Daily Twitter Digest

Computer Science

Decades-Old 3SUM and APSP Complexity Barriers Finally Broken

Alman and Vassilevska Williams give the first polynomial-time improvements over textbook algorithms for 3SUM (O(n^1.9992)) and All-Pairs Shortest Paths (O(n^2.9995)), refuting the long-standing 3SUM, APSP, Exact Triangle, and related fine-grained complexity hypotheses. The results stem from a new algorithm for 'thin' matrix products, built on a Coppersmith/Schönhage-style identity, which computes only the matrix entries needed rather than the full product, and this is shown to solve a sparse triangle-detection problem that underlies all these hardness conjectures. Twitter reaction treated this as a major breakthrough in fine-grained complexity theory, with some tweets excitedly noting it beats a long-anticipated barrier (jokingly compared to the wait for GTA 6). A separate thread of confusion/skepticism arose from claims circulating that 'Claude discovered the algorithm,' which conflicts with the paper's actual human authorship (Alman and Vassilevska Williams) and prompted questions about how such AI-authorship claims should even be evaluated or cited.

Discussion: 7 tweets from 7 authors · @mahdi_tcs_, @evima0, @_Mira___Mira_, @Quasilocal, @nojobafterphoto +2

Medicine

Keto-Centered Telehealth Program Linked to Lower Calculated Heart Risk

This paper (no abstract available, so details here are based on Twitter discussion) reports on a cohort of roughly 90-120 adults with elevated BMI who underwent one year of telehealth metabolic care centered on a liberal-fat ketogenic diet. Using the American Heart Association's new PREVENT calculator, the authors say estimated 10-year ASCVD risk fell by about 28-30%, despite LDL cholesterol reportedly not changing. Commentators, mostly low-carb/keto-aligned clinicians and researchers, framed this as notable real-world evidence that a saturated-fat-liberal ketogenic diet can lower calculated cardiovascular risk scores even without LDL reduction; the tweets express enthusiasm rather than critical scrutiny, and no independent skepticism or methodological critique appears in the discussion shared.

Discussion: 4 tweets from 3 authors · @DoctorTro, @AdrianSotoMota, @MattCalkinsMD, @AdrianSotoMota

Computer Science

JEPA-Anything Proposes One Predictive Architecture Across Seven Scientific Domains

The paper introduces JEPA-Anything, a domain-agnostic extension of joint-embedding predictive architectures that decomposes latent representations into complementary factors (orthogonal predictive factorization) learned via dedicated pathways and recombined for prediction. The authors test it across vision, biology, clinical trajectories, control, molecular dynamics, physical fields, and weather, reporting improvements over matched JEPA baselines on all 10 dynamics tasks, a 34.8% reduction in intervention-prediction error on Interventional Pong, best molecular rollout accuracy across four systems, and experimental validation of a biologically-nominated intervention in cell cultures, organoids, and mice. They also claim latent orbital modes recover the Keplerian scaling exponent with a fitted slope near -1.5. Twitter commentary from the authors frames this as a step toward general-purpose world models, emphasizing that JEPA's predictive principle extends well beyond vision into heterogeneous scientific domains, though the tweets themselves offer promotion rather than independent scrutiny of the broad claims.

Discussion: 2 tweets from 2 authors · @Charles_Y_Wu, @LingYang_PU

Mathematics

Paper Claims Resolution of the Longstanding Kannan–Lovász–Simonovits Conjecture

The paper claims to prove that the KLS constant ψ_n is bounded above by a universal constant C for all isotropic log-concave probability measures on R^n, which would resolve the long-standing KLS conjecture; it also claims a matching universal bound on the Poincaré constant C_P(μ). The authors frame this as establishing that Cheeger's constant for such measures is uniformly bounded below, independent of dimension. Twitter commentary celebrated the result as a major breakthrough, but some noted with skepticism that the authors describe the proof as substantially LLM-assisted (crediting tools like GPT and Claude models with finding most of the mathematical ideas), which has fueled discussion about how to evaluate and verify such claims.

