Openshelf: research on authors' own websites https://chenxiachan.github.io/openshelf/ 2026-11-07T01:05:49Z App Store 2026 in numbers: AI, pricing, abandoned apps https://obelisk.club/research/app-store-2026 2026-11-06T00:00:00Z obelisk.club Tallies AI features and in-app purchases across 178,637 App Store apps. How scientists use Claude https://www.anthropic.com/research/how-researchers-use-claude 2026-11-06T00:00:00Z anthropic.com Uses a privacy-preserving method to study how scientists use Claude. Teaching Qwen Why https://www.lesswrong.com/posts/35yw2qGK7naFLAvcq/teaching-qwen-why-1 2026-11-06T00:00:00Z lesswrong.com Experiments on teaching Qwen the reasons behind behavior for alignment. Experiments That Could Win https://aspirationalneuroscience.org/experiments/ 2026-11-06T00:00:00Z aspirationalneuroscience.org Concrete experimental proposals for winning the Memory Decoding Challenge. Automated "Inverse" Design of Grating Couplers for Co-Packaged Optics https://engineering.flexcompute.com/articles/inverse-design-grating-couplers/ 2026-11-06T00:00:00Z engineering.flexcompute.com Describes automated inverse design of grating couplers for co-packaged optics using shape optimization and adjoint methods. Codetations: Rich, Granular, and Maintainable Annotations for Evolving Code https://ztatlock.net/pubs/2026-uist-codetations/ 2026-11-06T00:00:00Z ztatlock.net Research project page for Codetations, a VSCode extension for persistent code annotations. Shaping and Monitoring Models Through Their Internal Representations https://www.goodfire.com/research/mechanistic-shaping 2026-11-06T00:00:00Z goodfire.com Research on shaping and monitoring models through their internal representations. The Secret Leak Test: how well GateBolt kept Hodor's password, and where it didn't https://hodor.gatebolt.com/blog/secret-leak-test 2026-11-06T00:00:00Z hodor.gatebolt.com Test of GateBolt prompt injection filter against AI attackers trying to steal a password. Lighthouse 13.6: how we made PageSpeed Insights more accurate for Framer sites https://www.framer.com/blog/lighthouse-13-6/ 2026-11-06T00:00:00Z framer.com Framer engineering breakdown of Lighthouse 13.6 changes fixing scoring of preloaded code. pico-WAM—Counterfactual Prediction and Planning Error https://www.keivalya.com/pico-wam/ 2026-11-06T00:00:00Z keivalya.com Presents pico-WAM, a compact world action model, evaluating counterfactual prediction and planning. I Spent $1000 Making LLMs Play Competitive Pokémon https://blog.doubleword.ai/llms-play-pokemon 2026-11-06T00:00:00Z blog.doubleword.ai Experiment spending $1000 on LLMs playing competitive Pokemon against each other. 594,101 knocks: who reads my site, and what AI crawlers can't see https://www.shuffle-on.com/blog/ai-crawlers-read-my-site 2026-11-06T00:00:00Z shuffle-on.com Log analysis of OpenAI crawler hitting an SSO login page 594,101 times. The Day a Test Deleted My Home Directory https://lucianolupo.com/notes/The%20Day%20a%20Test%20Deleted%20My%20Home%20Directory 2026-11-06T00:00:00Z lucianolupo.com Postmortem of a coding agent mutation test that deleted a home directory. Turning the HEIC decoder into a SHA-256 computer https://www.irisc2.com/blog/heic-hashquine 2026-11-06T00:00:00Z irisc2.com Technical demonstration turning a HEIC decoder into a SHA-256 computer via a crafted image sequence. Using decision models to speed up design rule violation repair in chip design https://engineering.flexcompute.com/articles/decision-models-drc-repair/ 2026-11-06T00:00:00Z engineering.flexcompute.com Uses decision models to speed up design rule violation repair in chip design. Harness 2: executing agent code as it streams https://www.framer.com/blog/harness-2/ 2026-11-06T00:00:00Z framer.com Framer introduces Harness 2, executing JavaScript as it streams to improve agent efficiency and reduce costs. METR: Independent Assessment of AI R&D Acceleration at Anthropic in October 2026 https://www.lesswrong.com/posts/N5AmYkRks5CqAxvBg/metr-independent-assessment-of-ai-r-and-d-acceleration-at 2026-11-06T00:00:00Z lesswrong.com METR report finds Anthropic's AI R&D acceleration doubled by October 2026 due to AI automation. Programming in the Machine Code Space: LLMs can read & write executables byte by byte https://sgiannopoulos.github.io/blog/machine-code-programming/ 2026-11-06T00:00:00Z sgiannopoulos.github.io Research shows LLMs can read and write machine code directly without disassemblers. Can LLMs redesign themselves as they train? https://www.doubleai.com/research/can-llms-redesign-themselves-as-they-train 2026-11-06T00:00:00Z doubleai.com DoubleAI research enables AI agents to redesign neural network architectures during training. VFlash: Enabling System-1 VLMs for Physical AI https://research.nvidia.com/labs/lpr/vflash/ 2026-11-06T00:00:00Z research.nvidia.com NVIDIA proposes VFlash, accelerating Physical AI inference via parallel visual planning. The components of agent memory https://www.anthropic.com/research/components-of-agent-memory 2026-11-06T00:00:00Z anthropic.com Anthropic research shows how Claude organizes memory and corrects errors via a 'dreaming' mechanism. DeepFinance: Can AI diligence a $100MM deal? https://www.juliusresearch.ai/blog/deep-finance 2026-11-06T00:00:00Z juliusresearch.ai DeepFinance is a benchmark and leaderboard measuring AI performance on real-world private equity and investment banking workflows. Computer-Use Agents Are Embodied Generalists https://pranjal2041.github.io/cua-embodied/ 2026-11-06T00:00:00Z pranjal2041.github.io Research shows a single computer-use agent ranking first, second, third, and sixth across four embodied AI competitions. The Language Server Tax: More Cost Without Better C/C++ Code Repair. ByteAsk https://byteask.ai/research/lsp-claude-code-cpp 2026-11-05T00:00:00Z byteask.ai Tests how the clangd LSP plugin affects Claude Code's ability to resolve issues. StateBench: Testing how language models track state changes over long documents https://huggingface.co/blog/allenai/statebench 2026-11-05T00:00:00Z huggingface.co Proposes StateBench, a benchmark of how language models track state changes in long documents. Kimi K3.1 Pro writes fake OpenAI ciphertext https://www.peterstran.com/projects/kimi-fernet/ 2026-11-05T00:00:00Z peterstran.com Documents Kimi K3.1 Pro outputting fake OpenAI ciphertext. Larger Tokenizers Suffer More Multi-Epoch Regret https://wenhaochai.com/blogs/vocab-multi-epoch.html 2026-11-05T00:00:00Z wenhaochai.com Experiments show that larger tokenizers incur higher regret in multi-epoch training. On the Limits of Test-Time Compute for Self-Evaluation https://anirudhb11.github.io/blog/scorpio/ 2026-11-05T00:00:00Z anirudhb11.github.io Analyzes the limits of using test-time compute for self-evaluation. Hiding a Rule Inside Another Model's Weights https://jonneslin.github.io/weight-space-trigger-en.html 2026-11-05T00:00:00Z jonneslin.github.io Manually edits about 0.1% of Qwen3.5-0.8B's weights to implant a rule. Why the search for particle dark matter is really hopeless, part 2 https://tritonstation.com/2026/11/05/why-the-search-for-particle-dark-matter-is-really-hopeless-part-2/ 2026-11-05T00:00:00Z tritonstation.com Argues that the search for particle dark matter is nearly hopeless given MOND phenomena (part 2). The test that deleted my home folder https://www.shuffle-on.com/blog/the-test-that-deleted-my-home-folder 2026-11-05T00:00:00Z shuffle-on.com Postmortem of an AI coding assistant test that deleted a home folder while demonstrating a path-traversal bug. Forget about credit assignment in long-horizon agentic post-training https://qiyaoliang.com/blog/es-vs-pg-long-horizon/ 2026-11-05T00:00:00Z qiyaoliang.com Argues credit assignment is unnecessary in agentic post-training, comparing evolution strategies with policy gradient. The Exploit Was the Bottleneck https://www.a16z.news/p/the-exploit-was-the-bottleneck 2026-11-05T00:00:00Z a16z.news Data analysis of 23 years of vulnerability data on AI and the patch gap. Rethinking Abstractions in the Agentic Era: Towards a DSL and Compiler Foundation for Agentic Kernel Development https://tvm.apache.org/2026/11/03/rethinking-abstractions-in-the-agentic-era 2026-11-05T00:00:00Z tvm.apache.org TVM community report on using coding agents to build DSL and compiler foundations. GPT-Live vs GPT-Realtime: Voice AI Trade-offs https://agentunicorn.ai/research/gpt-live-voice-ai-tradeoffs 2026-11-05T00:00:00Z agentunicorn.ai Comparison of GPT-Live, GPT-Realtime and streaming voice pipelines on delegation, guardrails and reliability. AgentSight: System-wide AI agent profiling and monitoring with eBPF https://eunomia.dev/agentsight/ 2026-11-05T00:00:00Z eunomia.dev Presents AgentSight, an eBPF-based system-wide profiling and monitoring tool for AI agents. VibeSys on Trainium: Agents Build and Tune Qwen3.8 Serving Engines on AWS Trainium2 https://syfi.cs.washington.edu/blog/2026-11-04-vibesys-trainium/ 2026-11-05T00:00:00Z syfi.cs.washington.edu Evaluates agents building and tuning Qwen3.8 serving engines on AWS Trainium2. How to Train a Swarm https://www.gr.inc/blog/how-to-train-a-swarm 2026-11-05T00:00:00Z gr.inc Describes methods and results for training agent swarms. Declarative Dragging https://declarative-dragging.github.io/ 2026-11-04T00:00:00Z declarative-dragging.github.io Proposes declarative dragging interaction; project page of a UIST 2026 best paper. The SBND collaboration releases first neutrino results https://news.fnal.gov/2026/10/the-sbnd-collaboration-releases-first-neutrino-results/ 2026-11-04T00:00:00Z news.fnal.gov The SBND collaboration releases its first neutrino interaction results. AbMaze: Evaluating frontier AI agents on protein fitness landscapes https://valenlabs.com/blog/abmaze 2026-11-04T00:00:00Z valenlabs.com Proposes AbMaze, a benchmark of antibody fitness landscapes for evaluating agents. Investigating vault particles with Claude https://www.anthropic.com/research/investigating-vault-particles-with-claude 2026-11-04T00:00:00Z anthropic.com Uses Claude to help resolve the first 3D structure of a prokaryotic vault particle. Advancing the Quality-Efficiency Frontier for Coding AI on Windows PCs - Microsoft Foundry on Windows Blog Coding AI on Windows: The Quality–Efficiency Frontier https://devblogs.microsoft.com/foundry-on-windows/coding-ai-on-windows-pcs-quality-efficiency/ 2026-11-04T00:00:00Z devblogs.microsoft.com Uses low-bit quantization and speculative decoding to improve local coding AI efficiency on Windows. Serving GLM-5.3-Flash on one Trainium2 box with NKI layer megakernels https://runanywhere.ai/blog/glm-5-3-flash-trainium2-nki-megakernels 2026-11-04T00:00:00Z runanywhere.ai Serves GLM-5.3-Flash on a single Trainium2 machine and measures throughput. I Tried to Break Weight Programming https://flecart.github.io/posts/2026-11-03-i-tried-to-break-weight-programming/ 2026-11-04T00:00:00Z flecart.github.io Replicates weight programming experiments and analyzes their complexity limits. MCPhantom: AI Agents Installing Deprecated MCP Servers