and not using a Hadamard matrix of order 668 and with a mathematical theorem I didn’t but belief sufficient to name a outcome.
Working with GPT-5.6 Sol, parallel brokers, exact-arithmetic packages and a proof assistant, I attacked two laborious issues over two days. One resisted each route I attempted. The opposite produced a proof candidate: a standard argument written out in full, precise checks on chosen inputs, a educating information and a partial formalisation in Lean.
That second end result remains to be a candidate. No specialist has reviewed the argument, its novelty just isn’t established, and Lean checks solely its algebraic core moderately than the whole theorem. I might not publish it as a brand new theorem in its present state.
The sample I took from the weekend is that mathematical experiments have gotten plentiful. The Hadamard venture closed 44 exactly outlined areas of the search house and audited 5 customary routes to a nonexistence proof. The Maxwell venture produced a whole proof candidate with a machine-checked algebraic core. Accepted mathematical data didn’t get cheaper alongside any of that, as a result of proof, understanding, novelty and impartial evaluation are separate obligations, and producing one other route settles none of them.
This can be a concrete sequel to 2 earlier concepts. In AI Has Made Analysis Low-cost. Understanding Is Nonetheless Costly, I argued that AI was making experimentation far cheaper than understanding. In From Tokens to Theorems: Constructing a Neuro-Symbolic AI Mathematician, I constructed a easy neuro-symbolic loop by which an LLM proposed mathematical formulation, SymPy checked them precisely, and failed candidates grew to become suggestions for the following try. That experiment was intentionally modest, however it uncovered an structure that reappeared right here at a a lot bigger scale.
Over this weekend, the identical sample prolonged past symbolic sequence issues. As a substitute of 1 mannequin and one symbolic checker, I used parallel brokers, exact-arithmetic packages, adversarial critics and a proof assistant. Candidate constructions and arguments have been generated, attacked, checked the place attainable and both discarded or retained with an express standing. The loop had grow to be richer as AI techniques improved via 2025 and 2026, however its primary construction remained the identical.
Wanting again afterwards, I realised the workflow carefully resembled a broader sample that Jeff Dean lately described for science and engineering. Quite than viewing AI as a device for fixing one drawback at a time, he characterised analysis itself as a recurring cycle: suggest an experiment, implement and run it, consider the result, then use that analysis to generate a greater experiment. His argument was that AI ought to automate this whole loop, decreasing iteration time, operating many experiments in parallel and studying from each analysis.
That was nearly precisely what had occurred over my weekend, albeit on a a lot smaller scale. The experiments have been mathematical moderately than bodily. “Implementation” meant exact-arithmetic packages, constraint solvers or proof assistants as a substitute of laboratory gear. Analysis meant integer verification, counterexample search and partial formalisation moderately than measuring a bodily system. Parallel brokers proposed constructions, generated verifiers, searched the literature, criticised arguments and translated concepts into totally different mathematical representations earlier than the following iteration started.
Seen this manner, probably the most placing growth was not that AI produced a proof candidate or did not assemble a Hadamard matrix. It was that the experimental cycle itself had grow to be partially automatable. As soon as precise evaluators existed, proposing, executing and refining mathematical experiments grew to become dramatically cheaper.
The necessary distinction was not merely considered one of scale however of validation. Within the earlier experiment, a candidate components may very well be examined immediately in opposition to the sequence, though reproducing the noticed values didn’t essentially reveal the underlying arithmetic. Right here, verification grew to become layered. Precise arithmetic may reject incorrect constructions, Lean may certify components of an argument, and literature searches may assist set up prior work, however none may alone decide novelty, validate each bridge in a common proof or resolve whether or not a outcome deserved to grow to be accepted mathematical data.
Hadamard 668 resisted a precise search stack
The primary drawback asks for a 668 by 668 grid crammed solely with +1 and -1, organized in order that any two totally different rows cancel out fully: multiply them entry by entry, add up the outcome, and you need to get precisely zero. In linear-algebra phrases, the rows are mutually orthogonal. This can be a Hadamard matrix of order 668, and matrices of this type present up in error-correcting codes, sign processing and the design of experiments. In line with Epoch AI’s present catalogue of the issue, 668 is the smallest order for which no person is aware of whether or not one exists.
