AI Discovery Ledger

OpenAI · OpenAI internal agent system

Reported as a resolution of one of the seven Millennium Prize Problems: a configuration in which a vortex tightens and spins faster while energy stays bounded (finite-time blowup). OpenAI said it does not intend to claim the prize. Mathematicians had not yet inspected the argument at announcement.

OpenAI · Astra

A batch of ten results across geometry, cryptography, and complexity from one agentic system, cited in the AI Discovery Ledger spec as the low-cost baseline (about $2,000 total, vendor stated) for the cost-per-breakthrough curve.

OpenAI · Astra

Reported as the first explicit construction of a non-sofic group, closing a question in group theory open since Mikhail Gromov introduced the concept of soficity in 1999 (27 years). One of ten results OpenAI attributed to its unreleased Astra model on 2026-08-01; this specific result is the one named in a subsequent research-misconduct allegation over citation of prior work. No source describes Astra's method as an agent swarm (unlike the separate, explicitly 10,000-agent Navier-Stokes effort), so approach is left null rather than inherited from the parent seed.

OpenAI · Astra

One of ten results OpenAI attributed to its unreleased Astra model on 2026-08-01; closes Erdos problem 183 in his catalogue of open problems. No citation or credit dispute found specific to this result (unlike the non-sofic group and a sphere-packing result from the same batch). No source describes Astra's method as an agent swarm, so approach is left null.

Google DeepMind · AlphaProof Nexus

Reported to have autonomously resolved 9 of 353 open problems it attempted from Thomas Bloom's Erdos Problems database, including two open for 56 years, and proved 44 of 492 open OEIS conjectures, at an inference cost reported qualitatively as 'a few hundred dollars per problem' (no single total figure given, so not recorded as a computeCostUsd figure). The arXiv paper itself (fetched directly) uses Lean-based verification but does not, in the text retrieved, name the system 'AlphaProof Nexus' or the model 'Gemini 3.1 Pro'; both of those specifics come from The Decoder's reporting, which is used as this row's source. validationStatus is kept at claimed_only rather than promoted to formally_verified: the paper states Lean was used, but no source read directly states an independent check or confirms the 'zero sorry' claim that circulates in secondary press, consistent with how the Navier-Stokes seed treats a Lean formalization that nobody outside the lab has yet confirmed.

OpenAI · OpenAI internal reasoning model

OpenAI reported that an internal reasoning model disproved Paul Erdos's 1946 conjectured growth rate for the planar unit distance problem, producing point configurations via algebraic number theory (class field towers) that exceed the conjectured bound. Nine mathematicians, including Fields medalist W. T. Gowers, co-authored a companion arXiv paper giving a human-verified, digested version of the AI-generated counterexample. The result narrows but does not close the gap to the known upper bound (about n^1.333). openai.com returned HTTP 403 on direct fetch, so OpenAI's own page was never read directly; the arXiv companion paper and secondary press stand in as source.

Google DeepMind · Aletheia

Answers a question of Erdos and Graham: shows that a double-exponential growth condition on a strictly increasing integer sequence is sufficient for the series of reciprocal products to have an irrational sum. Per the paper's own abstract, the original problem was solved autonomously by DeepMind's internally codenamed agent Aletheia (running on Gemini Deep Think); related generalizations in the same paper were produced through human-AI interaction, which is the basis for describing this as mostly-autonomous rather than human-directed, though no numeric AI-contribution share is given by any source.

Baker Lab, Institute for Protein Design, University of Washington · RFdiffusion (antibody fine-tune)

A fine-tuned RFdiffusion network combined with yeast display screening generated de novo VHHs, scFvs, and full antibodies binding user-specified epitopes with atomic-level precision. Cryo-EM confirmed the binding pose of designed VHHs against influenza haemagglutinin and Clostridium difficile toxin B, with a high-resolution structure verifying atomic accuracy of the designed CDRs. Initial computational designs had only modest affinity (tens to hundreds of nanomolar Kd); affinity maturation using OrthoRep produced single-digit nanomolar binders retaining the intended epitope selectivity. Separately reported success rates for the underlying design method are low (0 to 2 percent across targets per press coverage), so this is a discrete set of validated hits rather than a routine design method.

Broad Institute of MIT and Harvard, with the Collins Lab, Massachusetts Institute of Technology (Krishnan, Anahtar, Valeri et al.; senior author James J. Collins) · fragment-based and unconstrained de novo antibiotic generation framework

A generative AI framework designed de novo antibiotic candidates via a fragment-based method (screening more than 10 million chemical fragments in silico) and unconstrained de novo compound generation, both using genetic algorithms and variational autoencoders. Of 24 synthesized compounds, seven showed selective antibacterial activity. Two lead compounds, reported by name as NG1 and DN1 in Karolinska Institutet's coverage, showed bactericidal efficacy against multidrug-resistant isolates and reduced bacterial burden in vivo in mouse models of N. gonorrhoeae vaginal infection and methicillin-resistant S. aureus skin infection.

Google DeepMind · AlphaEvolve

AlphaEvolve found a procedure multiplying two 4x4 complex-valued matrices with 48 scalar multiplications, one fewer than Strassen's 1969 recursive algorithm (49), described by DeepMind as the first improvement in this (recursively-applicable, noncommutative) setting in 56 years. The novelty of that '56 years' framing is publicly disputed; see creditDispute.