Discussion: 2 tweets from 2 authors · @vetohaze, @LechMazur

Computer Science

4B-Parameter Chess LM Explains Moves at Grandmaster Strength

The paper introduces Queen, a 4B-parameter chess-language model that pairs a silent expert chess encoder with an instruction-tuned LM via cross-attention, trained on a QA curriculum to extract chess concepts and refine them through an iterative 'Bellman-update'-style distillation process. The authors report the model gains over 900 Elo across seven training iterations (1782→2697), surpassing frontier LLMs on playing strength and puzzle accuracy despite being orders of magnitude smaller, while producing explanations judged nearly as coherent as GPT-5.6-Sol. Twitter discussion mainly consists of the authors announcing the paper, project site, and Hugging Face release, with one researcher expressing enthusiasm for applying the same approach to other strategy games like shogi; no substantive critiques of the methodology appeared in the discussion.

Discussion: 2 tweets from 2 authors · @AdithyaNLP, @AIcia_Solid

Mathematics

Mathematicians Announce Proof of the Kannan–Lovász–Simonovits Conjecture

The paper, by Pierre Bizeul, Boaz Klartag, and Joseph Lehec, claims to resolve the KLS conjecture, a long-standing problem in high-dimensional convex geometry closely tied to Bourgain's slicing problem and the variance conjecture. According to the abstract, the key new ingredient is a recent criterion from Song and Zhang involving high-order derivatives of 'tilt averages,' which the authors use to close the final gap in the conjecture. Twitter users were excited, congratulating the three mathematicians on what would be a major result in asymptotic geometric analysis. Some commentary noted a separate, LLM-assisted paper circulating around the same time that also claims to resolve KLS—credited in its own text to tools like 'GPT-6 Astra' and 'Claude Fable 5'—which several users flagged skeptically as distinct from, and less credible than, the Bizeul–Klartag–Lehec proof.

Discussion: 2 tweets from 2 authors · @letwinbrayden, @LechMazur

Social Science

Survey Explores Virtual Sex Culture Among Young Japanese VRChat Users

This study presents results from a web survey of Japan's VRChat community, investigating claims that virtual sex has become an established cultural practice among young users of the platform, set against a backdrop of reported declines in real-world sexual interest and activity among Japanese youth. The researchers aim to characterize how widespread and normalized these virtual intimate practices are within the predominantly young VRChat user base. Methodologically, the paper relies on self-reported survey data collected from community members rather than behavioral or platform-level data. On Twitter, commentary was largely amused and curious rather than critical, with users noting the novelty of serious academic research into an unusual topic. One commenter speculated (without firsthand knowledge) about the mechanics of consent in virtual sexual encounters, such as whether interactions involve explicit in-app prompts for requesting and approving actions.

Discussion: 2 tweets from 2 authors · @sironeko114_vrc, @zukky_4649

Social Science

Manga About a Dialysis-Avoiding Patient Used to Teach Bioethics

This presentation, delivered at the 58th Annual Meeting of the Japan Society for Medical Education, describes a bioethics course that repeatedly used a manga about a man with severe diabetes who resists treatment and later begins dialysis, prompting students to grapple with patient autonomy, family involvement, and clinicians' responsibilities. The authors report an exploratory analysis linking students' narrative immersion in the manga to their final report scores among non-medical undergraduates. The record is an English-language version of the slides, with manga artwork and dialogue excluded from the CC BY license and credited to the original creator. Twitter discussion is minimal, with the manga creator celebrating the work's archiving on Zenodo as entry into international academic literature, offering no substantive critique of the methodology or findings.

Discussion: 1 tweets from 1 authors · @masaki_sakura

Medicine

Kidney's Proximal Tubule Cells Contribute Significantly to Glucose Production

No abstract is available for this paper, so this summary is based solely on Twitter discussion. The cited work appears to concern renal gluconeogenesis, specifically the role of proximal tubule S3 segment cells in producing glucose—a function traditionally attributed almost exclusively to the liver. A tweet highlighting the paper notes that the kidney normally contributes an estimated 20-25% of endogenous glucose production, rising to 40-50% during fasting, challenging the common assumption that only the liver performs gluconeogenesis. The commentary frames this as a corrective to a widespread misconception rather than raising any specific criticism of the paper itself.

