PYTHON · NUMPY

Agency and Action Efficiency

Implement the action-efficiency score that measures goal-directed intelligence, and see why brute force scores near zero while an efficient, guided agent scores high.

01Challenge

Brute-forcing a game by trying thousands of actions can eventually win, but a benchmark that squares your inefficiency drives that score to nearly zero. Implement the ARC-AGI-3 efficiency score and show why a guided, curious agent beats brute force.

02Model

The frontier of embodied intelligence is not just doing a task; it is acquiring the goal and realizing it efficiently. ARC-AGI-3 makes that measurable by scoring the SQUARE of how your action count compares to a human's, so being twice as wasteful costs three-quarters of your score and brute-forcing the thousands-wide action space scores essentially nothing. This is why raw frontier models score near one percent on the benchmark, and why a good harness that supplies general thinking patterns lifts them toward a third: intelligence, in this framing, is respecting the constraints of a problem, not searching harder. You implement the metric and confirm what it rewards.

03Guided practice

1. The metric. Efficiency is the SQUARED ratio, capped at 1: min(1.0, (human/actions)**2). 2. Fix score. Square the ratio. 3. Check the shape. score(20, 40) must equal 0.25 (twice the actions, quarter the score). 4. Confirm. The guided agent scores above 0.5; brute force scores below 0.01.

04Feedback

PASS WHEN PASS when your squared score gives 0.25 for double the actions, above 0.5 for the guided agent, and near zero for brute force.

If your run fails, check:

Score for 40 actions is 0.5, not 0.25: you have not squared the ratio yet. Return `min(1.0, (human/actions)**2)`.
Guided score too low: with the square, 20/26 gives about 0.59, which clears 0.5; check the exponent.
Brute not near zero: 20/4000 squared is tiny; make sure you divide human by actions, not the reverse.

05Retrieve & space

How does ARC-AGI-3 score action efficiency, and why square it? (squared ratio; punish inefficiency adversarially).
Why does a good harness raise a model's score without being game-specific? (it supplies general thinking constraints).
In this framing, what is intelligence? (respecting and acquiring a problem's constraints, not searching harder).

06Mastery & project

Implement a goal-efficiency metric and explain why efficient, constrained action, not brute force, is the measure of embodied intelligence.

The evaluation lens for any agent that has to acquire a goal and realize it well, the standard the play half holds itself to.

← Curiosity: Learning Before Rewards Where the Robot Earns Its Place →

PYTHON · NUMPY

Agency and Action Efficiency

Implement the ARC-AGI-3 action-efficiency score and compare brute force with a guided agent.

You read

the arrays and values already in scope

You change

the code you write in each cell

Fixed

the dataset and the checks that grade you