Google DeepMind · AlphaEvolve

AlphaEvolve found a configuration of 593 non-overlapping unit spheres touching a central sphere in 11 dimensions, improving the known lower bound on the kissing number in that dimension, on a problem where DeepMind says there had been no progress for two decades in dimensions below 16.

authors of the Advanced Science 2025 study, in collaboration with Google's AI co-scientist team · AI co-scientist

A multi-agent AI co-scientist system was used to recommend candidate anti-fibrotic drugs targeting epigenomic modifiers. Efficacy and toxicity of 25 drugs (3 previously known anti-fibrotics plus AI co-scientist recommendations) were tested in multi-lineage human hepatic organoids grown in microwells (microHOs). One AI-recommended, FDA-approved anti-cancer drug (Vorinostat) reduced TGF-beta-induced chromatin structural changes by 91% and promoted liver parenchymal cell regeneration. Wet-lab validation was performed in organoids, not in animals or humans; no clinical trial is reported.

Jose R. Penades lab (Imperial College London), with Google DeepMind's AI co-scientist team · AI co-scientist

The Penades lab posed the AI co-scientist a question about cf-PICI host-range expansion that their own team had already resolved experimentally over years of work but had not yet published. The system's top-ranked hypothesis matched the lab's own confirmed, unpublished mechanism (cf-PICIs hijack diverse phage tails to expand host range) within days. This is parallel recapitulation of an already wet-lab-validated but unpublished finding, benchmarked by the discovering lab itself, not an independent new discovery; the underlying wet-lab validation predates the AI run and is attributed to the human team by the paper's own abstract.

Microsoft Research, with the Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences · MatterGen

MatterGen is a diffusion-based generative model trained on roughly 608,000 stable materials from the Materials Project and Alexandria databases that designs novel crystal structures conditioned on target properties. As an experimental validation, Microsoft and SIAT synthesized a MatterGen-generated structure, TaCr2O6, conditioned on a target bulk modulus of 200 GPa; the synthesized material's structure matched the proposed one, with compositional disorder between Ta and Cr, and had a measured bulk modulus of 169 GPa against the 200 GPa design target, a relative error below 20 percent.

Mario Krenn's group (Max Planck Institute for the Science of Light / University of Tubingen), with Xiao-Song Ma's group at Nanjing University · PyTheus

PyTheus, an AI-driven quantum-optics design tool that represents candidate experiments as graphs, produced a configuration for entangling independent photons that did not require pre-existing entanglement or Bell-state measurements, contrary to the standard entanglement-swapping recipe it was originally asked to reproduce. A team at Nanjing University led by Xiao-Song Ma built the experiment and confirmed it worked as designed; the result was published in Physical Review Letters.

Princeton Plasma Physics Laboratory / Princeton University, with the DIII-D National Fusion Facility · deep reinforcement learning tearing-instability controller

Researchers designed an AI/deep-reinforcement-learning model to reconstruct kinetic plasma equilibria and adjust magnetic confinement fields in real time, then tested the controller on the DIII-D tokamak, integrating data from many sensors to maintain plasma stability under conditions including the ITER baseline scenario, where tearing instabilities usually arise.

Google DeepMind · FunSearch

FunSearch paired a pretrained code LLM (PaLM 2) with an evolutionary loop and an automated correctness evaluator, and found cap set constructions larger than any previously known, described by DeepMind as the largest increase in known cap set sizes in 20 years.

Google DeepMind · FunSearch

Same FunSearch run (same paper, same announcement) also produced a new online bin-packing heuristic that used fewer bins than established heuristics on the distributions tested. Coded as a separate row from the cap set result because they are two distinct named problems from the same system and paper.

Lawrence Berkeley National Laboratory / UC Berkeley (Ceder group), with Google DeepMind's GNoME predictions · A-Lab

A-Lab combined GNoME's crystal-stability predictions with an autonomous, active-learning-guided robotic synthesis loop to attempt synthesis of computationally predicted inorganic materials without human intervention. After independent critique and a January 2026 Author Correction, the confirmed success count stands at 36 of 40 originally reported compounds, with the 'new to science' framing walked back to 'new to the prediction platform.' GNoME's separate 2.2-million-crystal database release carries no discrete validated synthesis result of its own and is quarantined separately in this run.

Google DeepMind · AlphaDev

AlphaDev, built on the AlphaZero reinforcement-learning algorithm, discovered shorter assembly-instruction sequences for sorting small fixed-size sequences than the human-optimized routines in use, and DeepMind reports the resulting algorithms were adopted into LLVM's C++ standard library.

Stokes Lab, McMaster University, with the Collins Lab, Massachusetts Institute of Technology, and the Broad Institute of MIT and Harvard · neural network for antibacterial activity prediction

Researchers screened about 7,500 molecules for growth inhibition of A. baumannii in vitro, trained a neural network on the resulting dataset, and used it to predict structurally new candidates with activity against A. baumannii. The top candidate, abaucin, showed narrow-spectrum activity by perturbing lipoprotein trafficking via LolE, and controlled an A. baumannii infection in a mouse wound model.