# Generalised Superintelligence > Generalised Superintelligence (GSI) is a research lab building superintelligence for scientific discovery. Its system finds the structure in a hard problem, compiles a solver for that structure, and accepts a result only when multiple independent verifiers re-derive it. Each discovery leaves tools behind that make the next one cheaper. The aim is to make discovery a hundred times faster in every field where an answer can be checked. Generalised Superintelligence (GSI) was founded in 2026 by Kishan Kumar (Chief Scientist) and Yashraj Shukla (CEO). It works from Bengaluru and San Francisco. Its research log labels each result by how it was checked, including results that rest on one engine or whose novelty is not yet established. ## Pages - [Vision](https://www.generalised.si/vision) - [Research](https://www.generalised.si/research) - [Approach](https://www.generalised.si/approach) - [Contact](https://www.generalised.si/contact) - [Full text for language models](https://www.generalised.si/llms-full.txt) ## Selected results - No cyclic (465, 145, 45) difference set exists; it was the smallest open case in the La Jolla repository and a case of Ryser’s conjecture. - IBM’s two-gross [[288, 12, 18]] memory circuit has exact circuit distance 16 (published: 18); the matching lower bound is GSI’s, confirmed by multiple independent code bases. - CW(120, 36) exists, contradicting Theorem 4.24 of Tan (2018). - More than 200 Hadamard, skew-Hadamard and symmetric Hadamard matrices at orders listed as open. - Three new terms of OEIS A003111 (complete mappings of Z_27, Z_29, Z_31). - 860 exact minimum distances for 5G NR LDPC codes. - [All results with evidence](https://www.generalised.si/research) ## Papers and notes - [GSI-R-0001: We construct Hadamard matrices of 34 orders for which Table 4 of Cati and Pasechnik lists no known matrix](https://www.generalised.si/research/hadamard-matrices) - [GSI-R-0002: We construct skew-Hadamard matrices of order 4v for 106 odd v < 2500, and symmetric ones for 85, each listed as undecided in the 2025 census](https://www.generalised.si/research/skew-symmetric-hadamard) - [GSI-R-0003: IBM’s two-gross [[288, 12, 18]] memory circuit has circuit distance exactly 16, below the published upper bound of 18](https://www.generalised.si/research/two-gross-circuit-distance) - [GSI-R-0004: No cyclic (465, 145, 45) difference set exists, and so none exists in any abelian group of order 465](https://www.generalised.si/research/cyclic-difference-set-465) - [GSI-R-0005: Proper circulant weighing matrices exist for CW(120, 36), which Tan (2018) excludes; CW(180, 36), listed “Open”; and 209 pairs listed “No” in La Jolla version 1.3](https://www.generalised.si/research/circulant-weighing-matrices) - [GSI-R-0006: Circulant good matrices of orders 53 and 55 exist; one explicit set of each is given](https://www.generalised.si/research/good-matrices-53-55) - [GSI-R-0007: An exhaustive computer search finds exactly one inequivalent set of circulant Williamson matrices of order 75 and no circulant good matrices of order 75](https://www.generalised.si/research/williamson-good-order-75) - [GSI-R-0008: We computed that ℤ_27, ℤ_29 and ℤ_31 have 164,135,223,171,169,785; 18,674,231,142,723,249,003; and 2,427,915,096,775,283,929,115 complete mappings that fix 0, values that would extend OEIS A003111 to a(15)](https://www.generalised.si/research/complete-mappings-cyclic-groups) - [GSI-R-0009: We computed the numbers of inequivalent binary self-dual codes of even length from 42 to 56 and of inequivalent stabilizer states on 13 to 20 qubits](https://www.generalised.si/research/self-dual-codes-stabilizer-states) - [GSI-R-0010: In Table 4 of arXiv:2607.04716v1 on 5G NR LDPC sub-codes, d(C′(E[4], 48)) is 5, not 6, and we compute d(C(E[4], 192)) = 10, not 8](https://www.generalised.si/research/5g-nr-ldpc-minimum-distances) - [GSI-R-0011: Two non-interleaved schedules and the period-11 schedule of G. Pan (to 40 cycles) give [[288, 12, 18]] memory circuits of circuit distance 18](https://www.generalised.si/research/bivariate-bicycle-circuit-distances) - [GSI-R-0012: No cyclic relative difference set with nontrivial kernel lifts a cyclic (364, 121, 40), (255, 127, 63), (1464, 133, 12) or (2380, 183, 14) difference set](https://www.generalised.si/research/singer-relative-difference-sets) - [GSI-R-0013: Eleven SAT Competition 2025 main-track formulas listed as ‘unknown’ in the Global Benchmark Database have checkable refutation certificates](https://www.generalised.si/research/sat-competition-refutations) - [GSI-R-0014: We enumerate minimal cut sets of the 95-reaction E. coli core network up to size 4, 87–215 times faster than our mixed-integer baseline in single runs](https://www.generalised.si/research/metabolic-minimal-cut-sets) - [GSI-R-0015: We find exact ground states of 24 compact Ag–Pd pair-model cells with 8.5 to over 7,000 times less CPU time than Huang et al.’s cell solver](https://www.generalised.si/research/cluster-expansion-ground-states) - [GSI-R-0016: Our symmetry-reduced transfer-matrix program took 1.6 times as long as our engine to compute the periodic 4 × 4 × 4 Ising density of states exactly](https://www.generalised.si/research/ising-3d-density-of-states) - [GSI-R-0017: Five explicit binary linear codes of length 111 to 174 have minimum distance above the lower bounds listed at codetables.de on 16 November 2026](https://www.generalised.si/research/candidate-binary-codes) - [GSI-R-0018: The lab’s solver system gave the right class for 5 of 8 blind in-class models and fired on none of 3 negatives](https://www.generalised.si/research/solver-compiler-and-factory) - [A sharp ceiling on T_c in