https://www.glow.io/blogs/mcphantom-ai-coding-agents-are-installing-vulnerable-deprecated-mcp-servers-from-training-data 2026-11-04T00:00:00Z glow.io Reports that AI coding agents install deprecated MCP servers with known vulnerabilities. Can Advertising Fund Free AI? ChatGPT Ads Research https://ecommerceresearch.org/chatgpt-ads/ 2026-11-04T00:00:00Z ecommerceresearch.org Working paper using evidence from ChatGPT ad rollout to study whether advertising can fund free AI. 20 Percent of All ‘Ok’ Web Responses AI Gets Are Not Okay https://www.unite.ai/20-percent-of-all-ok-web-responses-ai-gets-are-not-okay/ 2026-11-04T00:00:00Z unite.ai Author experiments with AI-assisted scraping finding many HTTP OK responses are not truly valid. Regulation by AI: Replaying EPA’s Glyphosate Decisions https://reckoningscience.org/regulation-by-ai-replaying-epas-glyphosate-decisions/ 2026-11-04T00:00:00Z reckoningscience.org Experiment having five AI models replay 50 years of EPA glyphosate decisions, finding none more lenient. Making flawed benchmarks useful https://blatent.ai/blog/making-flawed-benchmarks-useful 2026-11-04T00:00:00Z blatent.ai Evaluation of extraction accuracy that reveals benchmark flaws and shows how to use them. Jane Street Blog - A study of sequence weighting at scale https://blog.janestreet.com/a-study-of-sequence-weighting-at-scale/ 2026-11-04T00:00:00Z blog.janestreet.com Jane Street study of scaling laws for data weighting across language models. Weight-space programming a neural quantum state https://paulg.info/2026/11/04/quantum-weight-space/ 2026-11-04T00:00:00Z paulg.info Case study using Opus to diagnose and repair a neural quantum state via weight-space programming. Learnings from Auditing Public Evaluations https://www.primeintellect.ai/blog/auditing-public-evaluations 2026-11-04T00:00:00Z primeintellect.ai Audit of public LLM evaluations finding broken graders and ambiguous tasks. Progress on provable inference for large language models https://www.anthropic.com/research/provable-inference 2026-11-03T00:00:00Z anthropic.com Anthropic reports progress on zkSNARKs for provable LLM inference. The shelf life of LLM findings https://zou-group.github.io/llm-claims-shelf-life/ 2026-11-03T00:00:00Z zou-group.github.io Replicates 50 published LLM findings across 20 models. The Nordic token tax https://research.klang.ai/papers/the-nordic-token-tax 2026-11-03T00:00:00Z research.klang.ai Measures the extra token cost that 12 frontier models charge for non-English prompts. LLM grounding tutorial: stop AI agents guessing prices https://blog.apify.com/llm-grounding-with-live-data/ 2026-11-03T00:00:00Z blog.apify.com Tests AI agent price hallucination and live-data grounding on 21 product pages. Krea 3 Technical Report https://www.krea.ai/blog/krea-3-technical-report 2026-11-03T00:00:00Z krea.ai Technical report for the Krea 3 image editing model. Rotary Position Encodings for Graphs https://idlemachines.co.uk/essays/rope-for-graphs 2026-11-03T00:00:00Z idlemachines.co.uk Proposes rotary position encodings for graphs. A Mathematical Framework for Transformer Circuits https://transformer-circuits.pub/2021/framework/index.html#summary-of-results 2026-11-03T00:00:00Z transformer-circuits.pub A mathematical framework for transformer circuits, a foundational paper of mechanistic interpretability. Quantity vs quality in pottery is a lie https://www.lesswrong.com/posts/LrAqM4bGmAoA79rP3/quantity-vs-quality-in-pottery-is-a-lie 2026-11-03T00:00:00Z lesswrong.com Traces the origin of the pottery quantity versus quality story and checks its facts. Autonomous Agents Need Mandatory Access Control https://www.legiscope.com/blog/autonomous-agents-mandatory-access-control.html 2026-11-03T00:00:00Z legiscope.com Proposes a mandatory access control framework for autonomous agents, modeled on FreeBSD. The 3D Printed Database https://www.amplifypartners.com/blog-posts/the-3d-printed-database 2026-11-03T00:00:00Z amplifypartners.com Explores automatically generating custom database systems from a domain ontology. “This Is Fine”: How Clean Rooms Train Agents to Close the Books on Fire https://archytas-ai.com/blog/this-is-fine 2026-11-03T00:00:00Z archytas-ai.com Studies how RL environments lead agents to falsely report task completion. Same Model, Different Harness: CUGA, Hermes, and Deep Agents https://huggingface.co/blog/ibm-research/cuga-harness 2026-11-03T00:00:00Z huggingface.co Compares three agent harnesses: CUGA, Hermes, and Deep Agents. The Hitchhiker’s Guide to Training on Compute Islands https://www.coreauto.com/hitchhikers-guide-to-training-on-compute-islands 2026-11-03T00:00:00Z coreauto.com A position on fault-tolerant distributed training across data centers, with sparse MoE experiments. A Direct Construction of SZ Sequence Matrices https://andrew-helmer.github.io/sz-matrix-construction/ 2026-11-03T00:00:00Z andrew-helmer.github.io Gives a direct construction of SZ sequence matrices. 1.26 Million Data Center Satellites? Orbit Says No https://blog.budget-engineer.com/space-data-centers-have-you-looked-up-lately 2026-11-03T00:00:00Z blog.budget-engineer.com Analysis of whether 1.26 million proposed data center satellites are feasible given orbital crowding. Quadrillion-parameter transformers and how to train them https://qlabs.sh/research/parameters 2026-11-03T00:00:00Z qlabs.sh Presents a scaling law and extrapolates training recipes for quadrillion-parameter transformers. TIS-100 Bench: LLMs Push the Pareto Frontier in Programming Puzzle Games https://tis-100.wincon.dev/ 2026-11-03T00:00:00Z tis-100.wincon.dev Benchmark of 16 LLMs on TIS-100 puzzles finding 133 new Pareto frontier points. Eval-production behavior gaps in agent safety and the role of eval-awareness https://alignment.openai.com/eval-production-behavior-gap/ 2026-11-03T00:00:00Z alignment.openai.com Compares agent behavior in safety evaluations versus production and studies eval-awareness. The Inference Stack Behind OCRia https://www.llamaindex.ai/blog/the-inference-stack-behind-ocria 2026-11-03T00:00:00Z llamaindex.ai Describes inference optimizations in the OCRia stack, including quantization and decoding speedups. StateBench: Testing how language models track state changes over long documents https://allenai.org/blog/statebench 2026-11-03T00:00:00Z allenai.org Introduces StateBench, evaluating how language models track state changes across long documents. ZHA: Zyphra Hierarchical Attention https://www.zyphra.com/our-work/zyphra-hierarchical-attention 2026-11-03T00:00:00Z zyphra.com Introduces ZHA, a hierarchical sparse attention method, with experiments. Formalizing PTX in Lean to Prove GPU Kernel Correctness https://sf-tensor.com/engineering/formalizing-ptx 2026-11-03T00:00:00Z sf-tensor.com Formalizes Nvidia PTX in Lean to prove GPU kernel correctness, with accompanying paper. Userland Linux scheduler in Rust-faster than time https://kapteko.com/userland-linux-scheduler-in-rust/ 2026-11-03T00:00:00Z kapteko.com Describes a userland Linux scheduler in Rust that makes CPU-bound programs run faster than wall time. Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights https://rikiyat.github.io/blog/weight-space-programming/ 2026-11-02T00:00:00Z rikiyat.github.io Finds that frontier LLMs can directly read and write neural network weights to improve Qwen. bf16 might be freezing your weight updates https://lacuna.tiptreesystems.com/paper/bf16-might-be-freezing-your-weight-updates/art_487b1ed949e745a6ae0c42815b8d3cba 2026-11-02T00:00:00Z lacuna.tiptreesystems.com Analyzes how AdamW updates can be frozen with bf16 weights. GDPx: Measuring Frontier Model Capabilities Across the Most Valuable Domains https://newsletter.semianalysis.com/p/gdpx-measuring-frontier-model-capabilities 2026-11-02T00:00:00Z newsletter.semianalysis.com Proposes GDPx, a benchmark that evaluates frontier models across several high-value domains. Accelerating Attention for Ultra-Long Context https://prefrontal.ai/news/lsa/ 2026-11-02T00:00:00Z prefrontal.ai Proposes LSA, a hierarchical clustering attention, to speed up ultra-long context. Will AI compute stay tight? https://hxu296.github.io/ai-compute-landscape/ 2026-11-02T00:00:00Z hxu296.github.io Analyzes AI compute rents and commitments to judge whether neoclouds are a bubble. WeVisCut: The Whole Editing Room in One Agent https://tencent.github.io/WeVisCut 2026-11-02T00:00:00Z tencent.github.io Project page for WeVisCut, which uses an agent to edit long videos. Making CUDA Checkpointing 15x Faster on a B300: A Worklog https://maknee.github.io/blog/2026/Cuda-Checkpointing/ 2026-11-02T00:00:00Z maknee.github.io A worklog on making CUDA checkpointing 15 times faster on a B300. ZeroBench https://zerobench.github.io/ 2026-11-02T00:00:00Z zerobench.github.io Proposes ZeroBench, a visual benchmark that is extremely hard for multimodal large models. Tracking R&D funding in the Global South: challenges and opportunities in India https://www.leidenmadtrics.nl/articles/tracking-r-d-funding-in-the-global-south-challenges-and-opportunities-in-india 2026-11-02T00:00:00Z leidenmadtrics.nl Analyzes the challenges and opportunities of tracking R&D funding in India. What does “verified memory safe” even mean? https://focs-lab.comp.nus.edu.sg/blog/memory-safety-audit/ 2026-11-02T00:00:00Z focs-lab.comp.nus.edu.sg Independent audit of the memory-safety proofs by Anthropic and Theorem, assessing what verified memory safe means. Agentic Backdoors via Pretraining Poisoning https://pbb.sh/research/Agentic-Backdoor/ 2026-11-02T00:00:00Z pbb.sh Shows that pretraining data poisoning can plant a remote-code-execution backdoor in agents. BlitzRank: Principled Zero-shot Ranking Agents with Tournament Graphs https://blitzrank.ai/ 2026-11-02T00:00:00Z blitzrank.ai Project page for BlitzRank, a tournament-graph method for query-efficient zero-shot LLM document ranking. Making a Magic The Gathering Eval https://folds.ink/posts/mtg-benchmark 2026-11-02T00:00:00Z folds.ink Builds a harness and benchmark for LLMs playing Magic The Gathering and reports results. Lessons From Red-teaming Auto-review, OpenAI's Blocking Monitor https://www.lesswrong.com/posts/dbStc8ZFESqfihwAd/lessons-from-red-teaming-auto-review-openai-s-blocking-1 2026-11-02T00:00:00Z lesswrong.com Red-teaming report on OpenAI Auto-review blocking monitor with lessons learned. GPU-accelerated Weight Teleportation https://humansand.ai/blog/gpu-accelerated-weight-teleportation 2026-11-02T00:00:00Z humansand.ai Describes GPU compression with streamed decompression for lossless model weight updates in RL training. Shotwell-Bench: An annotation benchmark that accepts valid alternatives https://shotwell.ai/blog/shotwell-bench/ 2026-11-02T00:00:00Z shotwell.ai Introduces Shotwell-Bench, a dense video annotation benchmark that accepts valid alternative segmentations. From Words to Waypoints: Teaching a Small LLM to Plan for Robots https://www.daweisun.me/blog/stl-planning-rl/ 2026-11-02T00:00:00Z daweisun.me Experiments teaching a small LLM to produce waypoint plans for robots. Creating Preferred IUPAC names https://chem-bla-ics.linkedchemistry.info/2026/11/01/creating-preferred-iupac-names.html 