That is an interesting goal for computation as a result of a proposed reply is straightforward to test. Multiply the candidate by its personal transpose utilizing precise integers, then learn off the outcome: each diagonal entry have to be 668 and each different entry precisely zero. That could be a single matrix multiplication, and there’s no subjective judgement on the remaining gate.
Discovering one is one other matter. The grid has 446,224 entries and every is both +1 or -1, so a naive search has 2⁴⁴⁶,²²⁴ candidates to get via. No quantity of compute touches a quantity that measurement. All of the helpful work goes into not looking out it, by imposing sufficient construction that the variety of free selections collapses.
That collapse is what the identified development households purchase you. The primary route right here makes use of the truth that 668 is 4 instances 167. Discover 4 quick +1/-1 sequences, of lengths 84 and 83, whose correlation errors cancel at each offset, and a normal recipe assembles them into the complete matrix. That recipe is carried out within the repository and was checked finish to finish on the smaller orders 28 and 36, the place precise integer checks affirm the ensuing matrices. This takes the search from 2⁴⁴⁶,²²⁴ candidates right down to 2³³⁴. Nonetheless far too many to enumerate, which is the sincere cause the issue is open, however now sufficiently small that construction and symmetry arguments can chew.
Different routes undergo totally different households, every with its personal identify: Legendre pairs of size 333, Williamson-type quadruples, cocyclic and transposed-Ito constructions over teams of order 167, and distinction households with prescribed symmetries. The names are secure to skim previous if they’re unfamiliar. What issues is that every is a distinct means of buying and selling one huge unstructured seek for a small structured one, and that every carried precise verifiers for the intermediate objects it accepted. Failed candidates have been rejected by integer arithmetic, not by approximate similarity.
A number of of the exclusions are exact sufficient to state numerically. A correlation sure proves that no answer lies inside Hamming distance 31 of 1 specific 334-bit place to begin: flipping any 31 or fewer of its bits can’t attain a legitimate reply. The evaluation bundle will embrace its full bit strings and digest. Individually, a precise enumeration covers all 4,096 quadruples constructed from circulant sequences, those the place every row is the earlier row rotated by a single place, restricted to these whose sample of destructive entries survives multiplying each index by 2 modulo 167. None of them works.
A counting argument over odd and even positions guidelines out another form: the case the place the third and fourth sequences are the primary two with the signal of each different entry flipped. Every of those exclusions covers solely the neighbourhood or symmetry sample it names, and nothing past it.
The venture additionally labored the issue from the opposite aspect and tried to show that no such matrix exists on this household. That failed too, and the failure is restricted sufficient to be helpful. Nonexistence outcomes on this space draw on a normal toolkit, and the audit labored via it merchandise by merchandise: Leung-Schmidt area descent, splendid factorisation and 2-adic valuation, Hasse-Minkowski, the multiplier theorems, and Bruck-Ryser-Chowla. Right here too the names matter lower than the end result. Each one both fails to use at these parameters or applies and produces no obstruction. What stays is a brief record of unresolved mathematical questions; none of those 5 customary routes provides the lacking obstruction.
Calibration mattered greater than any single run. The constraint solver was given a a lot smaller occasion of the identical four-sequence drawback, one already identified to have an answer: sequences of size 15 and 14, in opposition to the goal’s 84 and 83. It didn’t discover that identified answer inside two minutes. The timeout proved nothing concerning the feasibility of both occasion. It confirmed that this solver setup couldn’t clear up even the smaller occasion throughout the two-minute calibration price range, so I ended treating longer runs of the identical setup as the primary route ahead.
By the top, each closed area of the search house had been documented, along with an express record of 9 remaining open questions. A number of promising heuristics proved ineffective. No Hadamard matrix of order 668 was discovered, and nothing right here advances the overall conjecture. The issue stays open, and the eventual development might lie in a household this venture by no means explored. Even so, the journey was immensely rewarding. Exploring new arithmetic alongside an AI, making use of identified strategies to a longstanding open drawback, gave a glimpse of upper arithmetic changing into more and more accessible and opening the door for extra folks to take part in mathematical discovery.