linearised Eliashberg theory](https://www.generalised.si/research/superconducting-tc-ceiling): In linearised Eliashberg theory without Coulomb terms, T_c < 0.18273 √(λ⟨ω²⟩) for all spectra, with a sharp constant. Kondratev posted a proof on 19 September 2026. We add sharp versions for several bands, anisotropy and a static Coulomb term in stated cases. Some Coulomb kernels break the bound. - [Published electron-phonon data against the Eliashberg ceiling](https://www.generalised.si/research/tc-ceilings-published-data): None of 23,608 computed Tc values in a zero-pressure database, or of 201 reported Tc values, is shown to exceed the Eliashberg ceiling ω₂f(λ). In that database no ceiling from reported moments reaches 300 K. The ceiling alone does not exclude 300 K. - [Lower bounds for jellium, and the limit of smearing bounds](https://www.generalised.si/research/jellium-lower-bounds): A computer-assisted proof that classical three-dimensional jellium has energy per particle at least −1.448178733, a proved ceiling of −1.44800699488 for smearing bounds, a two-dimensional bound within 6.4×10⁻²⁹ of the triangular lattice, and a Lieb–Oxford certificate, which Ye Li's improve on. - [Ground states of the 2D dipolar Ising model and of dipolar spin ice](https://www.generalised.si/research/dipolar-magnet-ground-states): We prove which phase is a ground state of the 2D dipolar Ising model for every coupling in [0, 3.95] except a window of width 2.8×10^-6, and that the MHG state is the unique translation-invariant ground state of tin-foil Ewald dipolar spin ice with nearest-neighbour exchange only for J1/D ≤ 9/2. - [Certified ground states of long-range lattice gases](https://www.generalised.si/research/lattice-gas-ground-states): On the triangular lattice, five crystals are ground states of long-range lattice gases among all configurations, at 82 potentials. The method also gives certified windows on three other lattices and, among periodic configurations, Na/vacancy orderings of an in-plane sodium cobaltate model. - [Certified global optima for periodic ion allocation](https://www.generalised.si/research/certified-crystal-structure-prediction): Exact certificates prove the global minimum of a classical ion-allocation energy for 41 periodic instances and for bixbyite Y2O3, with no symmetry restriction. Spinel, pyrochlore, garnet and a 135-ion perovskite on 512 sites stay open, with bounds 5.6 to 90.3 eV below the best allocation we hold. - [Perfect quasirandom alloy cells and certified diffusion brackets](https://www.generalised.si/research/perfect-sqs-and-diffusion): Smallest supercells whose pair correlations equal those of a random equiatomic alloy, with complete catalogues for quaternary fcc and bcc, and certified two-sided bounds on the variational diffusion coefficient of an exclusion-process tracer and of an alloy vacancy. - [Log-concave turnover frequency and a dynamic cap for oxygen evolution](https://www.generalised.si/research/catalysis-limits): For a catalytic cycle of Eyring steps the turnover frequency is log-concave in the state free energies, so with convex barrier laws a volcano top is a certified maximum. For oxygen evolution at 1.5 V with floor barriers and OOH fixed 3.2 eV above OH, open-loop forcing gains at most 2.6945 decades. - [What circuit distance buys in bivariate-bicycle syndrome cycles](https://www.generalised.si/research/syndrome-cycles-bivariate-bicycle-codes): A depth-8 cycle for the 108-qubit bivariate-bicycle code keeps circuit distance 10 and lowers the logical error rate to 0.76 of IBM's at p = 0.0025 (SD6, BP-OSD). For the gross code, distance 11 at IBM's fault count gives no detectable gain at p = 0.004 and 0.005; the 756-qubit code has distance 34. - [Certified pointwise EUV failure probabilities and a photon dose floor](https://www.generalised.si/research/euv-stochastic-failure-certificates): In a stated Poisson shot-noise model of EUV contact printing, two pointwise failure probabilities are bracketed within a factor 1.07 to 1.24 from 10⁻¹² to 10⁻⁹, and a Neyman–Pearson dose floor for local resists is evaluated. The model is not calibrated to any process. - [Steep-switch limits and minimum energy per switching event](https://www.generalised.si/research/steep-switch-limits): For a ballistic source whose mode density falls no faster than an Urbach tail, we prove a closed-form bound on the on/off ratio and find its exact extremiser. The minimum supply energy per switching event reduces to the class's exact on/off frontier; the cold-source gain depends on two floors. - [Certified wirelength lower bounds for the ISPD 2005 benchmarks](https://www.generalised.si/research/placement-lower-bounds): In orientation N and without overlaps, every placement of the eight ISPD 2005 netlists has half-perimeter wirelength at least 0.338 to 0.604 of the lowest value printed in one 2025 preprint. A spreading term gives 0.556 and 0.448 on two designs. The optimum's distance from the bounds is unknown. - [Graphs two below the Moore bound, and the median of nine inputs](https://www.generalised.si/research/moore-bound-defect-two): No graph of maximum degree d, diameter 2 and d²−1 vertices exists for 6 ≤ d ≤ 3000, and none of diameter 3 and d³−d²+d−1 vertices for 4 ≤ d ≤ 150, by computer-assisted proof. A comparator network for the median of nine inputs needs exactly 19 comparators. ## Contact - Kishan Kumar, Co-founder, Chief Scientist: https://www.linkedin.com/in/kishan-kumar03 - Yashraj Shukla, Co-founder, CEO: https://www.linkedin.com/in/whynesspower - Schedule a call: https://cal.com/kishan-kumar-kgfq7w+whynesspower - LinkedIn: https://www.linkedin.com/company/generalised-superintelligence - X: https://x.com/GeneralizedSI