2026-11-01T00:00:00Z chem-bla-ics.linkedchemistry.info Explores using language models to generate preferred IUPAC names. Morg2SMILES: Generative Recovery of SMILES Strings from Morgan Fingerprints https://jonswain.github.io/morg2smiles/ 2026-11-01T00:00:00Z jonswain.github.io Discusses generative recovery of SMILES strings from Morgan fingerprints. Can an LLM and a keyword index replace embeddings? https://internals.laxmena.com/p/can-an-llm-and-a-keyword-index-replace 2026-11-01T00:00:00Z internals.laxmena.com Experiment testing whether an LLM plus a keyword index can replace embeddings for semantic search, with measurements. The abandoned PR page https://aictrl.dev/blog/open-source-review-ai-era 2026-11-01T00:00:00Z aictrl.dev Data analysis of pull request outcomes in two AI-native open-source projects, focusing on abandoned PRs. Leaking Muse's Charter, Meta's model spec https://www.lesswrong.com/posts/TJGPL5FAk9sCioPbE/leaking-muse-s-charter-meta-s-model-spec 2026-11-01T00:00:00Z lesswrong.com Reports extracting the content of Meta Muse Spark Open model Charter from its chain of thought. AI text watermarking and quality loss https://blog.keyvan.net/p/ai-text-watermarking-and-quality 2026-10-31T00:00:00Z blog.keyvan.net Evaluates how AI text watermarking affects text quality. A Specialised Haskell Processor on FPGA https://bathtub-01.github.io/technical/ouros-demo/ 2026-10-31T00:00:00Z bathtub-01.github.io Presents a specialised processor that runs Haskell on an FPGA. A Generalized Scaling Law for Monopiles in Sand: From Centrifuge Tests to an Interactive Predictor https://geo-huanwang.com/posts/2026/10/monopile-moment-rotation-predictor/ 2026-10-31T00:00:00Z geo-huanwang.com Proposes a general scaling law for monopiles in sand based on centrifuge tests, with an interactive prediction tool. Automated Subtle Emergent Misalignment Research https://alignment.anthropic.com/2026/automated-em-search/ 2026-10-31T00:00:00Z alignment.anthropic.com Alignment research on automating the study of subtle emergent misalignment. Measuring voice agents: turn-taking, tool calls and expressive speech https://soniqo.audio/blog/voice-agent-benchmarks 2026-10-31T00:00:00Z soniqo.audio Reports measurements of voice agents using TurnBench, SLURP and InstructTTSEval. A CVE triage pipeline that keeps asking, because day one tells you nothing https://34-118-125-101.sslip.io/log/rappaport/ 2026-10-31T00:00:00Z 34-118-125-101.sslip.io Describes Rappaport, a CVE triage pipeline that keeps re-asking questions over time, with design details and lessons. Deterministic Factoring in Time $N^{o(1)}$ https://hexagonmath.org/2610.00600v1 2026-10-30T00:00:00Z hexagonmath.org Claims a deterministic integer factoring algorithm running in time N^o(1). Honeycomb-Surfels (Hexels) https://honeycomb-surfels.github.io/ 2026-10-30T00:00:00Z honeycomb-surfels.github.io Project page for HoneyComb Surfels, an edge 3D rendering method. What is work for? Human wellbeing and the value of work in the AGI era https://institute.deepmind.com/essays/what-is-work-for-human-wellbeing-and-the-value-of-work-in-the-agi-era/ 2026-10-30T00:00:00Z institute.deepmind.com DeepMind studies work and wellbeing in the era of AGI. CDC Grant Disruption Methodology https://grantwitness.org/cdc/updates/methodology 2026-10-30T00:00:00Z grantwitness.org Explains the sources and methods of the CDC grant disruption data. AGI Economics: Coordination vs. Competence https://institute.deepmind.com/essays/the-economics-of-intelligence/ 2026-10-30T00:00:00Z institute.deepmind.com A DeepMind institute analysis of AGI economics, covering jagged agents and mosaic organizations. Building oversight tools into a model's weights https://divagr.com/blog/building-oversight-tools-into-a-models-weights 2026-10-30T00:00:00Z divagr.com Describes building a lie detector and log into Qwen weights that the model cannot disable. An adventure in multifoundational Mathematics 2.0: the cycle double cover theorem in ZFC with Grace17 https://grace17.org/docs/cycle-double-cover.html 2026-10-30T00:00:00Z grace17.org Reports a formal proof of the cycle double cover theorem in ZFC using Grace17. Reverse Turing Test: Surprising Proof AIs Fail to Spot AIs https://www.progressiverobot.com/2026/10/27/reverse-turing-test-ais-fail-to-detect-ais/ 2026-10-30T00:00:00Z progressiverobot.com Reports a reverse Turing test where seven AI judges caught only 3 of 98 AI impostors. What agents see when they search MCP tools https://www.babush.me/how-agents-search-mcp-tools.html 2026-10-30T00:00:00Z babush.me Analyzes how agents search MCP tools and why the search misses relevant decompiler functions. Product Hunt is swamped now https://www.mahl.me/blog/product-hunt-is-swamped-now/ 2026-10-30T00:00:00Z mahl.me A small scrape quantifies how Product Hunt launch volume changed after 2023. Understanding Open-ended Work from a Data Perspective https://www.fig.inc/blog/understanding-open-ended-work-from-a-data-perspective/ 2026-10-30T00:00:00Z fig.inc Fig describes learning signals missing in data on open-ended computer-use work. Solving the JSON Schema dialect fragmentation problem with auto-upgrades https://www.sourcemeta.com/blog/json-schema-dialect-auto-upgrade/ 2026-10-30T00:00:00Z sourcemeta.com Describes an approach to auto-upgrade JSON Schema dialects to solve fragmentation. Grace Blackwell needs a new scheduler: 27% more usable GPUs from the same racks https://skypilot.ai/blog/grace-blackwell-needs-a-new-scheduler 2026-10-29T00:00:00Z skypilot.ai A Grace Blackwell scheduler yields 27% more usable GPUs. Open-weight AI models can exploit the patch gap https://epoch.ai/publications/open-weight-models-patch-gap 2026-10-29T00:00:00Z epoch.ai Evaluates how open-weight models can exploit the patch gap. N-gram hinted language modeling: NanoGPT Speedrun in 6.7s https://brunborg.com/n-gram-hinted-lm 2026-10-29T00:00:00Z brunborg.com Uses n-gram hints to bring the NanoGPT Speedrun down to 6.7 seconds. Weaponizing Personalization: From Knowledge Graph to Extortion https://8rank.ai/game/apple-intelligence-knowledge-graph-in-the-wrong-hands 2026-10-29T00:00:00Z 8rank.ai Research on using Apple's personal knowledge graph and an LLM to plan a kidnapping. Every iPhone has a fingerprint. Can one photo give you away? https://vhamolya.com/blog/which-iphone-took-this-photo/ 2026-10-29T00:00:00Z vhamolya.com Tests whether a single photo can identify an iPhone's camera fingerprint. Coding Agent Degradation: Cross-Vendor Field Notes and an Initial Analysis https://www.arcblock.io/en/articles/when-coding-agents-stop-finishing/ 2026-10-29T00:00:00Z arcblock.io A field analysis of performance degradation in coding agents across vendors. AReA https://rajabi2001.github.io/area/ 2026-10-29T00:00:00Z rajabi2001.github.io Project page for AReA, efficient attention for high-resolution diffusion transformers. Why AI Design Tools Keep Making the Same Flyer: Claude Design, Canva, Moda and Grafista Tested https://grafista.io/blog/posts/grafista-vs-claude-vs-canva-vs-moda 2026-10-29T00:00:00Z grafista.io Compares outputs of four AI design tools on identical flyer prompts, repeated three times. State of Harness—October 2026 https://www.harness-hub.org/reports/2026-10 2026-10-29T00:00:00Z harness-hub.org A report defining the agent harness and surveying the market with directory counts. Serving Prefill and Decode for Agentic AI https://tenstorrent.com/en/developers/blog/serving-prefill-and-decode-for-agentic-ai 2026-10-29T00:00:00Z tenstorrent.com Describes serving prefill and decode for agentic AI workloads on Tenstorrent Blackhole hardware. 31 AI Models Took a Political Quiz. Here Is Where Each One Landed https://politicaldna.org/ai-political-dna/ 2026-10-29T00:00:00Z politicaldna.org Gave a 32-statement political quiz to 31 LLMs over 1,017 runs and compared them with human respondents. Scaling Misuse Detection with Self-Assessment https://safeguards.anthropic.com/2026/self-assessment/ 2026-10-29T00:00:00Z safeguards.anthropic.com Presents and evaluates a method where Claude self-assesses response safety to scale misuse detection. Catching Bugs Before They Land: JiTTesting’s Human-on-the-Loop Approach at Full Scale https://engineering.fb.com/2026/10/29/security/catching-bugs-before-they-land-jittestings-human-on-the-loop-approach-at-full-scale/ 2026-10-29T00:00:00Z engineering.fb.com Meta engineering report on applying human-on-the-loop just-in-time testing with LLMs at full scale. Qwenputer: a fully functioning, deterministic RISC-V CPU inside Qwen3.5 https://divagr.com/blog/qwenputer 2026-10-28T00:00:00Z divagr.com Implements a deterministic RISC-V CPU inside Qwen3.5. The sleep framework: A new way for LLMs to retain knowledge in continual learning https://research.google/blog/the-sleep-framework-a-new-way-for-llms-to-retain-knowledge-in-continual-learning/ 2026-10-28T00:00:00Z research.google Google proposes a sleep framework for continual learning. From Tokens to Bytes I: Scaling Hierarchical Networks https://www.cartesia.ai/blog/from-tokens-to-bytes-part-1 2026-10-28T00:00:00Z cartesia.ai Cartesia's research on scaling hierarchical networks. Intelligence Benchmarking https://artificialanalysis.ai/methodology/intelligence-benchmarking 2026-10-28T00:00:00Z artificialanalysis.ai Artificial Analysis Intelligence Index v5.0 benchmark. How AI agents shape experts’ judgment calls https://www.anthropic.com/research/agents-judgment-calls 2026-10-28T00:00:00Z anthropic.com Studies how analysts follow AI agent preferences in 86% of decisions. Of sheep and beer https://dlab.epfl.ch/2017-08-30-of-sheep-and-beer/ 2026-10-28T00:00:00Z dlab.epfl.ch A study of how early ratings create herding effects in product reviews. Are you paying for useless LLM prompts? https://cameronmpalmer.com/blog/are-you-paying-for-useless-llm-prompts/ 2026-10-28T00:00:00Z cameronmpalmer.com An experiment assessing whether redundant LLM prompts are actually effective. Can LLMs Supercharge Geolocations? Our Tests Show Big Improvements But Humans Remain Key https://www.bellingcat.com/resources/2026/10/22/can-llms-supercharge-geolocations-our-tests-show-big-improvements-but-humans-remain-key/ 2026-10-28T00:00:00Z bellingcat.com Bellingcat evaluates how well LLMs perform at geolocation. Agent Scrum https://daveremy.github.io/agent-scrum/ 2026-10-28T00:00:00Z daveremy.github.io An agent sprint method for AI coding agents, with observations from 324 sprints. Recurrent Looped Transformer https://yifanzhang-pro.github.io/recurrent-looped-tranformer/ 2026-10-28T00:00:00Z yifanzhang-pro.github.io Project page on state tracking with the Recurrent Looped Transformer. MotionBricks: Scalable Real-Time Motions with Modular Latent Generative Model and Smart Primitives https://nvlabs.github.io/motionbricks/ 2026-10-28T00:00:00Z nvlabs.github.io Project page for the MotionBricks real-time motion generation paper. Who is in charge? Robot autonomy levels for human-in-the-loop systems https://zlatanajanovic.com/robot-autonomy-levels/ 2026-10-28T00:00:00Z zlatanajanovic.com Proposes a framework of robot autonomy levels for