DeepMind introduces AlphaProof Code for Signal cryptography
DeepMind says AlphaProof Code used AI agents and human review to verify Signal cryptographic code, with verification materials planned for release.
Published entries filed under “Research” in News, newest first by publication date on this site.
39 entries
DeepMind says AlphaProof Code used AI agents and human review to verify Signal cryptographic code, with verification materials planned for release.
NASA confirmed both ESCAPADE spacecraft completed powered Earth flybys on November 7 and 9, 2026. Mars orbit insertion is planned for September 2027.
MIT CSAIL reported on Nov. 2 that GPT-6 Astra changed 0.016% of Qwen3.5-0.8B’s weights; MATH-500 rose from 7.9% to 48.1%.
Wired’s Nov 8 report maps the world-model field: models that simulate physics for robotics, self-driving and games, and LeCun’s startup past $1B raised.
Menlo Ventures' 2026 report: 89% of US enterprise open-source model spend goes to Chinese models — Kimi 61%, Llama 7%. Survey of ~400 US decision-makers.
OpenAI's autonomous pipeline reproduced 1,724 Linux kernel bugs and wrote 948 working root exploits at about $1,279 each; 61 patched upstream in 10 days.
DeepSeek's dsh-superagent-prototype (MIT, Oct 31) listens to GitHub and Feishu, takes work proactively, and self-reports a simulated p95 of 15.7 minutes.
GitHub's Octoverse 2026: over 40% of pull requests now start as an agent draft (17% in March), and 17.9B pushes in 12 months beat 2024 and 2025 combined.
OpenAI claims all-dimensions Kakeya, a Lean-verified ETH refutation and QCD existence, released via external shepherds a day after its panel objected.
SkeLAM (arXiv 2610.31474) grounds latent actions in robot skeletons: ~10 hours of robot data, no extra large-scale pretraining, 7 of 8 DROID metrics first.
Cammarata's October 15 essay: the polysemantic-neuron evidence was misread; ~99% of InceptionV1's last layer is interpretable. A peer raises five objections.
Sebastian Raschka put his full from-scratch reasoning-model course on YouTube for free on Oct 24: 8 episodes, 11+ hours, PyTorch code under Apache-2.0.
A dermatologist's open test on 200 biopsied lesions: a fine-tuned 2015 Inception v3 scored 62% top-1 vs 42% for GPT-6 Astra and Claude Fable 5.1.
Per Google DeepMind (arXiv, Oct 22), softmax and cross-entropy in PyTorch, JAX and TensorFlow hide two floating-point bugs; a log-space rewrite fixes both.
CAICT's token-service report (Oct 23): China burned ~500T tokens a day in June; normalized prices nearly doubled to 13.46 yuan, overseas models ~117.8 yuan.
In a forced pick across 479 midterm races, Grok 4.7 chose Republicans ~92% of the time, GPT-6.1 Sol Democrats ~97%; Grok's API carried ~1,240 hidden tokens.
1,946 GPT-Rosalind agents, one per gene, reanalyzed a mouse whole-brain CRISPR atlas in 32 minutes; OpenAI, PerturbAI and Scripps published the reports.
RoboColiseum put four frontier models through one Agent2Act harness on Oct 20: Opus 5.5 leads at 76.28, while manipulation still trails specialist policies.
An open-access Scientific Reports paper runs four vibration sensors and three LLM/VLM agents over a 16th-century minaret, scoring six zones separately.
Anthropic's Oct 16 report: junior technical staff fell from 21% to 14%, junior hires from 27% to 9%, and agent sessions ran 6x staff hours.
Manifold AI's R²SI scored 46.58 and 39.35% on RoboDojo's agent board Oct 17, running the same GPT-6 Astra it beat; weights are still private.
Anthropic says an internal research model traced the five NIST prime-curve seeds to lrand48 output seeded with 1997 timestamps; the $36,864 bounty went to ISRG.
Two Nature protease-design papers appeared October 14; Huxiu reports 13 of 69 cysteine designs were active and discusses a second metalloprotease study.
Shanghai AI Lab announced a 3D Ghys conjecture proof on Oct 15. The paper credits Intern-S2 with helping develop ideas; PKU formalized its theorems in Lean.
An arXiv paper proposes CNF-MCE tokenization: 26.7% better compression than o200k_base at 250K vocab, 67.0% at a million, downstream on par.
LangChain's Oct 14 post turns context-rot fixes into Deep Agents built-ins: results offloaded to files, summaries at 85% window use, memory one argument away.
MIT, ECMWF and Microsoft published probabilistic bias correction in Nature on Oct 14: 2-6-week forecast skill doubles, gains skew tropical, code open source.
Scott Helme rebutted the viral "Passkeys are a trap" essay on Oct 14: synced passkeys return attestation "none", and Evilginx relays TOTP in real time.
Anthropic's red team ran a 26B model across 15 $55 TV boxes; GLM-5.2 hit all 8 sandbox nodes in 4 of 5 trials, and small models may cross the line by end-2026.
Safe Rust can bypass a struct's lifetime bounds via covariant fields plus Drop, causing use-after-scope — an 11-year-old hole, triaged P-high, unfixed.
DeepMind's AlphaProof Nexus solved 9 of 353 open Erdős problems and 44 of 492 OEIS conjectures; the work cleared peer review at Science on October 8.
Yann LeCun said in Seoul on Oct 12 that AI's next breakthrough will come from world models, not bigger LLMs; his H-JEPA paper lifts AntMaze from 18% to 73%.
Agent leaderboard data through Oct 8 puts Claude Opus 5.5 (Max) first at 15.53% and Opus 5.5 (High) second; DeepSeek, Kimi and Tencent models rank 17th–19th.
US software developer postings are up ~15% since Claude Code launched, but 71% of the growth is senior roles and junior developer employment is down ~20%.
The openai/math GitHub release carries Apache-2.0 manuscripts with Lean proofs, over IAS advisory-group objections.
Anthropic says 950 Claude agents found ART, a reverse-transcriptase family resembling CRISPR; Wired notes the function is unconfirmed and unreviewed.
Open-source SparkDiffusion pushes 720P-14B video generation to 265x on one RTX 5090 via sparse attention, distillation and FP8; VBench-2.0 drops 0.5 points.
In a guest post published by Anthropic, physicist Matt von Hippel reports that Claude, running largely unsupervised for days on Claude Science, computed the nine-loop six-particle amplitude in planar N=4 super-Yang-Mills. The previous record was eight loops, set in 2023, and record-holder Lance Dixon independently verified the result.
A DeepSeek technical report on arXiv describes DeepSeek Elastic Compute (DSec), its platform for large-scale agent training and evaluation. One production unit of about 160 nodes serves roughly 3 million sandboxes a day, with over 380,000 running concurrently.