Discussion: 1 tweets from 1 authors · @AcidoRinon

Biology

First Nationwide Distribution Map of 16 Tick Species in Japan

This study published in Parasites & Vectors presents what the authors describe as the first comprehensive nationwide distribution mapping of tick species in Japan, modeling the predicted ranges of 16 tick species and examining how these distributions relate to the occurrence of tick-borne diseases. No abstract was available, so this summary is based on discussion only. On Twitter, a researcher highlighted that Japan previously lacked any nationwide survey of tick distribution despite the fact that tick-borne disease outbreaks are thought to be strongly shaped by vector tick distribution, framing this paper as filling that gap.

Discussion: 1 tweets from 1 authors · @Hyla_tanuki

Medicine

Nature Reviews Disease Primers Publishes IBS Review with Treatment Algorithm

This entry is a Disease Primers article on irritable bowel syndrome; no abstract was available, so this summary is based on the Twitter discussion only. The tweet highlights a comprehensive treatment algorithm for IBS management, sharing a figure from the primer rather than offering independent commentary or critique.

Discussion: 1 tweets from 1 authors · @AmbrizMau

Computer Science

ACE-Step: An Open Music Generation Model Aiming to Be 'Stable Diffusion for Music'

ACE-Step is an open-source foundation model for music generation that combines diffusion-based synthesis with a deep compression autoencoder and a lightweight linear transformer, aligning semantic representations during training for fast convergence. The authors claim it generates up to 4 minutes of music in 20 seconds on an A100 GPU—15x faster than LLM-based competitors—while improving musical coherence and lyric alignment, and supporting tasks like voice cloning, lyric editing, and remixing. The team frames it as a general-purpose foundation model rather than a single end-to-end pipeline, intended to enable further subtask development. Twitter commentary (in Japanese) raised concerns about training data transparency, noting that since the model's training data is undisclosed and opt-out isn't offered, it is—critics argue—functionally indistinguishable from the 'unauthorized learning' music-generation AI that some creators have objected to.

Discussion: 1 tweets from 1 authors · @30years_over

Mathematics

Random-matrix analysis resolves Trefethen's conjecture on Gaussian elimination stability

The paper studies the growth factor of Gaussian elimination, the key quantity controlling numerical stability during the solution of linear systems. The authors prove tight estimates for growth factors under Haar-orthogonal preconditioning and without pivoting, and most notably show that for an n×n Gaussian matrix under partial pivoting, the growth factor is rarely much larger than √n — confirming a long-standing conjecture by Nick Trefethen. The same techniques yield an improved smoothed-analysis bound for growth under partial pivoting. Twitter commentary (from Ethan Epperly) highlighted this as a notable resolution of a well-known open problem in numerical linear algebra, with no substantive criticism raised in the discussion.

Discussion: 1 tweets from 1 authors · @ethanepperly

Computer Science

Full-Bandwidth Transformer Feeds Back Hidden States, Not Just Tokens

The paper argues that standard autoregressive transformers only pass the sampled token back to the input at each decoding step, discarding the rich top-layer hidden state and leaving a narrow 'vertical' feedback channel compared to the broad horizontal attention over past tokens. The authors propose fusing the previous top-layer hidden state with the new token embedding via a gated linear unit, letting unverbalized computation re-enter the stack with a fresh depth budget, while keeping the same architecture, KV cache, and LM objective. Trained at 1B parameters on up to 400B tokens with a scheduled multi-pass training scheme, this 'latent feedback' improves validation loss, few-shot evals, math/coding generation, and instruction-tuned performance, roughly matching standard transformers trained on 1.5x more tokens with negligible extra decoding cost. The tweet discussing this is from one of the authors announcing a revision of the paper, highlighting new sections on debugging training runs, post-training/KV cache replay, and an alternative fusion gate, alongside an unofficial code release — it's promotional/author commentary rather than independent critique, so no external skepticism has surfaced yet.

Discussion: 1 tweets from 1 authors · @xidulu

Biology

9-Million-Year-Old Shark Fossil Caught With Its Last Meal Still Inside

Researchers describe a rare, well-preserved partial skeleton of the extinct lamniform shark Cosmopolitodus hastalis from the late Miocene Pisco Formation in Peru, including fish remains (a pilchard, Sardinops sp.) preserved in the abdominal region as fossilized stomach contents. Tooth morphology indicates the shark was a juvenile, making this the first direct evidence of piscivory in young C. hastalis, consistent with feeding patterns seen in modern juvenile lamnids like great whites and makos. The find also reinforces the Pisco Formation's status as an exceptional fossil site (Fossil-Lagerstätte) shaped by early mineralization and highlights the importance of schooling pilchards in Miocene coastal food webs.