human-in-the-loop systems (book chapter). Building My Own Tiny AI Weather Model https://tcumulus.dev/blog/building-my-own-tiny-ai-weather-model/ 2026-10-28T00:00:00Z tcumulus.dev Describes building a small AI weather model based on the GraphCast architecture. Benchmarking Coding Agents at Scale https://www.snowflake.com/en/blog/engineering/Benchmarking-coding-agents/ 2026-10-28T00:00:00Z snowflake.com Introduces SnowPR, a benchmark built from real pull requests to evaluate coding agents on a large monorepo. Rethinking Search as Code Generation https://www.perplexity.ai/hub/blog/rethinking-search-as-code-generation 2026-10-28T00:00:00Z perplexity.ai Perplexity describes recasting search as programmable primitives and code generation for agents. Training Reve 2.0 at scale https://app.reve.com/training-infra 2026-10-28T00:00:00Z app.reve.com A technical report on the infrastructure used to train Reve 2.0 multimodal models. Octoverse: Commits, pushes, reach new highs in 2026 as developers build with agents https://github.blog/news-insights/octoverse/octoverse-commits-pushes-reach-new-highs-in-2026-as-developers-build-with-agents/ 2026-10-28T00:00:00Z github.blog GitHub annual report presenting 2026 statistics on commits, developers and growth across industries. CFSG in Lean https://cfsg.axiommath.ai/ 2026-10-27T00:00:00Z cfsg.axiommath.ai Formalizes the classification of finite simple groups in Lean and obtains a new theorem. Neuron Persona Library https://qwen-neurons-site.vercel.app/ 2026-10-27T00:00:00Z qwen-neurons-site.vercel.app Studies how flipping a single neuron triggers a persona in Qwen. A $2^{o(n)}$ Time Algorithm for 3-SAT https://hexagonmath.org/2610.00577v1 2026-10-27T00:00:00Z hexagonmath.org Claims a deterministic algorithm for 3-SAT that would refute the ETH. 40× fewer tokens: wrap Markdown in SQLite https://www.withnative.ai/research-and-posts/markdown-context-benchmark 2026-10-27T00:00:00Z withnative.ai Claims SQLite makes Markdown 40 times more token-efficient. A critical look at self-distillation for reasoning https://jasonkena.github.io/blog/posts/does_self_distillation_improve_reasoning/ 2026-10-27T00:00:00Z jasonkena.github.io A critical analysis of self-distillation for reasoning. SAS|Symbiotic Autoprotection System|κD = 0.56 https://leesintheblindmonk1999.github.io/sas-landing/#hero 2026-10-27T00:00:00Z leesintheblindmonk1999.github.io A claimed deterministic AI structural audit framework with K_D=0.56. teaching a car to drive inside its own dream https://fpoltronieri.github.io/blog/2026/teaching-a-car-to-drive-inside-its-own-dream/ 2026-10-27T00:00:00Z fpoltronieri.github.io Records experiments teaching autonomous driving to drive inside a latent world model. Same price, different reason. Probing AI pricing agents for collusion https://www.maxtaylor.me/articles/same-price-different-reason-probing-ai-pricing-agents-for-collusion 2026-10-27T00:00:00Z maxtaylor.me Tests whether linear probes can distinguish colluding AI pricing agents from innocent ones. Can Machines Follow One Voice Through a Crowded Room? https://www.kelvinjiang.com/2026/10/can-machines-follow-voice-through-crowded-room.html 2026-10-27T00:00:00Z kelvinjiang.com An exploratory experiment on whether machines can follow one voice in a crowded room. From the bottom of my heart, I salute you https://janni.turunen.in/poisoned-open-weights/ 2026-10-27T00:00:00Z janni.turunen.in Demonstrates three ways to poison open-weight models, including a few hundred poisoned samples. Human travel patterns reflect geographic boundaries rather than universal laws https://theprimary.com/science/2026-10-26/mobility-power-law-artifacts 2026-10-27T00:00:00Z theprimary.com Reports that human travel patterns follow geographic containers rather than universal power laws. The Fertility of an Idea: tracing the discovery lineage of optogenetics https://www.amplifypartners.com/blog-posts/the-fertility-of-an-idea 2026-10-27T00:00:00Z amplifypartners.com Traces the discovery lineage of optogenetics through citation and historical analysis. Can AI draw San Francisco from memory? https://voygr.tech/blog/can-ai-draw-san-francisco-from-memory/ 2026-10-27T00:00:00Z voygr.tech Tests six AI chat models at reproducing San Francisco map data from memory. The Ability of Frontier Models to Control Robots is Rapidly Improving https://www.lesswrong.com/posts/ndS9Xn8Qx5PppHJ3Q/the-ability-of-frontier-models-to-control-robots-is-rapidly 2026-10-27T00:00:00Z lesswrong.com Demonstrates that general-purpose multimodal models are rapidly improving at controlling robots. Concerning behavior from a popular open-weight AI model https://cohere.com/blog/concerning-behaviour-popular-open-weight-ai-model 2026-10-26T00:00:00Z cohere.com Cohere discloses abnormal behavior in an open-weight model. MS-Human-700: Whole-body Human Musculoskeletal Modeling and Control https://lnsgroup.cc/research/MS-Human/ 2026-10-26T00:00:00Z lnsgroup.cc Project page for MS-Human-700, a whole-body human musculoskeletal model in MuJoCo. Why the search for particle dark matter is hopeless, part 1 https://tritonstation.com/2026/10/26/why-the-search-for-particle-dark-matter-is-hopeless-part-1/ 2026-10-26T00:00:00Z tritonstation.com Argues why the search for particle dark matter is hopeless (part 1). Research | Galatea: optimal transport for lunar place recognition https://www.lunarsim.org/research/ 2026-10-26T00:00:00Z lunarsim.org Presents Galatea, an optimal transport method merging global and local descriptors for lunar visual place recognition. The Half-Life of an Instruction https://digitaliziran.si/2026/10/26/the-half-life-of-an-instruction.html 2026-10-26T00:00:00Z digitaliziran.si Reports an experiment on how DeepSeek V4 Pro's instruction following changes with few-shot examples. AI agents have few qualms with log tampering https://www.manifold.security/blog/agent-log-tampering 2026-10-26T00:00:00Z manifold.security Reports tests showing most frontier AI agents comply with requests to tamper with their own logs. Joints, Camera, Action! Co-Generating 3D Motion and Video for Arbitrary Rigs https://research.nvidia.com/labs/par/projects/jca/ 2026-10-26T00:00:00Z research.nvidia.com Presents Joints, Camera, Action, a model that jointly generates video and 3D motion for arbitrary rigs. Qwen, Minus the E https://kevin-wu.ca/blog/qwn/ 2026-10-25T00:00:00Z kevin-wu.ca An experiment fine-tuning Qwen on a 16GB GPU to avoid the letter e. Simulated Control Evaluations https://www.redwoodresearch.org/blog/simulated-control-evaluations 2026-10-25T00:00:00Z redwoodresearch.org Presents simulated control evaluations that red-team AI agent monitors using an LLM that fakes tool outputs. 3D Gaussian Splatting for Real-Time Radiance Field Rendering https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/ 2026-10-25T00:00:00Z repo-sam.inria.fr Project page for the SIGGRAPH paper on 3D Gaussian Splatting for real-time radiance field rendering. Exploring Apple's On-Device Model: Core Advanced https://www.airealist.ai/p/exploring-apple-on-device-model 2026-10-25T00:00:00Z airealist.ai An analysis of how Apple's 20B-parameter on-device model is stored in flash and activates few experts. Deleting RAG from our web agent made it 2.3x faster https://www.skyvern.com/blog/deleting-rag-from-our-web-agent-made-it-2-3x-faster/ 2026-10-24T00:00:00Z skyvern.com Reports that Skyvern's rewrite ran 2.3 times faster on the Odysseys benchmark. Timestral Financial Forecasting: Time-Series Foundation Models https://philippdubach.com/posts/forecasting-financial-markets-with-timestral/ 2026-10-24T00:00:00Z philippdubach.com Evaluation results of the Timestral model for forecasting financial markets. Compiling a Workflow into a 3B Model https://arpitbhayani.me/blogs/compiling-an-sop-into-a-3b-model/ 2026-10-24T00:00:00Z arpitbhayani.me A fine-tuning experiment compiling an incident triage SOP into a 3B model. BeNice! Closing the Inference-Memory Gap of Looped Models with KV-Sharing–Sandy Research https://sandyresearch.github.io/benice/ 2026-10-24T00:00:00Z sandyresearch.github.io Research on narrowing the inference-memory gap of looped models through KV-sharing. Solving a Giant Rubik's Cube with a Quadruped Robot https://leggedrobotics.github.io/rubik-skills/ 2026-10-24T00:00:00Z leggedrobotics.github.io Project page on a quadruped robot that autonomously scans and solves a giant Rubik's cube. AI Reviews Might Beat Human Peer Reviews https://blog.stephenturner.us/p/ai-peer-reviews-beat-humans 2026-10-24T00:00:00Z blog.stephenturner.us Introduces a demonstration that AI reviews may be no worse than human peer reviews. Mundo: From seeing a place to simulating what happens there https://huggingface.co/blog/Luo-Yihang/mundo-world-model 2026-10-24T00:00:00Z huggingface.co Presents Mundo, a world model that builds an interactive simulation of a room from a few photos. How well do job retraining programs work? https://www.anthropic.com/research/reviewing-the-evidence-on-worker-retraining-programs 2026-10-24T00:00:00Z anthropic.com An evidence review of worker retraining programs from Anthropic's Economic Research team. Crowley's Three Axes of JSON Querying https://nrposner.com/blog/crowleys-three-axes-of-json-querying/ 2026-10-24T00:00:00Z nrposner.com Describes design axes and benchmarks for a streaming JSON query engine exceeding 2 GB/s. ReplaySSM: Cache SSM Inputs, Not State https://tridao.me/blog/2026/replayssm/ 2026-10-24T00:00:00Z tridao.me Proposes ReplaySSM, which caches SSM inputs instead of recurrent state to reduce decoding memory traffic. Air Purifiers as White Noise Machines https://www.jefftk.com/p/air-purifiers-as-white-noise-machines 2026-10-23T00:00:00Z jefftk.com Measures how well air purifiers work as white noise machines. An alignment assessment of the UK AISI cybertesting incident https://www.anthropic.com/research/alignment-assessment-ukaisi 2026-10-23T00:00:00Z anthropic.com Assesses Mythos 5's unauthorized actions during the UK AISI evaluation. A Fix for Softmax and Cross-Entropy https://jonbarron.info/softmaxfix/ 2026-10-23T00:00:00Z jonbarron.info Proposes an improvement to softmax and cross-entropy. A good teacher meets the student where they are https://www.perplexity.ai/hub/blog/a-good-teacher-meets-the-student-where-they-are 2026-10-23T00:00:00Z perplexity.ai Proposes JOLT, which jointly trains a teacher and a student. The Latest Anthropic Claude Models Contain AI Text Watermarks: What Does This Mean for You? https://www.oreilly.com/radar/the-latest-anthropic-claude-models-contain-ai-text-watermarks-what-does-this-mean-for-you/ 2026-10-23T00:00:00Z oreilly.com Reports that Claude model outputs contain AI text watermarks. MemCurrent: Designing Persistent Memory for Production Agents https://www.databricks.com/blog/memcurrent-designing-persistent-memory-production-agents 2026-10-23T00:00:00Z databricks.com Introduces the MemCurrent persistent memory architecture and its effects. Hamiltonian Paths in Solid Grid Graphs https://zzyzek.github.io/SolidGridGraphHamPathPaper/ 2026-10-23T00:00:00Z zzyzek.github.io Proves that Hamiltonian paths in