Maxwell’s drawback produced a candidate that’s a lot more durable to test
The second venture started with an previous drawback from mathematical physics. Place some level prices in house. Their mixed potential creates equilibrium factors the place the forces steadiness. The aim is to sure what number of such factors there could be.
The final story modified lately. A July 2026 paper by Arathoon, Ball and Kvalheim constructed 5 level prices with at the very least 24 nondegenerate equilibria, disproving Maxwell’s proposed common components of (okay − 1)² for okay prices. A second July paper by Gabrielov, Novikov, Novikov and Shapiro sharpened the sure for 3 optimistic prices from twelve to 6 nondegenerate equilibria, for each optimistic Riesz exponent. Each papers can be found on arXiv: The Maxwell Conjecture is False and From 12 to six: Sharpening the Three-Cost Certain in Maxwell’s Downside.
The precise candidate declare begins with three distinct supply factors in Euclidean house of dimension at the very least two, three optimistic prices and a optimistic exponent α. Their potential at some extent p away from the sources is:
The candidate declare is that this potential has at most 4 nondegenerate equilibrium factors for each α > 0. It treats collinear sources individually and claims precisely two equilibria there. The acquainted Coulomb potential is the case α = ½. That’s beneath the sure of six the July paper establishes, which is why the argument wants a specialist reader, not my very own confidence in it.
The geometric thought behind the candidate could be described with out the technical equipment. For 3 noncollinear prices, each equilibrium lies inside their triangle. Its place could be represented by three optimistic weights, known as barycentric coordinates. The candidate argument rewrites the bodily drawback as a query concerning the peaks of a associated mathematical floor.
If two of the related peaks existed, draw the straight chord becoming a member of them. Each endpoints must curve downwards alongside that chord. The candidate argument derives precise formulae for these two endpoint curvatures and makes use of a second inequality to indicate that they can not each be destructive. If each bridge in that argument is sound, this leaves at most one nondegenerate native minimal of the bodily potential contained in the triangle. A planar index rely then provides the proposed higher sure of 4 equilibria.
The 2 initiatives require totally different validation. A single proposed Hadamard matrix could be settled by one precise calculation. The Maxwell candidate issues each supply triangle, each set of optimistic prices and each optimistic exponent, so it is determined by a sequence of quantified geometric, analytic and topological steps. A believable proof can cover an error in any bridge between them.
The traditional manuscript has been checked internally. Precise rational packages consider its derived identities on nonsymmetric rational inputs at α = ½ and α = 1. These are precise transcription controls on chosen examples, not a symbolic verification of the quantified identities.
The partial Lean growth accomplished efficiently, and a supply scan discovered no sorry, admit or added axiom. Its 33 named theorems cowl the central second inequality, endpoint-gap algebra and an summary two-dimensional Hessian signal argument. They don’t but formalise the barycentric correspondence, the differential identities linking the bodily drawback to the summary matrices, the endpoint-curvature derivation, the worldwide index and compactness arguments, or the collinear and higher-dimensional reductions. Lean is checking the algebra, and solely the algebra.
That boundary issues. Formalising the algebra that Lean accepts most readily can create an undeserved halo of confidence across the unformalised geometry. The subsequent formalisation goal ought to prioritise the steps almost definitely to comprise an error.
The analysis loop grew to become partially automatable
I used the mannequin as one part in a analysis harness, alongside precise packages, a proof assistant and express proof guidelines. The preliminary exploration ran over a weekend, adopted by extra checking and writing. I chosen the issues, redirected or stopped unproductive routes, required the standing labels and determined which claims may seem right here. Contemporary-context critics acquired an artefact and an adversarial guidelines as a substitute of the entire working dialog. The deliberate public bundle will document the mannequin configuration, prompts, code variations and commit hashes wanted for a fuller provenance path.
One agent proposed formulations. One other tried to interrupt them. Others wrote precise verifiers, looked for counterexamples, translated bounded questions into constraint techniques, in contrast an argument with the literature, or defined an unfamiliar definition from a number of instructions. The work accrued via a repeatable loop:
- state a exact declare;
- derive penalties;
- take a look at small or finite instances precisely;
- ask a fresh-context critic to assault the weak factors;
- formalise the parts the place machine checking provides confidence;
- replace the standing ledger;
- preserve, revise or discard the concept.
