EBM lab

A playground for learning energy-based models deeply, by building Ahmed's idea: a growing network of small energy-based models whose job is to make connections, that caches what it did during the day and retrains during sleep, generation after generation. See IDEA.md.

Started 2026-10-13. Separate from Genesis: ~/genesis-archive is a read-only research library we draw knowledge from. No Genesis code is carried over.

What this is, in plain words

A tiny creature lives in a 48 x 48 world full of smells. It has two antennae, a body that gets hungry, and a brain made of a small energy-based model. Nobody tells it what food is. The only teacher is its own body: eating makes it feel better, starving and poison make it feel worse. The question of the lab is simple: can a brain like this learn, within one life, to find food by smell?

As of EXP-02, the answer is yes.

The world and the senses

• Food and poison leak three chemical "notes" (A, B, C) that spread and fade. Food smells of A + B, poison of B + C. So no single note tells them apart.
• Each tick the creature reads the six numbers under its two antenna tips, and turns them into 12 inputs: which side smells stronger, how strong overall, what changed since last tick, and how its body feels.
• It can do three things: turn left, turn right, or go forward.

The brain

Every possible move gets an energy, F(x, a). Lower energy means "this move fits what I sense right now". The creature picks the lowest-energy move 90% of the time, and a random one the other 10%, so it keeps exploring.

Learning means reshaping that energy. The rule that worked (arm ebm-td) treats -F as a guess of how good each move is, and nudges it toward what actually happened: every tick while awake, and again at night by replaying what it did during the day. When a day surprises it, it grows a new hidden unit.

Watch it live

The run viewer replays a whole life from the log: the world, the body's energy and toxin, and the brain's energy for each move, tick by tick. This is day 10 of seed 203:

And the same day, from the teaching video:

The result

20 creatures, 15 days each (5250 ticks is the most a creature can live):

• The value-learning brain (ebm-td) lives 4113 ticks on average in the linear world and 3876 in the xor world. 15 of 20 creatures survive the whole 15 days.
• It reaches the level of tabular Q-learning, the classic textbook baseline. It does not clearly beat it.

Did it really learn?

Take one creature's brain from day 5, freeze it, and drop it into 15 new worlds:

• the day-5 brain survives 15 / 15
• the same brain with its antennae swapped left for right survives 0 / 15
• the brain it was born with survives 2 / 15

So the skill lives in the weights, and it depends on knowing left from right.

What it does not do (yet)

• It does not avoid poison. The survivors simply eat more food.
• The hidden units are not what makes it work. The xor world does not yet force the creature to combine notes.
• The full list, with every correction kept on record, is in findings/02.

How the lab works

Every experiment is pre-registered (written and committed before it runs). Every conclusion gets two hostile referees whose job is to break it. When a claim turns out wrong, the correction is added and the old claim stays visible (see findings/01 and research/07).

Try it yourself

uv sync
uv run pytest                              # the 83 laws the simulation must obey
uv run python weights/try_it.py            # drop the trained brain into a new world
uv run python -m http.server 8765          # then open the viewer link it prints

The trained weights and an honest table of how well they do in new worlds are in weights/.

A full teaching video (lesson 01, about an hour, every number in the brain explained) is coming.

Layout

Path	What
IDEA.md	the idea, verbatim, and its parts
RESEARCH-PLAN.md	the goal, hypotheses H1–H5, phases (v0.1; the creature-first reorder is in session 02's journal)
DESIGN.md + DESIGN-AMENDMENTS.md	creature v0's spec, and every change to it with its reason
creature/	the code: physics, world, body, senses, brain (EBM), baselines, sim, run, export
tests/	the laws (uv run pytest)
experiments/	harness, calibration, pilots, pre-registrations, results/ (json + png)
viewer/	the run viewer (see its README)
runs/	exported lives to watch
docs/media/	the pictures and clips in this README
weights/	the trained brains (final weights of seed 203, both worlds) + try_it.py to drop one into a new world
CLAUDE.md	ground rules for any agent working here
STATE.md	where the lab is right now (updated every session)
HANDOFF.md	what the next session does first
research/genesis-digest/	what Genesis learned, distilled and cited (A-E + 00-SYNTHESIS.md)
research/	research seats R1–R4 (03–06), the proof-of-possibility note 07 with its scripts
experiments/	one folder per experiment, each with its own question, code, and results
decisions/	numbered decision records
findings/	durable results, each with the command that re-derives it (01 negative, 02 the first positive learning result)
sessions/	one journal per session