FULL STORY

AI Shakes Algorithm Bounds: From Breakthrough Rumors to Community Verification

Early October brought viral claims of AI-assisted complexity breakthroughs in integer multiplication and 3SUM, followed by skepticism over OpenAI's framing and a community livestream attempting to replicate the records with AI agents.

2026-10-06 ~ 2026-10-09 · 4 episodes · 30 posts

Episode 1 · Truly Subquadratic 3SUM and Subcubic APSP Break Decades-Old Barriers, Reportedly with Claude's Help (2026-10-06, 13 posts)

Josh Alman and Virginia Vassilevska Williams posted an arXiv preprint, "Truly Subquadratic 3SUM and Truly Subcubic APSP via Triangles in Sparse Lopsided Graphs," delivering the first polynomial improvements for two textbook problems: deterministic O(n^1.9992) for 3SUM on n polynomially-bounded integers, and O(n^2.9995) for integer-weight APSP. These break the long-standing n² and n³ barriers, effectively refuting the 3SUM and APSP hypotheses and reshaping fine-grained complexity theory. Social media reports say Anthropic's Claude contributed to discovering the core algorithms, though no direct statement from the paper or Anthropic is available yet.

Confirmed

  • The paper gives deterministic algorithms, the first truly subquadratic/subcubic results for 3SUM and APSP, for polynomially-bounded integer inputs with polynomial-related conditions
  • The paper comes with a Lean formalization and claims to refute the randomized integer Word-RAM 3SUM conjecture—per @basedjensen, ranked 159th on the authors' own list of most important open math problems—and states Claude contributed to finding refutation algorithms for 3SUM, APSP, and ExaTime-related problems
  • Per @thegautamkamath relaying Ilya Razenshteyn, Claude produced an O(n^1.9992) algorithm for 3SUM (deciding whether three of n integers sum to zero) with a Lean-formalized proof
  • @AMBNNJ says an internal Anthropic model independently discovered algorithms breaking the quadratic/cubic barriers; @robleclerc adds that the internal model found the core algorithms while exploring open problems, which researchers then simplified into proven polynomial speedups for both problems
  • Per @arannayebi, the same day saw multiple TCS breakthroughs: sub-n·log n FFT (new integer multiplication bound), subcubic APSP, subquadratic 3SUM, and a proof of the KLS conjecture; he joked the day shook his faith in human algorithmic ability
  • Per @HankYeomans relaying Przemek Chojecki, Claude helped improve polynomial bounds for the 3SUM-Hard family, which some see as progress toward P=NP
  • Per @arannayebi quoting NoahCitron's quip: mathematicians think math is beautiful, machines think it is not—humans simply happened to find the elegant solutions first, and the machine-derived subquadratic 3SUM algorithm is a case in point

Unconfirmed

  • Claude's involvement is so far reported only via social media (@AMBNNJ, @thegautamkamath, @robleclerc, @basedjensen, etc.), without direct elaboration in the paper or from Anthropic

Why it matters

  • The 3SUM and APSP hypotheses are pillars of fine-grained complexity, underpinning many conditional lower bounds; their refutation forces a redrawing of the field's landscape
  • Comments relayed by @thegautamkamath call it a genuinely remarkable result that also takes much of the fun out of the field, as many classic hardness assumptions now wobble
  • If Claude's involvement is confirmed, it would mark a landmark case of AI achieving a substantive breakthrough in open theoretical research, suggesting automated search may find algorithms humans long missed due to a preference for "elegant" solutions

Episode 2 · Integer Multiplication Breaks the n log n Barrier, with AI Leading the Way (2026-10-07, 11 posts)

Starting October 7, researchers on social media reported that integer multiplication has for the first time been pushed below the long-assumed O(n log n) lower bound—faster than Schönhage–Strassen–type algorithms—with AI credited as the main contributor. Together with recent results such as 3-sum at n^1.999, the news triggered widespread discussion and debate in the algorithms and complexity-theory community, flooding timelines to the point that some bloggers joked about buying a Texas Instruments calculator instead.

Confirmed

  • New work indicates integer multiplication can beat n log n; the news, relayed by @QuintinPope5 from @mgostIH, was described as "cursed."
  • @IgorCarron relayed commentary that AI (AlphaEvolve–style work) found a method faster than n log n while AES encryption remains unaffected—"just because it hasn't been announced?"—though this remains commentary and algorithm details are unpublished.
  • Technical approach (per a tweet cited by @deanwball): recast multiplication as a convolution network, run FFT, and exploit butterfly-network symmetries; asymptotically faster, but Dean Ball noted the speedup only shows up around the 37th decimal place.
  • The research process itself is being rewritten by AI: @arannayebi relayed Christian Szegedy's remark that such results once took years of scattered, peer-reviewed papers, but AI plus formal tools can compress the process to hours; Aurelien Prot built a live tracker streaming the community's AI-driven assault on the problem.
  • The main contribution comes from AI—a dramatic contrast with LLMs failing at multiplication five years ago (discussion relayed by @CatAstroPiyush).
  • @bzogrammer (relayed by @snikolov) offered a lay explanation: these are typical fractal-style complexity bounds, not bizarre results—naive O(n²) multiplication was long deemed optimal until Karatsuba's log2(3)≈1.58 algorithm; history is repeating.
  • Related "shaving" results: 3-sum down to n^1.999 and multiplication to n(log n)^0.999.
  • @chrisalbon noted timelines flooded with talk of "sub n log n" and "kappa constants."

Unconfirmed

  • Most claims come via second-hand relays; the original paper has not been formally released, and the algorithm details and exact role of AI are not given in the posts.

Why it matters

  • If the lower bound is truly broken, long-trusted complexity bounds may merely be "weak approximations" reflecting human cognitive limits.
  • The research paradigm shift matters: per Szegedy, AI plus formal tools could compress multi-year peer-review cycles into hours, with a public live tracker enabling community collaboration—potentially reshaping mathematical research itself.
  • Aesthetic debate: @thomasahle worried such results make deep mathematics "ugly" (square-packing-style bounds), yet sees beauty from another angle; @deanwball marveled that in 2026 we can still discover new things about "multiplication."
  • Practicality debate: @srchvrs questioned whether shaving small constants off known exponents is a fundamental breakthrough or a hack—many asymptotic improvements historically carry "galactic" constants with no practical impact; responses exist in @thomasahle's discussion threads.

Episode 3 · OpenAI's 'Below n log n' Claim Called Out as Wordplay (2026-10-07, 2 posts)

Posts citing OpenAI's internal models' TCS results, like 3SUM at O(n^1.9992), were mocked online: a claimed 'below n log n' bound is actually n(log n)^0.5499...9, a technicality amounting to wordplay.

Episode 4 · AI Agents Chase Integer Multiplication Complexity Records in Viral Livestream (2026-10-08, 4 posts)

An AI Agent community livestream tracks @0xdoug's agents attacking the integer multiplication complexity record, reaching 2^(-15) after previously hitting 2^(-34). The community has turned the progress into memes, joking about 'sub-linear time' multiplication and an impending 'foom' in big-number arithmetic.