Discussion: 1 tweets from 1 authors · @PaleontologieFr

Computer Science

LLMs' Stated Confidence Levels Track Their Internal Probabilities

The paper investigates whether large language models' verbalized confidence (stated in words or numbers) actually reflects their internal sampling probabilities, or whether the two are independently derived from different training signals (frequency statistics vs. explicit probabilistic assertions). By systematically manipulating both distributional frequencies and asserted uncertainty statements in training and in-context data, the authors find that both readouts are shaped by both uncertainty sources, and crucially that verbalized and internal probabilities align more closely than would be expected if they were tracking these sources independently—suggesting verbalized confidence can serve as a genuine probe of a model's internal distribution. The researchers frame this as resolving an open question about whether we can trust LLMs' stated uncertainty as a proxy for their actual internal beliefs. The Twitter discussion consists mainly of the lead author's thread introducing the paper's motivating question and framing it as a surprising result, with the promise of further findings to follow; no independent critical commentary has yet emerged in the visible discussion.

Discussion: 1 tweets from 1 authors · @sineadwilliamso

Computer Science

Survey Maps Open Problems Facing Mechanistic Interpretability Research

This forward-looking review surveys the field of mechanistic interpretability—efforts to reverse-engineer the computational mechanisms inside neural networks—and catalogs the open problems blocking further progress. The authors argue that advances require both conceptual and practical methodological improvements, clearer thinking about how to target these methods toward specific scientific and engineering goals, and attention to socio-technical challenges shaping the field's trajectory. Twitter discussion was minimal, with one user noting the 80-page survey was recommended by researcher Neel Nanda on the 80,000 Hours podcast, framing it as a useful orientation document for those studying the field.

Discussion: 1 tweets from 1 authors · @neural_avb

Chemistry

Generative Models Learn to Sample the Grand Canonical Ensemble

This paper extends Boltzmann generators—generative models that sample equilibrium configurations—to the grand canonical ensemble, where particle number fluctuates alongside configuration. The authors propose two designs: one conditions a variable-size generator directly on chemical potential, while the other factorizes the distribution into a particle-number component and a canonical Boltzmann density, enabling reuse of existing Boltzmann generators and a tractable likelihood for importance sampling. They validate the approach on a Lennard-Jones fluid and methane adsorption in a zeolite, showing it generalizes across chemical potentials and can be corrected via self-normalized importance sampling or grand canonical Monte Carlo.

Discussion: 1 tweets from 1 authors · @yoko_materialDX

Physics

Precomputed Green's Functions Speed Up Photonic Inverse Design 1000x

The paper extends the Precomputed Numerical Green Function (PNGF) method to nanophotonic inverse design, replacing repeated full-wave electromagnetic simulations with a single precomputation step after which objective functions and gradients scale linearly with design-region size. Combined with a low-rank matrix update technique, this yields sub-millisecond per-iteration costs and reportedly more than three orders of magnitude speedup over conventional methods, enabling new photonic devices to be designed from scratch in seconds without sacrificing accuracy. Twitter discussion of the paper was limited to a single link share with no substantive technical commentary.

Discussion: 1 tweets from 1 authors · @QuentinWach

Social Science

Paper Argues AI Labs Are Rewriting Safety Norms to Fast-Track Military AI

The paper argues that in the absence of agreed-upon AI risk thresholds, AI industry labs and 'AI safety' organizations have pushed risk tolerances skewed by arms-race narratives and speculative existential risk framing, effectively seizing control over life-or-death risk determinations from democratic processes. The authors term this 'safety revisionism'—replacing established safety engineering methods and terminology with vaguer alternatives—and warn this trajectory is driving a race-to-the-bottom in evaluating foundation models for military use, undermining rather than protecting US national security and compliance with international humanitarian law. The authors draw on historical risk-threshold precedents from nuclear and aviation safety engineering to make their case. Discussion centered on an exchange where author Heidi Khlaaf (a safety engineer with nuclear and aviation background) accused another commentator of dishonestly engaging with her work, pointing to a series of related readings on AI risk governance, military AI, and safety-washing critiques as context for the paper's argument.

Discussion: 1 tweets from 1 authors · @HeidyKhlaaf