Inside this specific harness, producing one other believable route was quick relative to checking it. Recording its precise scope, finding its weakest inference, checking whether or not it was already identified and deciding whether or not it deserved professional consideration remained costly.
That loop may run throughout many branches in parallel. A failed strategy now not consumed a whole night earlier than I learnt that its assumptions have been inconsistent. A conjectured identification may very well be changed into a precise program and falsified shortly. A dense proof may very well be rewritten as a geometrical image, then as algebra, then as a listing of obligations for formalisation. The Hadamard venture subsequently left a helpful document regardless that it missed its goal: exact reductions, closed areas, failed strategies and calibration outcomes that may forestall the identical blind searches from being repeated.
This isn’t a conveyor belt working via a finite record of open issues. Proving a theorem modifications the encircling map. It exposes new constructions, suggests conjectures, connects questions that had regarded unrelated and creates new instructions to discover. Mathematical analysis is cyclical: a outcome closes one query whereas opening a number of others. Making every flip of that cycle cheaper can produce extra arithmetic to research, not an approaching finish to the topic.
The mannequin additionally accelerated studying
I got here to each initiatives with out specialist coaching of their underlying arithmetic. Over the weekend I bumped into aperiodic autocorrelation, algebraic norms, barycentric coordinates, Hessians, second inequalities and planar index principle.
The mannequin defined every thought repeatedly and from totally different angles. It moved between formulae, small numerical examples, visible instinct and the function an thought performed within the wider argument. When an evidence didn’t land, I may say so with out embarrassment and ask for an additional. Ultimately a number of items clicked into place.
Michael Nielsen describes a associated course of in Utilizing spaced repetition techniques to see via a chunk of arithmetic. His central level is that mathematical understanding just isn’t binary however could be deepened nearly indefinitely by decomposing a proof, restating its concepts in several kinds, probing variations and rebuilding the connections between them. With sufficient familiarity, the proof can start to really feel nearly clear: much less like a sequence of symbols to be adopted and extra like a set of mathematical objects whose relationships could be manipulated immediately.
My expertise with the mannequin felt like an interactive analogue of a part of that course of. As a substitute of developing spaced-repetition playing cards, I may repeatedly ask for a similar thought to be represented as algebra, geometry, a numerical instance, an instinct or a problem query. The necessary half was not that the mannequin provided an evidence as soon as, however that it made repeated reformulation low-cost. Nielsen calls the eventual state having the ability to “see via” a chunk of arithmetic. I didn’t attain something like specialist mastery over a weekend, however I may really feel the beginnings of the identical transition from following an argument line by line to recognising bigger chunks of its construction.
Earlier than the weekend, I realised, I had subconsciously considered mathematical analysis largely as fixing issues already ready on a listing. The work itself felt totally different. A lot of it consisted of studying unfamiliar concepts, following failed routes far sufficient to grasp why they failed, reformulating questions and noticing sudden connections. A proof was one milestone in that bigger exploratory course of, not the method as a complete.
The result was working literacy moderately than experience. I may observe the construction properly sufficient to ask higher questions, discover when two claims had been conflated and perceive what an exterior reviewer would wish to examine. A transparent clarification nonetheless offers no proof {that a} theorem is true, and I needed to preserve reminding myself of the distinction between following an argument and having the ability to reconstruct it.
For me, this was probably the most quick type of mathematical abundance. The mannequin met me on the level the place understanding stopped and saved altering the illustration till progress resumed. It made an unfamiliar topic explorable with out making me a specialist in it.
In Mathematicians are grappling with the likelihood that AI would possibly eclipse them for Understanding AI, Kai Williams experiences on interviews with greater than twenty mathematicians. What he heard most frequently was not fashions proving issues, however fashions used to discover a means into unfamiliar areas of the literature. That matches my expertise precisely. The mannequin provided a map and several other translations; the underlying papers, derivations and precise checks nonetheless needed to carry the declare.
Tasmin Chu’s essay Mathematicians must act identifies a danger in the identical workflow. She argues that the modest extensions, literature workout routines and first proofs most uncovered to automation are additionally how college students grow to be researchers. If a mannequin does that work as a substitute of directing it, the sphere might weaken that pipeline. My very own extension of her level is that it’s also the place future reviewers study judgement. One weekend can’t settle the priority. I learnt most once I needed to reconstruct an argument, ask what would falsify it and uncover why an evidence had did not land, not once I merely acquired a solution.
