Logged dailyLast checked

AI Discovery Ledger

A running record of scientific and technical results where an AI system did the discovering. Every entry carries a grade for how well the claim has been checked, because an announcement and a confirmed result are different things.

Get the weekly emailRSS

Entries logged
16
Peer-reviewed or confirmed
4 of 16
Latest entry
15 Nov
May ’26
Jun
Jul
Aug
Sep
Oct
Nov

One block per entry, by month, colored by evidence grade. Hatched blocks are contested.

1 entry

PhysicsMachine-checked

GPT-6 Astra credited with the proof behind a candidate lattice definition of the Standard Model

Nathaniel Craig (UC Santa Barbara, visiting scientist at OpenAI) · GPT-6 Astra

Putting the Standard Model on a spacetime grid has been stuck for decades because the weak force treats left- and right-handed particles differently, and Lüscher's fix for simpler theories had not been carried over to the full model in more than 25 years. Craig says he had a strategy but could not prove it in every case, and that Astra turned it into a construction and led the detailed proof. The preprint, posted October 6, builds a gauge-invariant fermion measure on four-dimensional lattices for a class of theories that includes the Standard Model, and comes with a Lean 4 formalization.

Still openThe model's role rests on Craig's own account, published on OpenAI's science blog. Three lattice theorists read the work privately, but no outside assessment has been published. Reflection positivity and a link to a continuum theory are not established, and Craig calls it an answer in principle more than in practice.

1 entry

MedicinePeer-reviewed

Deep learning finds an ECG signal that flags sudden cardiac death risk

UC Berkeley and collaborators · Nature · Deep-learning model trained on 440,000+ ECGs

A model trained on Swedish heart tracings linked to death records picked out a pattern in ordinary ECGs that marks people at high risk of sudden cardiac death. The group it flagged died suddenly at about 7% a year, against 4.6% for the group picked by the standard pumping-strength test, and the pattern held in patient records from the US and Taiwan.

Still openNobody yet knows what the signal is biologically, and it is not in clinical use.

Source: ScienceDaily

1 entry

MathematicsMachine-checked

Gemini agents help prove the Courtade–Kumar conjecture, with a batch of new algorithm results

Google Research with CUHK and other academic co-authors · Antigravity /teamwork agents on Gemini 4 models

The 2013 conjecture says no one-bit summary of a noisy signal carries more information than simply reading off a single input bit. Google says Gemini found counterexamples to a broader version, worked out most of the analytical bounds and interval-arithmetic certificates, and extended the proof from balanced to unbalanced functions without human input; the whole proof is formalized in about 50 million lines of Lean 4. The same post lists about ten further new results, including a deterministic Boolean matrix multiplication algorithm that answers an open question from STOC 2024 and the first n-bit functions needing more than n/2 quantum queries.

Still openGoogle and its co-authors are the only source, and nothing here has been peer reviewed or independently reproduced. People supplied key steps: the CUHK team proposed the function that cut the problem to a 14-variable optimization and gave one analytical bound. Google gives no count of problems attempted or failed.

1 entry

BiologyAnnounced

AlphaGenome 2 predicts what a DNA variant does in one chosen cell type

Google DeepMind · AlphaGenome 2

The model takes a stretch of DNA plus a tissue or cell state and predicts gene activity there. It recovered a known inflammatory bowel disease mechanism in gut macrophages and predicted that a variant tied to severe COVID-19 disrupts gene activity most in activated T cells.

Still openThe findings are computational predictions that still need lab confirmation, and the announcement names no journal.

1 entry

Computer scienceMachine-checked

AI agents formally verify the code behind Signal's encryption

Google DeepMind · AlphaProof Code, agents on Gemini 4 Carbon

In about four weeks, agents wrote millions of lines of machine-checked Lean proof showing that the compiled Signal protocol code keeps its core security promises. When researchers put two old, already-fixed bugs back into a copy of the code, the system proved each buggy version insecure.

Still openA proof is only as good as its specification. Those were written and reviewed by people but have not had an outside audit, and no previously unknown bugs were reported.

2 entries

Computer scienceMachine-checked

AI system proves the Aaronson–Ambainis conjecture on the limits of quantum speedups

Physical Superintelligence (PSI) · PSI research system running a non-public Anthropic model

The conjecture, open since 2008, says every bounded low-degree polynomial has at least one input bit with sizeable influence. PSI's system had a full proof by day three of its run and then tightened the bound twice, ending at an influence of at least about 0.0025·v^(5/2)/d^6 for variance v and degree d, where earlier bounds shrank exponentially with the degree. It follows that a quantum algorithm cannot save exponentially many input reads on typical random inputs to a yes/no problem.

Still openLean accepts the main theorem, but the paper sits on GitHub without peer review. The only outside readers so far are a Stanford group whose comments are known through PSI, and the model is not public. The result limits input reads, not total running time.

Computer scienceAnnounced

AI agents find two new weaknesses in retired RSA-1024 encryption

Google DeepMind with UC San Diego and Inria · Supervised AI research agents

The agents found a signature-forgery shortcut that is up to 1,600 times faster on paper when an attacker controls only a small part of the signed data, and a batch method that cuts the cost of factoring each additional key about 30-fold.