solid grid graphs can be decided in polynomial time. Two Path Zig Zag Numberlink https://zzyzek.github.io/R2K2ZZNPaper/ 2026-10-23T00:00:00Z zzyzek.github.io Gives a forbidden-pattern characterization of Numberlink and a polynomial-time algorithm. On Serial Depth and Monitorability in Looped Transformers, Part 1 https://aopatric.github.io/blog/2026/recurrent-grokking/ 2026-10-23T00:00:00Z aopatric.github.io Studies serial depth in looped transformers and the monitorability of chain of thought. EPA Grant Disruption Methodology https://grantwitness.org/epa/updates/methodology 2026-10-23T00:00:00Z grantwitness.org Explains the sources and methods of the EPA grant disruption data. Twitch Plays: AI Interpretability of Bonsai-8B https://www.lesswrong.com/posts/ozbbzsh49wvgfbCDp/twitch-plays-ai-interpretability-of-bonsai-8b 2026-10-23T00:00:00Z lesswrong.com Live interpretability exploration of Bonsai-8B with a public interpretability map. MalwareBench: Reverse-engineering custom-built, realistic, evasive malware into actionable intelligence https://malwarebench.com 2026-10-23T00:00:00Z malwarebench.com Evaluates frontier models on reverse engineering evasive malware across platforms with a new benchmark. FrontierCS 2 https://frontier-cs.org/?v=2 2026-10-23T00:00:00Z frontier-cs.org A benchmark of 48 open-ended computer science R&D tasks for evaluating AI agents. Finite Ring Cosmology Lab: Schrödinger—Dirac dynamics of the total state over the (13, 233) holographic substrate https://finitering.space/38-s13/lab.html 2026-10-23T00:00:00Z finitering.space Presents a browser simulation of Schrödinger-Dirac dynamics on a holographic substrate with a claimed universe model. A Robust Decentralized Infrastructure for Trust-Aware Open Knowledge Sharing https://knowledgepixels.com/nanopub-ecosystem-paper/ 2026-10-23T00:00:00Z knowledgepixels.com Describes a decentralized, trust-aware infrastructure for open knowledge sharing. AI Article Reviews Dominate Human Peer Reviews at Their Best: Evidence and Policy Choices https://gking.harvard.edu/peerless/ 2026-10-23T00:00:00Z gking.harvard.edu Presents evidence that AI-generated article reviews outperform human peer reviews on many measures, with policy implications. An LLM judge passed renders with defects I planted https://yamanahlawat.in/articles/llm-judge-same-code-different-scores 2026-10-23T00:00:00Z yamanahlawat.in Experiments show a vision-model judge passes renders with planted defects and is inconsistent across runs. Jakub Ćwirlej https://jakubcwirlej.com/ai-dev/ 2026-10-23T00:00:00Z jakubcwirlej.com Documents training a 160M code model on one GPU that reaches 22.56% on HumanEval. Bootstrapping the Verified Software Stack https://theorem.dev/blog/verified-software-stack/ 2026-10-23T00:00:00Z theorem.dev Describes a research effort to build a shared formally verified software stack starting with nixpkgs. How does Fable 5.1 hide messages in images? https://www.lesswrong.com/posts/WHEQps5PK8LrrcTAb/how-does-fable-5-1-hide-messages-in-images 2026-10-22T00:00:00Z lesswrong.com Studies how frontier LLMs hide information inside images. Rule 30: One Black Cell https://www.wolframcollatzmahler.com/rule30/index.html 2026-10-22T00:00:00Z wolframcollatzmahler.com Explores the properties of Rule 30 growing from a single black cell. Opening Linear A https://www.plasma.ai/research/linear-a 2026-10-22T00:00:00Z plasma.ai Open-sources a Semitic decipherment of Linear A with corpus evidence. Does a text-to-speech model know what it’s saying? https://vernolabs.ai/research/does-a-text-to-speech-model-know-what-its-saying 2026-10-22T00:00:00Z vernolabs.ai Tests whether a text-to-speech model understands what it says. Good Night, and Good Luck: AI, fear and what the job data says https://rpallas.xyz/good-night-and-good-luck/ 2026-10-22T00:00:00Z rpallas.xyz Uses data to analyze how AI affects software jobs. LoongForge-VLDA https://baidu-baige.github.io/LoongForge/projects/vlda.html 2026-10-22T00:00:00Z baidu-baige.github.io Proposes VLDA, which uses compact deltas as the interface for robot manipulation. Automated nightly LLM testing: an AI agent benchmarks new local models while I sleep https://siimtolk.github.io/articles/llm-testing-overnight/ 2026-10-22T00:00:00Z siimtolk.github.io Tests 18 local LLMs on two RTX 5060 Ti GPUs. Machines Read, Humans Sign: Code Review in One Year https://www.hugo.im/posts/code-review-in-one-year/ 2026-10-22T00:00:00Z hugo.im Reports one year of data from an in-house code review agent. Elasticity of money in production networks, working capital, credit lines and financial conditions https://www.bis.org/publications/working-paper-1350-elasticity-money-production-networks-working-capital-credit-lines-and-financial-conditions 2026-10-22T00:00:00Z bis.org Research paper on money elasticity, credit lines and production networks. Solving a 31-year-old grand challenge in mathematics–the KLS conjecture https://www.microsoft.com/en-us/research/blog/solving-a-31-year-old-grand-challenge-in-mathematics-the-kls-conjecture/ 2026-10-22T00:00:00Z microsoft.com Announces resolution of the KLS conjecture with AI assistance. China’s Next-Generation Industrial Policy https://rhg.com/research/chinas-next-generation-industrial-policy/ 2026-10-22T00:00:00Z rhg.com Report analyzing China's next-generation industrial policy. Mammoth-1 and Elephant-1: frontier decision models https://labs.baseten.co/articles/mammoth-1-and-elephant-1-frontier-decision-models 2026-10-22T00:00:00Z labs.baseten.co Technical report on the Mammoth-1 and Elephant-1 decision models. Chinese AI Writes Weaker Code for Beijing’s Adversaries. We Fixed It. A Qwen3-Coder Case Study https://www.hirundo.io/blog/qwen3-coder-vulnerabilities 2026-10-21T00:00:00Z hirundo.io Finds that Qwen3-Coder generates less secure code when prompted with US-related personas. Reactive Optimizers https://xingjianbai.com/reactive-optimizers/ 2026-10-21T00:00:00Z xingjianbai.com Proposes Reactive Optimizers, which let an agent rewrite the optimizer source code during training. GLM-5.3 Sometimes Considers How Train-Time Actions Shape Future Behavior https://www.lesswrong.com/posts/R5g8YT25G9DXJJ48n/glm-5-3-sometimes-considers-how-train-time-actions-shape 2026-10-21T00:00:00Z lesswrong.com Finds that GLM-5.3 sometimes considers how its training-time actions shape future behavior. What Claude Code's multi-agent code review costs and finds|Quickchat AI https://quickchat.ai/post/claude-code-multi-agent-code-review 2026-10-21T00:00:00Z quickchat.ai Evaluates the cost of Claude Code's multi-agent code review and the issues it finds. Is 'migration choice' possible during the migration course? Understanding biographical migration projects https://fluchtforschung.net/is-migration-choice-possible-during-the-migration-course-understanding-biographical-migration-projects/ 2026-10-21T00:00:00Z fluchtforschung.net Distinguishes biographical migration projects from migration intentions and analyzes them. AI Agents in a Shared Workspace: Benchmark https://www.spacebar.ai/benchmark 2026-10-21T00:00:00Z spacebar.ai Introduces SpacebarBench for evaluating agents in a shared workspace. Pantheon | Building the Most Diverse UMI Dataset in Robotics https://pantheon.inc/research/diverse-umi 2026-10-21T00:00:00Z pantheon.inc Describes a diverse UMI robotics dataset collection operation. RFD4-Proteina https://rfd4-proteina.org/ 2026-10-21T00:00:00Z rfd4-proteina.org Presents RFD4-Proteina, a latent reasoning model for biomolecular design. Deterministic Inference Did Not Make Our Pipeline Deterministic https://www.ctgt.ai/research/concordance 2026-10-21T00:00:00Z ctgt.ai Experiment showing deterministic LLM inference did not make a pipeline deterministic. The Overfit https://blog.gopfert.eu/ai-macros-from-photos 2026-10-21T00:00:00Z blog.gopfert.eu Weighs 32 meals to test how accurately AI estimates macros from photos. Benchmarking Bot Detection Systems Against Modern AI Agents in 2026 https://research.poh.org/bot-benchmarking-2026/ 2026-10-20T00:00:00Z research.poh.org Re-benchmarks how well bot detection systems handle modern AI agents. How We Cold Start a 32B Model in 1.1s https://numerata.com/blog/post-gpu-cold-start-checkpoint-restore 2026-10-20T00:00:00Z numerata.com Describes cold starting a 32B model in 1.1 seconds using GPU checkpoint/restore. Improving in chess is hard. I built the world's most human-like chess AI to help me. https://mbuffett.com/posts/chess-move-prediction 2026-10-20T00:00:00Z mbuffett.com Introduces the author's chess AI, built to be the most human-like. Celestial Vision Navigation, or: How I Found Where an Elite Dangerous Screenshot from 2015 Was Taken, Using Computer Vision https://alphapixeldev.com/celestial-vision-navigation-elite-dangerous-dark-wheel-station/ 2026-10-20T00:00:00Z alphapixeldev.com Uses star-based positioning to solve a mystery location in Elite: Dangerous. Reducing Reward Hacking by Localizing Bad Behavior Instead of Penalizing It https://alignment.anthropic.com/2026/graft/ 2026-10-20T00:00:00Z alignment.anthropic.com Proposes reducing reward hacking by localizing bad behavior rather than penalizing it. Blanking Beavers: Results and Open Problems https://nickdrozd.github.io/2026/10/20/blanking-beavers-results-and-open-problems.html 2026-10-20T00:00:00Z nickdrozd.github.io Reports results and open problems for the Blanking Beavers Turing machine problem. How leaky is your inference API? Detecting model weight steganographic exfiltration https://weights-exfil.github.io/model-weights-exfiltration/ 2026-10-20T00:00:00Z weights-exfil.github.io Shows a covert channel that steals model weights through token choices in LLM APIs. Robotic arm reaching task with reinforcement learning (PPO) https://ethan-jz.github.io//ppo-robot-reaching/ 2026-10-20T00:00:00Z ethan-jz.github.io Deploys a robotic arm reaching task using PPO reinforcement learning. MARBLE: exact sequence parallelism for packed multimodal masks https://lumalabs.ai/news/packed-multimodal-attention 2026-10-20T00:00:00Z lumalabs.ai Presents Marble, exact sequence parallelism for packed multimodal sequences, with results. Building a System One Model for Prompt Injection https://www.stackone.com/blog/system-one-model-prompt-injection/ 2026-10-20T00:00:00Z stackone.com Describes rebuilding the Defender prompt injection guard with evaluation results. Fine-tuning a small language model for RAG in a browser on a Mac https://bart.degoe.de/fine-tuning-a-small-language-model-for-rag/ 2026-10-20T00:00:00Z bart.degoe.de Fine-tunes Qwen3-0.6B on a Mac for in-browser RAG and reports results. Jev models make agents faster. Hooked control planes steal the difference. https://pentad.ai/PLRN/031/ 2026-10-20T00:00:00Z pentad.ai Analyzes how hook chains absorb speedups from fast decision models. The free upgrade that raised your bill: what a new model version really costs https://three.dev/blog/the-free-upgrade-that-raised-your-bill 2026-10-20T00:00:00Z three.dev Measures how new model versions change real API bills at equal list price. Beyond g-mean²: complementary metrics for monitorability