Bigger initiatives present the identical uneven automation
Two a lot greater efforts level the identical means.
In 2025, Google DeepMind reported that AlphaEvolve had been utilized to greater than 50 open mathematical issues. In line with DeepMind, it rediscovered the very best identified answer in roughly three quarters of the instances and improved the very best identified lead to a few fifth. The strategy applies to a helpful however slim class of issues whose proposed options could be expressed as algorithms and scored robotically.
In Might 2026, OpenAI reported {that a} general-purpose mannequin discovered a counterexample to the Erdős unit distance conjecture. 9 exterior mathematicians then produced a brief, digested, human-verified model of the argument, which put the human verification into the revealed document moderately than leaving it as a declare concerning the mannequin.
Each examples present that automation is uneven. Machine-readable scoring lets a system reject unhealthy candidates and iterate with out ready for an individual to learn every one. A common proof candidate nonetheless requires cautious work on ideas, translations and quantified arguments.
Tom Zahavy’s ICML 2026 place paper LLMs can’t bounce presents a vocabulary for that cut up. Borrowing Peirce’s three modes of inference, he argues that machine studying has mechanised induction, which finds a rule by compressing many examples, and is quickly mechanising deduction, which derives penalties from premises already mounted. What it has not mechanised is abduction: proposing the premise itself to clarify one thing stunning. His case examine is common relativity, and his central commentary is that Newtonian gravity confronted no measurable disaster on the time. The equivalence of inertial and gravitational mass had been checked to a precision of10⁻⁹, and the one anomaly, Mercury’s orbit, was broadly attributed to an undiscovered planet. An optimiser would have discovered nearly nothing to enhance. He grants {that a} mannequin given Einstein’s 1915 postulates may plausibly derive the sphere equations, since that half is deduction; the 1913 model failed as a result of the axioms have been mistaken, not as a result of the logic was.
The paper is a place piece and says so, and Zahavy is express that his argument is aimed on the bodily sciences, the place the uncooked materials is sensory, and that arithmetic grounds its intuitions otherwise. I might not push the analogy previous that. However the division of labour was recognisable from the within. The precise packages and Lean did deductive work properly, and the Maxwell candidate progressed as a result of a framing already existed to infer inside. Hadamard 668 had no such framing at hand, and what it lacks just isn’t extra search however a premise: a development household or a theorem no person has written down. His studying of AlphaEvolve makes the identical level from the opposite aspect: it optimises properly inside a hard and fast framework as a result of it has a gradient to observe.
The tempo has not slowed since. Williams opens his article with a Fields Medallist becoming a member of OpenAI’s security crew and an inner mannequin that the corporate says solved ten main open issues; Chu opens hers with the identical announcement, noting that the reported token value of these outcomes was round two thousand {dollars}. I’ve not checked both declare, and neither had been via evaluation once I wrote this.
Jordan Ellenberg captured the optimistic historic response in his 2014 e book How To not Be Incorrect, in a passage Williams quotes: “We’ll reclassify that analysis as ‘computation.’” Arithmetic has repeatedly moved its frontier after machines absorbed an previous job. Reclassification doesn’t, by itself, clarify how folks will study to decide on the following questions, choose the solutions or maintain a group round them.
AI might subsequently improve mathematical abundance in a broader sense than proving extra theorems. It could generate extra conjectures, partial theories, proposed connections between fields and instructions value investigating. My weekend doesn’t set up how typically these solutions can be deep or genuinely new; it reveals solely how cheaply candidate routes can already be generated and examined. Even a modest improve would shift extra of the burden towards understanding, curation, evaluation, clarification and prioritisation. The problem just isn’t solely to provide extra arithmetic, however to resolve what deserves consideration.
Jeff Dean, in founding Discoveryloop, frames this as a programme extending far past arithmetic. Fusion, medication, cybersecurity and supplies science all share a model of the identical experimental cycle: suggest an experiment, run it, consider the outcome and use that analysis to decide on the following experiment. If AI more and more automates that loop throughout science and engineering, arithmetic could also be an early illustration of a broader transition moderately than a particular case. The bottleneck then strikes from producing experiments to deciding which outputs are dependable, important and worthy of changing into trusted data.
