Research theses
01

The polymath becomes operational again

AI lowers the cost of entering new fields, holding provisional maps, and testing bridges across domains.

Read the essay at Yuki Capital
02

Production becomes cheap; verification becomes central

As agents write code, proofs, experiments, and papers, the human bottleneck moves toward specification, audit, judgment, and responsibility.

03

Research needs new trust infrastructure

The hard question is how to know which AI-generated discoveries are interesting, correct, reproducible, novel, and worth attention.

How I work
Explore widelyUse models to map literatures, compare frameworks, draft code, and surface cross-domain analogies worth testing.
Verify narrowlyPush promising claims through reproducible checks, tests, solver traces, formal tools, or expert review before treating them as real.
Keep memoryArchive useful failures, contradictions, source trails, and negative evidence so later work compounds instead of resetting.
Current areas
Autonomous research systemsEval-driven lab loops, candidate archives, benchmark adapters, and hardening protocols for AI-assisted discovery.
Cosmic topology and CMB structureCompact-universe models, low-ell CMB correlations, and falsifiable topology tests.
Holography and black-hole informationIsland formulas, Page curves, and entanglement bounds on compact spaces.
Casimir energy and dark-energy testsVacuum-energy calculations, holographic dark energy, and public cosmology data.
Extremal combinatoricsFrankl-type union-closed set problems, entropy barriers, and proof reductions.
Scattering geometry and entanglementAmplituhedron, celestial holography, and positive-geometry bridges.
Formal verification pipelinesLean, PB certificates, solver traces, reproducibility, and expert review gates.

More research at yukicapital.com/research