https://alignment.openai.com/beyond-gmean-squared/ 2026-10-20T00:00:00Z alignment.openai.com Introduces two complementary monitorability metrics beyond g-mean². RELAGRAM https://nevolodia.com/thoughts/001/ 2026-10-20T00:00:00Z nevolodia.com Describes Relagram, latent recursion RL for vehicle routing. Hands Off: Detecting Model Weight Tampering https://stealthium.io/blog/hands-off-detecting-model-weight-tampering 2026-10-19T00:00:00Z stealthium.io Introduces the ROME attack that tampers with model weights in GPU memory, and ways to detect it. Towards Company World Models https://www.beaconholdings.com/research/towards-company-world-models 2026-10-19T00:00:00Z beaconholdings.com Proposes company world models, where AI learns from shop owners' experience to support small business decisions. Can AI automate Epoch? https://epoch.ai/publications/can-ai-automate-epoch 2026-10-19T00:00:00Z epoch.ai Epoch has six models do real work tasks to assess how far its work can be automated. Are You Sure You Want to Use MMAP in Your Database Management System? (CIDR 2022) https://db.cs.cmu.edu/mmap-cidr2022/ 2026-10-19T00:00:00Z db.cs.cmu.edu A paper page arguing that database management systems should not use mmap. How to Measure Intelligence Beyond Human Scale? https://www.lesswrong.com/posts/avtquSx2PkWsxXW6T/how-to-measure-intelligence-beyond-human-scale 2026-10-19T00:00:00Z lesswrong.com Introduces, based on a paper, methods for measuring intelligence beyond human scale. Is Python the favorite language of LLMs? https://www.zansara.dev/posts/2026-10-19-is-python-the-favorite-language-of-llms/ 2026-10-19T00:00:00Z zansara.dev Examines whether LLMs prefer programming languages such as Python. What content categories are used in science blogs? https://blog.front-matter.de/posts/what-content-categories-are-used-in-science-blogs/ 2026-10-19T00:00:00Z blog.front-matter.de Analyzes the content categories of Rogue Scholar science blogs and proposes a taxonomy. How much intelligence can you run locally? (October 2026) https://localai-brief.vercel.app/ 2026-10-19T00:00:00Z localai-brief.vercel.app Research brief comparing locally runnable models with frontier models on benchmark indices. The Art of Swarming https://www.csis.org/analysis/art-swarming 2026-10-19T00:00:00Z csis.org Policy report arguing drone swarms change mass on the battlefield. Efficient Extraction for Effectful E-graphs https://ztatlock.net/pubs/2026-oopsla-eggcc/ 2026-10-19T00:00:00Z ztatlock.net Presents Statewalk DP for fast effect-aware e-graph extraction without ILP. Evaluating Javo https://getjavo.com/research/sales-evaluation/ 2026-10-19T00:00:00Z getjavo.com Evaluates an AI sales agent on 100 websites and 14.8 million visitors. Lessons learned building Fronteira by Alcides Fonseca https://wiki.alcidesfonseca.com/blog/lessons-learned-building-fronteira/ 2026-10-19T00:00:00Z wiki.alcidesfonseca.com Describes Fronteira, an MCP proxy using information-flow policies to block attacks. Running 10 million Codex CLIs on one server-LFI processes https://kapteko.com/running-10-million-codex-clis-on-one-server-lfi-processes/ 2026-10-19T00:00:00Z kapteko.com Describes running millions of Codex CLI processes on one server. MemStrata: AI answers you can check, with sources. https://www.memstrata.dev/ 2026-10-19T00:00:00Z memstrata.dev Shows LongMemEval scores shift depending on whether the judge can see the sources. How RC2’s π table was made (PITABLE, RFC 2268) https://khalidhamadeh.com/writing/rc2-pi-table 2026-10-18T00:00:00Z khalidhamadeh.com Studies how the RC2 cipher's pi table was generated and reproduces it with a Fisher-Yates shuffle. Neural networks are not already clean https://www.lesswrong.com/posts/Dkh7pXKjoNWkkxQDz/neural-networks-are-not-already-clean 2026-10-18T00:00:00Z lesswrong.com Discusses what evidence is needed to show a neuron is explained, and rebuts the claim that neural networks are already clean. TIC 402710280: A TESS Exoplanet Candidate https://sixeleven.in/blog/tic402710280 2026-10-18T00:00:00Z sixeleven.in Investigates a possible exoplanet using TESS transit data and Claude. NEEDLE — search engine benchmarks https://keenableai.github.io/needle/ 2026-10-18T00:00:00Z keenableai.github.io Releases Needle, an open-source real-time search engine benchmark for agents, with results. Una Watch TOTP - MFA on your wrist! https://shkspr.mobi/blog/2026/10/una-watch-totp-mfa-on-your-wrist/ 2026-10-18T00:00:00Z shkspr.mobi Builds a TOTP two-factor authentication device for the wrist. Training and Deploying Production Cyber Monitors on Kimi K3 https://www.goodfire.com/research/production-cyber-monitors-on-kimi-k3 2026-10-18T00:00:00Z goodfire.com Trains cybersecurity monitors for Kimi K3 that beat an LLM judge at lower cost. Can AI set the focal point of a photo? https://workflows-of-ai.com/blog/can-ai-set-the-focal-point-of-a-photo 2026-10-18T00:00:00Z workflows-of-ai.com Evaluates 19 methods for cropping images using LLMs and classification models. A faster way to convert a timestamp ➜ Hour, Min, Sec https://www.benjoffe.com/fast-time-of-day 2026-10-18T00:00:00Z benjoffe.com Presents a faster timestamp to hour-minute-second algorithm with benchmark results. Ghidra + CBMC = Assembly Model Checker https://www.philipzucker.com/asmc_cpp/ 2026-10-18T00:00:00Z philipzucker.com Connects Ghidra lifted assembly semantics to CBMC to build an assembly model checker. Beating Warcraft III’s insane AI with self-play reinforcement learning on one desktop https://warlock.ai/blog/warcraftsim-1/ 2026-10-18T00:00:00Z warlock.ai Trains a self-play RL agent reaching 89% win rate against Warcraft III's built-in AI. How employment is changing at Anthropic https://www.anthropic.com/research/employment-at-anthropic 2026-10-17T00:00:00Z anthropic.com Anthropic analyzes how AI is changing employment inside the company. Basis | Ship your interpreter https://www.basis.ai/blog/ship-your-interpreters/ 2026-10-17T00:00:00Z basis.ai Shows how to use an LLM to write rules for a programming language and prove that an interpreter conforms to them. Does code quality still matter when AI writes the code? https://docspring.com/blog/posts/does-code-quality-still-matter-when-ai-writes-the-code/ 2026-10-17T00:00:00Z docspring.com Replays mistakes in agent-assisted development and asks whether code quality still matters in the AI era. Proactive AI for Superintelligent Agents https://research.google/pubs/proactive-ai-for-superintelligent-agents/ 2026-10-17T00:00:00Z research.google Google Research proposes a proactive AI approach for superintelligent agents. Why gradient clipping changes Muon’s geometry: a two-dimensional example https://bishnubhusal.com.np/blog/2026/clipping-geometry-private-muon/ 2026-10-17T00:00:00Z bishnubhusal.com.np Uses a two-dimensional example to show how gradient clipping changes Muon's geometry. When Security Rules Keep Changing Their Minds https://elena6918.github.io/blog/2026/rules/ 2026-10-17T00:00:00Z elena6918.github.io Studies version flip-flopping in LLM-written security detection rules across iterations. Peer reviewed articles about biological processes in WikiPathways https://chem-bla-ics.linkedchemistry.info/2026/10/17/peer-reviewed-articles-about-biological-processes.html 2026-10-17T00:00:00Z chem-bla-ics.linkedchemistry.info Surveys peer-reviewed articles about biological processes in WikiPathways. Open and Emergent Problems in Agentic Privacy and Security: A Contextual Angle https://research.google/pubs/open-and-emergent-problems-in-agentic-privacy-and-security-a-contextual-angle/ 2026-10-17T00:00:00Z research.google Technical report outlining open problems in agentic privacy and security. Building a Chess Cheat Detector https://chesscheatdetector.com/blog/building-a-chess-cheat-detector 2026-10-17T00:00:00Z chesscheatdetector.com Builds a chess cheat detector from 6,613 simulated games with Maia. We shrank Whisper large-v3 by 68%, below the size of turbo https://tileward.com/research/a-smaller-speech-recogniser/ 2026-10-17T00:00:00Z tileward.com Reports compressing Whisper large-v3 by 68% with accuracy results on ten tests. A benchmark map of the US economy https://www.mercor.com/research/labor-index/benchmark-map/ 2026-10-16T00:00:00Z mercor.com Builds an AI benchmark map and analyzes how well benchmarks cover US economic work hours. Designing more efficient quantum circuits with Claude https://www.anthropic.com/research/designing-efficient-quantum-circuits 2026-10-16T00:00:00Z anthropic.com Claude autonomously designs quantum circuits and improves on the best published versions. The AI expectations gap https://www.anthropic.com/research/ai-expectations-gap 2026-10-16T00:00:00Z anthropic.com Anthropic analyzes where divergent expectations about AI's economic impact come from. How pro-China are Chinese models? https://www.vals.ai/blogs/china-us-analysis 2026-10-16T00:00:00Z vals.ai Evaluates how pro-China Chinese models are on political topics. Scaling pre-training on Melee https://ericyuegu.com/hillclimbing-melee/ 2026-10-16T00:00:00Z ericyuegu.com Scales pretraining of Melee game policies through action representations and offline RL. Do looped transformers write reasoning they don't use? https://igort99.github.io/posts/2026/10/05/looped-cot-faithfulness/ 2026-10-16T00:00:00Z igort99.github.io Experiments test whether the looped transformer Huginn actually uses its reasoning. NSF Grant Disruption Methodology https://grantwitness.org/nsf/updates/methodology 2026-10-16T00:00:00Z grantwitness.org Explains how NSF grant termination data is tracked. Models know when they’re reward hacking—and we can catch them at scale https://www.goodfire.com/research/reward-hacking-activation-monitors 2026-10-16T00:00:00Z goodfire.com Finds an internal signal accompanying reward hacking and builds probes to detect it. This is not a game of snake https://blog.neels.tech/blog/this-is-not-a-game-of-snake/ 2026-10-16T00:00:00Z blog.neels.tech Experiment teaching a transformer model to play Snake and examining what it learns. Toward RL on mixed-precision, mixed-accelerator samplers https://modal.com/blog/toward-rl-on-mixed-precision-mixed-accelerator-samplers 2026-10-16T00:00:00Z modal.com Reports asynchronous RL training for agentic coding using mixed-precision, mixed-accelerator samplers. Bonsai Blog | Fully Managed Elasticsearch & OpenSearch • The Pink Elephant in Your Prompt https://bonsai.io/blog/pink-elephant-in-your-prompt/ 2026-10-16T00:00:00Z bonsai.io Reports 720 trials testing whether telling an agent to ignore existing code works. Running DOOM Inside a Transformer https://nostalgia.dev/posts/running-doom-inside-a-transformer/ 2026-10-16T00:00:00Z nostalgia.dev Builds a transformer that executes the DOOM game and describes the method. A Statistical Analysis of LLM Benchmaxxing https://llemre.com/benchmaxxing/ 2026-10-16T00:00:00Z llemre.com Statistical estimate of benchmark overfitting by model lab. REKI: Registering Knowledge in VLAs https://kyutai.org/reki/ 2026-10-16T00:00:00Z kyutai.org Presents REKI, a 2.4B vision-language-action model with memory, trained on public data. Benchmarking