Abundance creates a human and institutional drawback
If fashions can generate many extra conjectures, proof candidates, counterexamples and partial formal developments than people can learn, storing them in chat transcripts won’t work. The identical acceleration additionally modifications who learns the craft, who receives credit score, who takes duty for errors and who’s requested to evaluation the output.
A helpful mathematical document wants greater than a title and a PDF. It ought to carry a normalised assertion, express assumptions, a standing label and the exact scope of any computation. It also needs to carry the formal artefacts the place they exist, the dependencies on earlier outcomes, the provenance of the human and machine contributions, the novelty standing, the identified assaults and a plain-language clarification.
Search ought to function over claims and dependencies moderately than solely key phrases. A helpful system may match a proposed lemma in opposition to equal statements in several notation, present which unreviewed claims would suggest a goal outcome, and establish arguments that depend on the identical unformalised bridge. Failed routes needs to be searchable too, offered their scope is precise. Over time, this might grow to be a shared corpus of profitable proofs, failed makes an attempt, deserted search branches, reductions, counterexamples, calibration outcomes and intermediate constructions. Each well-scoped failure may slim the remaining search house as a substitute of being unknowingly repeated.
With out that infrastructure, I count on repeated work and misplaced confidence. Fashions will rediscover previous outcomes, subtly range false proofs and generate extra materials than peer evaluation can take up. The bottleneck would transfer from producing arithmetic to curating a reliable map of it.
Williams’s interviews have been reporting, not a consultant survey, and the folks he spoke with provided no single response. Many anticipated AI to enrich their work within the close to time period and already used it in restricted methods, particularly to traverse unfamiliar literature. Others frightened concerning the coaching path for youthful researchers, future funding and a occupation organised round duties that fashions might study to carry out. Their disagreement displays a number of targets being bundled collectively. Fixing open issues is one intention of arithmetic, however so are understanding, clarification, theory-building, educating and sustaining a group.
The Leiden Declaration warns that present automated strategies produce believable however unreliable arguments which are laborious to tell apart from right proofs, that the identical drawback applies to formalisations via the interpretation between machine and human shows, and that AI-assisted papers make reviewing extra demanding. Timothy Gowers, writing about the declaration, pushes the thought additional: he imagines mathematicians choosing from an enormous physique of AI-generated arithmetic and writing it up in order that others can take up it, and he’s candid about how a lot of the present tradition that will displace.
Any such map will want popularity and incentives behind it. Gowers makes the purpose sharply: if one individual will get a mannequin to unravel an open drawback and one other digests the answer and explains it in order that mathematicians can study from it, the second individual ought to get many of the credit score. Explaining why a generated outcome issues, discovering the refined flaw in it, or connecting it to a uncared for theorem could also be value greater than producing the primary draft. As we speak’s publication tradition just isn’t designed to recognise these contributions cleanly.
The attribution drawback reaches additional again than the ultimate proof. Many celebrated breakthroughs are the end result of many years of definitions, lemmas, conjectures, failed approaches and strategies developed by many researchers. If an AI provides the final lacking argument, rewarding solely that final mile dangers obscuring the human work that made the step attainable and gave it which means. Provenance ought to subsequently document not solely who or what accomplished the proof, however the chain of concepts and outcomes on which it relies upon. Credit score ought to observe that chain moderately than acquire at its endpoint.
Chu’s essay sharpens all of this into an institutional argument. She presents correct attribution, duty for errors and open data as mathematical values that AI use might pressure. A generated proof can obscure the human work it is determined by, and a believable error can eat way more professional consideration than it took to provide. She additionally worries that researchers will share fewer unfinished concepts if someone else can convert these concepts into model-assisted outcomes earlier than their originators end. That final level is a forecast moderately than one thing my experiment demonstrates, however it means a declare registry would wish guidelines for consent and attribution round unpublished concepts, not solely higher search over completed ones. Among the many prescriptions she attracts from which are that mathematicians ought to use fashions judiciously or by no means, and that chat logs needs to be revealed alongside mathematical papers.
