, not a streak. Say a
concept back in your own words; what you miss returns in a day,
what holds returns in a week, then a month.
-
No score, ever. A miss is scheduling information,
not a grade — the app tells you what comes back sooner. There is
no percentage, no streak to break, and no line that goes down.
-
Bring your own material. Papers, lectures and
exercises you are already working through get scored on the same
axes and enter the same daily rotation.
For programs
Your curriculum, with retention attached
Cohort programmes are good at delivering material and structurally
bad at knowing what stuck. Bench runs a curriculum as a track — the
sequence you already teach — and puts a retrieval loop underneath it.
-
Tracks carry your sequence. 5
ship today, including ARENA 3.0 and MATS-oriented paths; a track is
a manifest, so yours is a file rather than a rebuild.
-
Closes the notebook gap. Working through a
notebook with autocomplete on teaches your editor, not you. Bench
asks for the same idea cold, later, away from the cell that
produced it.
-
Per-learner data stays per-learner. Row-level
security is on every table, scoped to the owner.
-
Early enough to shape. Cohort-level reporting is
not built yet. If you run a programme and want it built the way
your programme actually works, now is when that is cheap.
What ships today
- Problems
- 47
- Warmups
- 12
- Readings
- 18
- Exercises
- 52
- Tracks
- 5
- Skill axes
- 13
Written by a practitioner working through this material, not
licensed or scraped. It grows every week.
Have a look
Access is invite-only during alpha. Tell us what you are working
toward and we will get back to you.