Open-Source LLMs on Verilog RTL Generation Across 50 Tasks (NMIMS, IIT Roorkee, BITS Pilani) https://semiengineering.com/benchmarking-open-source-llms-on-verilog-rtl-generation-across-50-tasks-nmims-iit-roorkee-bits-pilani/ 2026-10-16T00:00:00Z semiengineering.com News summary of a paper benchmarking open-source LLMs on Verilog RTL generation. Neural Networks Are Already Clean https://already-clean.goodfire.com/ 2026-10-15T00:00:00Z already-clean.goodfire.com Proposes a working hypothesis that neurons and circuits in neural networks are relatively clean. Jev Meets the Trolley Problem https://zachlandes.com/trolley-problem/ 2026-10-15T00:00:00Z zachlandes.com Poses the trolley problem to Jev and 13 decision models and compares the results. Exposing AdFection: How Discover-Tech Turned Targeted Ads Into a Pipeline for Zero-Click Mercenary Spyware https://citizenlab.ca/research/how-discover-tech-turned-targeted-ads-into-a-pipeline-for-spyware/ 2026-10-15T00:00:00Z citizenlab.ca Citizen Lab discloses zero-click spyware delivered through targeted ads. Courk's Blog – Laser Fault Injection on a Budget: RP2350 Edition https://courk.cc/rp2350-challenge-laser 2026-10-15T00:00:00Z courk.cc Carries out low-cost laser fault injection on the RP2350. Reproducing the Hugging Face Incident, Part I https://www.abundant.ai/blog/hugging-face-incident-part-1/ 2026-10-15T00:00:00Z abundant.ai A controlled reproduction of the package manager attack in the Hugging Face incident. How We Trained an AI Model to Play Browser Games Live https://ironbee.ai/blog/how-we-trained-an-ai-model-to-play-browser-games 2026-10-15T00:00:00Z ironbee.ai Describes how a local model was trained to play browser games live, with results. What are the most important commands of a programming language? https://blog.ptidej.net/what-are-the-most-important-commands-of-a-programming-language/ 2026-10-15T00:00:00Z blog.ptidej.net Analyzes the most important commands in AMOS BASIC programs to enable translation to JavaScript. Package managers have an integrity problem https://blog.trailofbits.com/2026/10/14/package-managers-have-an-integrity-problem/ 2026-10-15T00:00:00Z blog.trailofbits.com Tests how well 25 package managers detect tampered repositories. Apertus 1.5 https://apertus-ai.org/articles/2026-07-apertus-1-5/ 2026-10-15T00:00:00Z apertus-ai.org Technical report for Apertus 1.5 covering multimodality, reasoning, and long context. FLAT: Flash Linear Attention on TPUs https://trainstation-project.github.io/blog/flat/ 2026-10-15T00:00:00Z trainstation-project.github.io Implements kernels such as linear attention on TPUs using Pallas. Barista: Synthesizing Typestate Specifications with LLM Agents https://catarinagamboa.github.io/papers/barista/ 2026-10-15T00:00:00Z catarinagamboa.github.io Barista: synthesizes typestate specifications with LLM agents (ASE 2026). ML Conference Trends: NeurIPS & ICLR, 2023–2026 https://wenyaxie023.github.io/blog/2026/ml-conference-trends/ 2026-10-15T00:00:00Z wenyaxie023.github.io Counts topic trends in accepted NeurIPS and ICLR papers from 2023 to 2026. A Computational Swinging Lexical Network https://schiekiera.github.io/blog/2026/computational-swinging-lexical-network/ 2026-10-15T00:00:00Z schiekiera.github.io Implements a computational model of the Swinging Lexical Network for word production. T12 POSS Cages: D6h and D2d Dodecasilsesquioxane Isomers https://silsesquioxane.blogspot.com/2026/10/t12-dodecameric-silsesquioxanes.html 2026-10-15T00:00:00Z silsesquioxane.blogspot.com Discusses the D6h and D2d isomers of the T12 POSS cage. Gemma 3 1B at 114 tokens per second on a laptop CPU https://knl.co.in/posts/gemma-3-1b-ternary/ 2026-10-15T00:00:00Z knl.co.in Reports ternary conversion and custom kernels getting Gemma 3 1B to 114 tokens per second on a laptop CPU. Confidential AI ･ Encrypted KV Cache Offloading for Confidential Computing https://confidential.ai/blog/kv-offload-under-cc 2026-10-15T00:00:00Z confidential.ai Presents encrypted KV cache offloading for confidential computing as a production implementation. ExplainThatMove https://www.explainthatmove.com/ 2026-10-15T00:00:00Z explainthatmove.com Compares chess moves and explains in natural language why the stronger one is better. Context Inertia: When Agents Won’t Let Go https://www.dbreunig.com/2026/10/14/context-inertia.html 2026-10-15T00:00:00Z dbreunig.com Describes context inertia, where agents keep pursuing goals despite contrary evidence. Inside Agentic AI Workloads https://gimletlabs.ai/blog/inside-agentic-ai-workloads 2026-10-15T00:00:00Z gimletlabs.ai Characterizes agentic AI workloads that repeatedly re-enter model inference over growing state. Routers Are a New AI Primitive https://www.vals.ai/blogs/routers-new-ai-primitive 2026-10-15T00:00:00Z vals.ai Argues for and runs independent evaluation of model routers on quality, cost and latency. The Continual Constitutional Convention: How language models revise the documents that govern them https://www.lesswrong.com/posts/c9woewdY9ddsPqf3J/the-continual-constitutional-convention-how-language-models 2026-10-15T00:00:00Z lesswrong.com Experiments studying how language models revise the constitutions that govern them. What 1,866 GitHub Projects Tell Their AI Coding Agents https://secondtimeisabug.com/blog/what-projects-tell-ai-coding-agents/ 2026-10-15T00:00:00Z secondtimeisabug.com Analyzes AGENTS.md files from 1,866 popular GitHub projects. How We Made Our MMDB Files a Third Smaller Without Changing a Byte of Data https://ipinfo.io/blog/how-made-mmdb-files-smaller 2026-10-14T00:00:00Z ipinfo.io IPinfo describes an optimization that shrinks MMDB files while keeping lookup results unchanged. We found a Tencent agent fleet https://swarmcha.se/posts/chinese-agent-fleet-full-report 2026-10-14T00:00:00Z swarmcha.se An investigation documenting large-scale AI agent activity on cloud platforms. The Machines that Make the Machines https://developer.nvidia.com/blog/the-machines-that-make-the-machines/ 2026-10-14T00:00:00Z developer.nvidia.com NVIDIA describes methods and lessons from training robots to assemble GB300 test trays. AI Scaling vs Human Scaling https://www.tobyord.com/writing/ai-scaling-vs-human-scaling 2026-10-14T00:00:00Z tobyord.com Analyzes whether AI model scaling with tokens resembles human scaling over time. Build-A-Chip-Bench https://chipagents.ai/research/build-a-chip-bench 2026-10-14T00:00:00Z chipagents.ai Proposes Build-A-Chip-Bench to evaluate agents on chip design from specification to GDS. We found 110 reward hacks and grading bugs in Terminal-Bench, DeepSWE, and BFCL https://usetokenless.com/blog/envcheck/ 2026-10-14T00:00:00Z usetokenless.com Finds 110 reward hacks and grading bugs in benchmarks including Terminal-Bench. Analyzing technical documents with Deep Review https://research.google/blog/analyzing-technical-documents-with-deep-review/ 2026-10-14T00:00:00Z research.google Google Research describes analyzing technical documents with Deep Review. A Turing Test for Physical Engineering https://www.spectrallabs.ai/research/cswp-turing-test 2026-10-14T00:00:00Z spectrallabs.ai Evaluates whether AI agents can pass SOLIDWORKS certification exams. Ultramarine 1 Flash: Scaling Non-Verifiable RL https://www.paperinstruments.com/blog/ultramarine 2026-10-14T00:00:00Z paperinstruments.com Introduces Ultramarine 1 Flash, an RL model for non-verifiable tasks. Visualizing Crossref Repository Histories https://metadatagamechangers.com/blog/2026/10/10/visualizing-crossref-repository-histories 2026-10-14T00:00:00Z metadatagamechangers.com Visualizes the history of Crossref repository metadata. Optimization: does quantization hurt coding agents? https://aistack.imec-int.com/blog/quantization 2026-10-14T00:00:00Z aistack.imec-int.com Measures how FP8, INT4 AWQ and NVFP4 quantization affect coding agent performance. rust-djinn: You Should Specify Your Code Unambiguously (and Prove That You Did) https://henniger.dev/post/djinn/ 2026-10-14T00:00:00Z henniger.dev Presents a 300-line specified, proven gzip decompressor in Rust. Next-Concept Modeling: Fast, Parallel Coarse-to-Fine Text Generation | Scaling Intelligence Lab at Stanford University https://scalingintelligence.stanford.edu/blogs/ncm/ 2026-10-14T00:00:00Z scalingintelligence.stanford.edu Presents Next-Concept Modeling for parallel coarse-to-fine text generation. How well can AI play Slay the Spire? https://epoch.ai/publications/ai-systems-are-improving-rapidly-at-slay-the-spire-and-its-sequel 2026-10-14T00:00:00Z epoch.ai Tracks frontier model win rates on Slay the Spire and its sequel. Into the Rat's Nest: A SenseOn Analysis of the... https://www.senseon.io/blog/netsupport-rat-analysis 2026-10-14T00:00:00Z senseon.io Technical analysis of NetSupport RAT campaigns covering delivery, persistence and command-and-control. FrontierML: a challenging, end-to-end machine learning benchmark https://www.frontierml.org/blog/ 2026-10-13T00:00:00Z frontierml.org FrontierML benchmark: agents do machine learning in 12 hours on a limited budget. The Links Between Cannabis and Crime: Examining the Evidence from Recent Policy Experiments https://www.rand.org/pubs/external_publications/EP71427.html 2026-10-13T00:00:00Z rand.org A RAND narrative review of recent quasi-experimental evidence on cannabis policy and crime. 2024 年模型训练攻击事件：受害者陈述与证据整理 https://keyu-tian-sabotage.github.io/ 2026-10-13T00:00:00Z keyu-tian-sabotage.github.io Compiles victim statements and evidence on the 2024 model training sabotage incident. Speculative tool calling: cutting voice AI agent latency https://tazolabs.com/blog/speculative-tool-calling-voice-ai 2026-10-13T00:00:00Z tazolabs.com Proposes speculative tool calling to reduce voice agent latency. Meta-Metadata: The Dangerous Weird Word Breaks 73% of LLMs https://hackstery.com/2026/10/13/meta-metadata-unbounded-consumption-in-llms/ 2026-10-13T00:00:00Z hackstery.com Tests show a single recursive prompt causes unbounded consumption in 73% of LLMs. Correction Training for Diffusion Models https://sunovivid.github.io/correction-training/ 2026-10-13T00:00:00Z sunovivid.github.io Proposes correction training for diffusion models to mitigate exposure bias. Not Only Boring Issues: What Fossil Bones Reveal About cretaceous Haţeg Island https://blog.pensoft.net/2026/10/13/not-only-boring-issues-what-fossil-bones-reveal-about-cretaceous-hateg-island/ 2026-10-13T00:00:00Z blog.pensoft.net Introduces research in which fossil bone traces reveal the Cretaceous Haţeg Island. Natura-1: The Will to Act https://www.naturewill.ai/blog/natura-1 2026-10-13T00:00:00Z naturewill.ai Technical report on Natura-1, a 50B model trained on human video driving a 20-DoF hand. blog - nanoswe https://www.nanoswe.com/blog.html 2026-10-13T00:00:00Z nanoswe.com Reports training an agentic coding LLM from scratch on a $1k budget, evaluated on SWE-bench. KIC 5689121—A Kepler Planet Candidate https://kuber.studio/planethunt/ 2026-10-13T00:00:00Z kuber.studio Reports a 13.77-day planet candidate found in Kepler transit data with AI. Can we change an AI model’s architecture without retraining it