A whole fly brain you can backprop through
45,661 trainable parameters over 163,903 neurons. The looming pathway responds 81.5x above shuffled wiring.
warpfield
Warpfield is an independent research lab building learning systems that have to answer to the physical world.
Most of our work starts from one question: what happens when a model has to act inside something real, a connectome, a molecule, a body in a physics engine, instead of predicting text about it?
Text is a record of what people already worked out. Physics, chemistry and biology still hold results nobody has written down, and the only way to get them is to run the experiment, in a lab or in a simulator faithful enough to disagree with you.
We work small: rented GPUs, open datasets, and a fleet of models that plan, write and check code alongside us. What keeps that honest is discipline about evidence. Every claim ships with the run that produced it, and every number is measured against a control.
This site is where the projects live. A few are finished. Most are in motion.
Projects
-
01
Connectome ping pong Neuroscience
Two fruit flies wired from the full male CNS connectome, 164k neurons and 6.2M synapses, placed in a MuJoCo arena and trained end to end to rally a ball. The whole brain is differentiable; only synapse gains, time constants and biases learn.
54-hit rally
-
02
sysone-vl Robotics
A fast decision model that looks at a scene and answers typed questions with calibrated probabilities instead of generated text. Qwen2.5-VL backbone, slot heads, and reinforcement learning on calibration.
Grasp calibration error 0.21 to 0.06
-
03
Reentry Materials
Watching a graphite heat shield come apart atom by atom. A 432-atom slab ramped from 300 K to 10,000 K under a machine-learned interatomic potential, rendered as it unzips into carbon chains.
Oxygen chemistry fails on this checkpoint, reported as such
-
04
termdesk Agents
A real Linux desktop inside the terminal, with a control socket and an indexed accessibility tree so any model can drive it by element instead of by pixel. Disposable sandboxes, recording built in.
How we work
Measure above a baseline. A response only counts once it clears a shuffled or stimulus-free control. Twice now a confident number turned out to be a comparison against the wrong zero.
Build the checker before the search. If a result cannot be replayed from a seed and a config, it is a lead, not a result.
Keep the negatives. A potential that cannot do oxidation chemistry is worth writing down; it saves the next person the same month.