01AI & Computation
Circuit Diffing as a Release Gate
Treat a model update like a code change: diff the circuits, not just the benchmark scores.
Read →
A working idea book · not peer reviewed
Each entry starts from a result published in the last eighteen months, names a gap that result opens, and proposes a falsifiable experiment. No idea here is a summary. All of them are arguments, and each one says what would kill it.
Showing all 24 ideas
Treat a model update like a code change: diff the circuits, not just the benchmark scores.
Before a cell world model nominates a target, make it predict the trials that already failed.
Reasoning and measuring are both purchases. Nobody has priced them against each other.
Lipid nanoparticles are aimed at organs. Isoform-resolved spatial atlases let them be aimed at cell states.
Partial reprogramming is dosed open-loop. The clock that measures its effect should also control it.
If a vaccine can be sequenced, it can be released by sequencing. Hospital-scale manufacturing needs a release test that is not a six-week wait.
Stop hunting for a phage that matches the bug. Design a pair whose resistance costs the bug something.
A complete fly nervous system is now a wiring diagram. Wiring diagrams are screening instruments if you simulate them.
An electrode that decodes speech all day can also tell you, months early, that cognition is slipping.
Cortical organoids learn without backpropagation. That makes them the only physical test of which learning rule actually scales.
Focused ultrasound can reach deep structures without surgery. Gate it to the sleep spindle and it can write, not just nudge.
A magnet with no net magnetisation cannot leak field into its neighbour. That makes it a candidate substrate for dense analog memory.
A lithium filament concentrates current before it shorts. Diamond magnetometry can see the field it makes.
Water harvesting and direct air capture both fight humidity with the same sorbents. Run them on one cycle.
Every grid buffer will be driven by a pulsed load. The cycling literature is almost entirely smooth current.
Enhanced weathering is verified by models and flux towers. A buried bag of known mineral is a cheaper, more honest measurement.
Fusion's first customer does not need a turbine. It needs 600 degrees and does not care about Q.
Surface codes protect computation. Run one in sensor mode and coherence grows with code distance.
Solvation shells are a quantum many-body problem nobody simulates well. Rydberg arrays are built for exactly that shape.
MRI reconstruction is a fixed linear inverse problem. That is exactly what an optical processor does cheaply.
Permanently shadowed craters sit near 40 kelvin and never see the sun. That is a biorepository with no power supply.
The circulation is monitored by a handful of expensive moorings. Telecom repeaters already sit on the seafloor.
Humanoids are entering factories to work. The valuable output may be the trajectory log, not the parts.
Tactile models predict relative feel. Surgery needs absolute newtons, and nothing calibrates them.
What changed is the published work, cited. The gap is what that work leaves undone. The idea is the proposal. The decisive test is the cheapest experiment that would settle it, with an explicit result that kills it. If you only read one section, read the test.
Nothing here has been peer reviewed. The sources are real and linked; the synthesis and the proposed experiments are the author's. Treat every entry as a hypothesis with a plan attached, not as a finding.
Built 30 October 2026. Standalone HTML, no external assets. WordPress fragments for every idea are in
wordpress/.
Publication records: source hashes · original README · original manifest · additive publication note