Still openBoth headline figures are theoretical or asymptotic, not measured attacks. Keys of 2048 bits and larger are unaffected.

1 entry

BiologyPeer-reviewed

Proteins designed from plain-language descriptions work in the lab

Westlake University · Nature · Pinal

Pinal turns a written description of a desired function into protein sequences. It produced working proteins on all four test tasks, a plastic-degrading enzyme among them, and one design lifted a carbon-fixation pathway to 1.7 times its natural output.

Still openHit rates were uneven. One task yielded a single working protein from 29 tested.

1 entry

MedicineAnnounced

AI-designed antibody roughly doubles grip strength in mice with a muscular dystrophy

Anthropic and Mithril Medicine · Claude, per Anthropic

Anthropic reported a preclinical antibody for LAMA2-related muscular dystrophy that went from idea to mouse data in under 18 months. One design about doubled grip strength compared with untreated mice.

Still openThis comes from a company technical report that has not been peer reviewed. Human trials are not expected before late 2028.

Source: The Next Web

3 entries

Computer scienceMachine-checked

OpenAI model claims the Exponential Time Hypothesis is false

OpenAI · Internal reasoning model

The hypothesis says 3-SAT cannot be solved in sub-exponential time, and a large part of complexity theory is built on it. OpenAI says its model wrote a disproof on October 16 that was checked in Lean the next day. The paper and the formal proof are public.

Still openOutside experts are still working out whether the formal statement matches the claim and what the algorithm means in practice.

MathematicsMachine-checked

OpenAI model claims a proof of local smoothing in every dimension

OpenAI · Internal reasoning model

The local smoothing conjecture for the wave equation implies the Kakeya and restriction conjectures in harmonic analysis. OpenAI says its model proved it for all dimensions from two up on October 18, formalized it in Lean a day later, and published both on an expert's advice.

Still openAnalysts have not yet confirmed the argument by hand.

PhysicsMachine-checked

OpenAI reports a quantum chromodynamics result with a formalized Yang–Mills existence proof

OpenAI, with a physics group led by David Gross · Internal reasoning model

The third of three results announced together. OpenAI says the Yang–Mills existence part was formalized in Lean on October 16 and handed to a group of physicists for review.

Still openOn November 9–10 the five-physicist group, which OpenAI chose, reported that the Lean code rebuilds cleanly and that its definitions match the standard statement of four-dimensional Yang–Mills existence, while saying it does not yet fully understand the proof. No fully independent check is on record, the 1,147-page written proof is still private, and the mass gap, the other half of the Millennium problem, is untouched. Critics including Peter Woit question whether the verified statement meets the prize problem's standard.

1 entry

MathematicsContested

OpenAI releases 722 AI-written manuscripts claiming hundreds of open problems

OpenAI · Internal frontier model

Posted to GitHub across 372 topics, the claims include a matrix multiplication exponent of 9/4, a proof of the unique games conjecture and a wide zero-free region for the Riemann zeta function. About half the manuscripts come with Lean proofs.

Still openThree papers were retracted and 14 revised within two days. Some formal proofs appear to establish less than the papers claim, and full review is expected to take months.

Source: Wikipedia

1 entry

MathematicsContested

OpenAI claims a solution to the Navier–Stokes Millennium Prize problem

OpenAI · Multi-agent system of roughly ten thousand agents

The company said its system settled existence and smoothness for the fluid equations with a smooth external force, one of the seven Millennium Prize problems.

Still openNo complete independent check has been published. A paper by Bastounis, Circelli and Hansen argues that the Lean code does not correspond to the written blow-up proof, citing at least two discrepancies; that does not by itself show the proof is wrong, and no OpenAI fix is on record. Credit is also disputed: mathematician Tristan Buckmaster says unpublished results of his reached OpenAI days earlier.

1 entry

BiologyPeer-reviewed

First working viruses built from AI-generated whole genomes

Arc Institute and Stanford · Science · Evo 1 and Evo 2 genome language models

The models wrote complete bacteriophage genomes modeled on ΦX174. Of nearly 300 that were synthesized, 16 produced viable phages, and a mixture of them overcame E. coli strains that resist the natural virus.

Still openThe hit rate was about 5% on one of the smallest genomes known. Larger genomes are still out of reach.

1 entry

AI Discovery Ledger Weekly

One email every Friday, about a three-minute read: the week's lead result, the grades that moved up or down, what else was logged, and what is still open.

Unsubscribe from any issue. Prefer a feed reader? Follow by RSS.

How entries are graded

Each entry gets the highest grade its public record supports on the day it is logged, and the grade is revised when the record changes.

What gets logged

A result belongs here when an AI system produced the central finding: a proof, a design that worked in the lab, a new algorithm, a pattern in data that people had missed. New model releases, funding rounds, benchmark scores and product launches are left out, as are results where AI only sped up routine work.

Summaries are written from the linked sources. Where a lab is the only source for its own claim, the entry says so.