01
5 environments · ongoing
System One / System Two
Can a tiny network, trained by a slow planner, make most decisions at a fraction of the planner's cost,
and know when to hand control back?
A fast network (System One) proposes every move with a confidence. When it is unsure, or when a cheap
safety check rejects its move, a slow planner (System Two) decides instead, and that decision becomes a
new training example. The same pipeline runs on five very different problems.
- Snake With a cheap safety check, the hybrid reaches 98% of the planner's score for 8.8× less compute, just short of the 10× target.
- Lander A hand-written autopilot beats the network at the same cost: sometimes a rule is enough.
- Warehouse 16 robots: with less compute than the planner's cheaper setting, the hybrid delivers 89.4 loads against 60.1; the 10× target is missed.
- Racing The network alone drives 99.9% as far as the planner, for 1,164× less compute.
- Quadruped A 3D walking robot you can push. With its safety check, the network walks 91% as far as the planner for 82× less compute, while agreeing with it on only 41% of decisions.
Open the experiment →
02
complete
Handwriting pad
Can a network with 22,298 parameters read handwritten letters as well as a classic recogniser that compares
your letter with every stored example?
Write one uppercase letter at a time with your finger, like on a car's touchpad. Both recognisers read it in
your browser, and the page shows how much arithmetic each one spent.
- Result On 20 writers it never saw, the network reads 83.9% of letters right, the classic recogniser 83.2%, for 1,302× less arithmetic. The 90% target is missed.
Open the experiment →
03
applied · in progress
Go1 in your browser
Can a model of a real robot, trained on a GPU, walk in a browser tab, and can a planner teach it where
it struggles?
The lab's applied side: established tools (MuJoCo, MuJoCo Playground, a free Colab GPU) instead of code
written from scratch, under the same method. The Unitree Go1 runs in your browser, and you steer it.
- So far Trained in 84 minutes on a Colab T4, the policy walks on rough ground without falling, and runs in the browser exactly as it did in training. A first GPU planner never improved on it.
Open the experiment →