Impartial evaluation could be the following pure step
I don’t intend to show this private weekend studying venture right into a public evaluation effort. If I did, I might first set up the literature and precedence place, invite specialists to assault the traditional argument, and formalise the steps most prone to undetected error. Till these gates are handed, it stays a proof candidate.
For Hadamard 668, a remaining precise verifier has been written however has not but been exercised on a real candidate. The lacking ingredient is a brand new structural development or theorem. The destructive document ought to information future makes an attempt with out being mistaken for an answer or a nonexistence proof.
Whether or not or not both venture goes additional, the sturdy object just isn’t a proof alone however its place on the map. A proof sitting on a server has surprisingly little worth if no person is aware of it exists, why it issues, the way it pertains to earlier work or whether or not it may be trusted. As mathematical artefacts multiply, navigation via them turns into the scarce useful resource: finding the related declare, understanding its dependencies, judging its standing and explaining the place it modifications the frontier.
Working the loop quicker doesn’t shut that hole. Low-cost experiments multiply the issues we will test and the questions we will ask; they don’t, on their very own, inform us which premises are value checking or flip an output into shared understanding. Experiments grow to be low-cost. Information doesn’t.
Disclaimer: This was an off-the-cuff, curiosity-driven venture undertaken in my private time. The views expressed listed here are my very own and don’t signify these of my employer or any affiliated organisation. The content material relies on private expertise and reflection, and shouldn’t be taken as skilled or educational recommendation.
📚References
- Gabrielov, A., Novikov, Dm., Novikov, T., & Shapiro, B. (2026). From 12 to six: Sharpening the Three-Cost Certain in Maxwell’s Downside. Proves
that three optimistic level prices admit at most six nondegenerate equilibria for each optimistic Riesz exponent, bettering the twelve-point sure from the authors’ earlier work by supplying a separation argument on the saddle of a separated-variable first integral. That is the revealed sure the proof candidate mentioned right here would sit beneath, and the rationale that candidate wants a specialist reader moderately than my very own confidence in it. - Arathoon, P., Ball, G., & Kvalheim, M. D. (2026). The Maxwell Conjecture is False. Displays 5 level prices whose electrostatic potential has at the very least 24 vital factors, all nondegenerate, refuting Maxwell’s proposed sure of (okay−1)² for okay prices. Along with the paper above it defines the state of Maxwell’s drawback in July 2026, which is the setting for the four-equilibrium declare described right here.
- Zahavy, T. (2026). Place: LLMs can’t bounce. ICML 2026, PMLR 306. Argues through Peirce’s three modes of inference that machine studying has mechanised induction and is quickly mechanising deduction, however not abduction — the invention of the premise itself — utilizing common relativity as a case examine by which no error sign existed for an optimiser to observe. Provides the vocabulary this text makes use of to separate the deductive work Lean did properly from the lacking premise that retains Hadamard 668 open.
- Chu, T. (2026, 2 August). Mathematicians must act. Units out attribution, accountability for error and open data as group values that AI use might pressure, argues that the tractable issues most uncovered to automation are additionally how analysis mathematicians are skilled, and takes the place that within the present local weather it’s unethical to ask a mannequin to show a brand new theorem. The usual in opposition to which this venture is measured within the closing part.
- Williams, Okay. (2026, 4 August). Mathematicians are grappling with the likelihood that AI would possibly eclipse them.Understanding AI. Reporting from the Worldwide Congress of Mathematicians based mostly on conversations with over twenty mathematicians, discovering the commonest use to be navigating unfamiliar areas of the literature moderately than proving outcomes, and quoting the passage from Ellenberg’s How To not Be Incorrect reproduced right here.
- Nielsen, M. A. (2019, 12 January). Utilizing spaced repetition techniques to see via a chunk of arithmetic. Cognitive Medium. Describes mathematical understanding as an open-ended strategy of repeatedly decomposing, reformulating and connecting the weather of a proof till they grow to be deeply internalised. Nielsen makes use of spaced repetition because the mechanism, arguing that the method can ultimately produce a way of having the ability to “see via” the arithmetic moderately than merely reproduce its symbolic steps. This offers a helpful analogue for the repeated, multi-representation explanations used through the studying course of described right here.