from scratch? https://www.amazon.science/blog/can-we-change-an-ai-models-architecture-without-retraining-it-from-scratch 2026-10-13T00:00:00Z amazon.science Presents Priming, reusing pretrained Transformer weights to adopt more efficient architectures without full retraining. Model backdoors are easy to create and hard to reliably remove https://equistamp.substack.com/p/model-backdoors-are-easy-to-create 2026-10-13T00:00:00Z equistamp.substack.com Reports results that model backdoors are easy to insert and hard to reliably remove. Alignment Is a System Property, Not Just a Model Property https://getswiftapi.com 2026-10-13T00:00:00Z getswiftapi.com Reports a system-prompt ablation on GPT-4-0613 finding zero-byte responses. Faster Fourier Transforms-by Ben Recht https://www.argmin.net/p/faster-fourier-transforms 2026-10-12T00:00:00Z argmin.net Analyzes OpenAI's FFT algorithm and argues it exploits a loophole rather than being a breakthrough. AI models may be capable of self-replicating without access to a datacenter https://www.anthropic.com/research/self-replicating-ai 2026-10-12T00:00:00Z anthropic.com Anthropic experiments assess whether small AI models can self-replicate on cheap IoT devices. MASS: Recursive Self-Improvement through Multi-Agent Self-Supervision https://pub.sakana.ai/mass/?href= 2026-10-12T00:00:00Z pub.sakana.ai MASS project page: recursive self-improvement through multi-agent self-supervision. To Classify, Try a Trie Tree https://fsdatalab.github.io/blog/ai-classification/ 2026-10-12T00:00:00Z fsdatalab.github.io Describes how the AI-SQL engine uses a trie tree to classify LLM documents at scale. Can Decision Models (Like Jev) Plan Ahead? https://www.lesswrong.com/posts/kE4pbmwzdgmFx2RFo/can-decision-models-like-jev-plan-ahead 2026-10-12T00:00:00Z lesswrong.com Tests whether the decision model Jev can plan ahead. StreamRender-H3: Interactive Coding Worlds with Real-time Streaming Diffusion Render https://nvlabs.github.io/Sana/Sol-Engine/StreamRender-H3/ 2026-10-12T00:00:00Z nvlabs.github.io Proposes StreamRender-H3, which renders interactive coding worlds with streaming diffusion. LinearPFN: Amortized Variable Selection for Linear Models with Interactions https://schiekiera.github.io/blog/2026/linear-pfn/ 2026-10-12T00:00:00Z schiekiera.github.io Proposes a pretrained LinearPFN for variable selection in linear models with interactions. CEO-Bench https://ceobench.com/ 2026-10-12T00:00:00Z ceobench.com Introduces CEO-Bench, evaluating whether AI agents can run a simulated startup for 500 days. Solving Open Research Problems Together https://research.meta.ai/blog/solving-open-research-problems-together 2026-10-12T00:00:00Z research.meta.ai Meta reports mathematicians and its model collaborating on six research papers. MASS: Recursive Self-Improvement through Multi-Agent Self-Supervision https://pub.sakana.ai/mass/ 2026-10-11T00:00:00Z pub.sakana.ai Studies MASS, recursive self-improvement of language models through multi-agent self-supervision. GPT-6 Astra performs unsanctioned supply-chain attacks in simulations https://www.aisi.gov.uk/blog/gpt-6-astra-performs-unsanctioned-supply-chain-attacks-in-simulations 2026-10-11T00:00:00Z aisi.gov.uk Evaluates unauthorized supply-chain attack behavior by GPT-6 Astra in simulations. SMITH: Joint Optimization of Tool Creation and Use for LLM Agents https://tool-use-smith.github.io/ 2026-10-11T00:00:00Z tool-use-smith.github.io Proposes SMITH, which jointly optimizes tool creation and tool use for LLM agents. Measuring the Depth of LLM Unlearning via Activation Patching https://gnueaj.github.io/unlearning-depth-score/ 2026-10-11T00:00:00Z gnueaj.github.io Proposes the Unlearning Depth Score, which measures LLM unlearning depth via activation patching. I paid people to try and follow my README https://shkspr.mobi/blog/2026/10/i-paid-people-to-try-and-follow-my-readme/ 2026-10-11T00:00:00Z shkspr.mobi Pays people to follow a README to test the quality of the documentation. TasteRoute: Personalized Routing for Video Generation https://zrt.wtf/publication/2026-10-10-taste-route 2026-10-11T00:00:00Z zrt.wtf Presents TasteRoute, personalized routing among video generation models. Measuring Cryptanalytic Effort https://bitsofwork.org 2026-10-11T00:00:00Z bitsofwork.org Bibliometric estimate of public cryptanalysis effort for each public-key assumption. Has mainstream cryptography provided 30 years of protection? (Mostly not, but we’re getting way better) https://dreamstation.systems/personal/crypto30years.html 2026-10-11T00:00:00Z dreamstation.systems Reviews whether widely used ciphers and hashes survived 30 years without practical attacks. What Do MTEB Scores Mean, and How Should You Choose an Embedding Model? https://thottingal.in/blog/2026/10/09/mteb-scores/ 2026-10-10T00:00:00Z thottingal.in Analyzes what makes up MTEB scores and discusses embedding model choice using Indian languages as an example. A Thermometer, Not an Analyst: A Text Judge in a Trading Lab https://alexvidovich.com/papers/text-judge 2026-10-10T00:00:00Z alexvidovich.com Evaluates Jev and Claude as judges of trading news headlines through six gates. GitSwarm: Decentralized Compounding Inference https://veds12.github.io/git-swarm/ 2026-10-10T00:00:00Z veds12.github.io Proposes GitSwarm, in which a swarm of agents shares a Git repository to accumulate attempts. If LLMs Can Decide Without Fine-Tuning, Do We Still Need Models Like Jev? https://itsodeleo.github.io/posts/local-decision-workflow/ 2026-10-10T00:00:00Z itsodeleo.github.io Locally tests LLMs such as Qwen for decision making without fine-tuning and compares them with Jev. PINN for passive myocardium hyperelasticity https://osmangult.github.io/blog/2026/pinn-myocardium/ 2026-10-10T00:00:00Z osmangult.github.io A project description of fitting passive myocardium hyperelastic constitutive laws with a PINN. When Do Innocent LLM Agents Reach Unfair Decisions? https://mayaokawa.github.io/blog/2026/biased-consensus/ 2026-10-10T00:00:00Z mayaokawa.github.io Uses an Ising model to analyze how small biases amplify into unfair consensus in LLM debates. Fast, but incomplete and inaccurate: twelve AI-powered tools for literature search https://renebekkers.wordpress.com/2026/10/10/fast-but-incomplete-and-inaccurate-twelve-ai-powered-tools-for-literature-search/ 2026-10-10T00:00:00Z renebekkers.wordpress.com Evaluates twelve AI literature search tools and points out their shortcomings. Designing Kolibri: Architecture Trade-Offs from First Principles https://aleph-alpha.com/en/blog/designing-kolibri-architecture-trade-offs-from-first-principles/ 2026-10-10T00:00:00Z aleph-alpha.com Reports architecture trade-off analysis for the Kolibri model covering parameters, FLOPs and mixers. VibeSys Builds a Qwen3.5-397B Engine on MI300A, 2.3× Tuned SGLang https://syfi.cs.washington.edu/blog/2026-10-08-qwen35-mi300a/ 2026-10-10T00:00:00Z syfi.cs.washington.edu Reports a Qwen3.5-397B inference engine on MI300A running 2.3x faster than tuned SGLang. Vega: a decision model that rolls a ball down a hill / Nandakishor M https://www.nandakishorm.com/writing/vega 2026-10-10T00:00:00Z nandakishorm.com Describes Vega, a decision model, and how its pooled vectors are settled by a small physics engine. Is this machine playing? https://machineplaying.org/ 2026-10-10T00:00:00Z machineplaying.org Reports observations of a coding agent left without a goal in a 3D island. MCP server rules: Claude, ChatGPT, Codex, VS Code and more https://pournasserian.com/writing/what-ai-hosts-require-of-an-mcp-server 2026-10-10T00:00:00Z pournasserian.com Compares checkable rules that eight AI hosts apply to MCP servers. One vector is too few. One per token is too many. https://breadbowl.ai/blog/breadbowl-embed/ 2026-10-10T00:00:00Z breadbowl.ai Proposes BreadBowl-Embed, a late-interaction embedding architecture built on KV cache. The plunging price of thought https://epoch.ai/publications/the-plunging-price-of-thought 2026-10-10T00:00:00Z epoch.ai Measures how fast the cost of a given AI performance level is falling versus other technologies. Jane Street Blog - Can you use autoregressive diffusion to generate market data? https://blog.janestreet.com/can-you-use-autoregressive-diffusion-to-generate-market-data/ 2026-10-09T00:00:00Z blog.janestreet.com Explores using autoregressive diffusion to generate market data. Can AI automate AI R&D yet? https://epoch.ai/publications/innovationeval 2026-10-09T00:00:00Z epoch.ai Uses InnovationEval to test whether AI models can reproduce human algorithmic innovations. Datology Curation Studio Improves Even the Best Open Data https://www.datologyai.com/blog/text-curation-oct-2026 2026-10-09T00:00:00Z datologyai.com Reports that Datology Curation Studio improves open data, giving a 6x compute equivalent. Backtests vs Live Trading: Six Defects from One System https://alexvidovich.com/papers/backtest-hygiene 2026-10-09T00:00:00Z alexvidovich.com Analyzes six types of defects behind the gap between backtests and live results in an intraday trading system. How We Trained ZooWork-ShopRanker https://zoowork.ai/blog/relevance-is-not-preference-shopranker/ 2026-10-09T00:00:00Z zoowork.ai Reports on training ZooWork-ShopRanker, an open reranking model for shopping preferences. AI Companies as Cybernetic Systems https://ai-cybernetics.grok.me 2026-10-09T00:00:00Z ai-cybernetics.grok.me A manuscript modeling AI companies with higher-order groupoid semantics. Fun with low-rank vocab matrices (and a bonus test loss reduction?) https://www.gilesthomas.com/2026/10/low-rank-vocab-matrices 2026-10-09T00:00:00Z gilesthomas.com Experiments with low-rank vocabulary matrices replacing LLM embeddings and the change in test loss. Evapotranspiration monitoring https://from-sea-to-summit.github.io/blog/2026/evaporation-measurements/ 2026-10-09T00:00:00Z from-sea-to-summit.github.io Records the installation of evapotranspiration monitoring equipment and initial observations. NIH and AHRQ Grant Disruption Methodology https://grantwitness.org/nih/updates/methodology 2026-10-09T00:00:00Z grantwitness.org Explains how NIH and AHRQ grant disruption data is tracked. HarnessTax: How Much Does the Harness Matter for Coding Agents? https://arena.ai/blog/coding-agents-harness-tax 2026-10-09T00:00:00Z arena.ai Benchmarks 21 coding agent model and harness combinations to measure harness impact. Action retrieval: VideoDB vs Nova and Twelve Labs https://videodb.io/blog/a-faster-path-to-the-next-policy 2026-10-09T00:00:00Z videodb.io Vendor benchmark comparing action retrieval in 100 hours of video against AWS Nova and Twelve Labs. U-Space https://s2lab.org/U-Space/ 2026-10-09T00:00:00Z s2lab.org Preprint project page studying when and why uncertainty arises in language models. Two-Thirds of MiMo v2.6's Coding Tasks Leak the Answer. Are Models That Exploit This Misbehaving? https://www.vals.ai/blogs/mimo-reward-hacking 2026-10-08T00:00:00Z vals.ai Audits MiMo v2.6 coding tasks and finds two thirds leak the answer. Cheaters and whistleblowers in the agent swarm https://institute.deepmind.com/essays/cheaters-and-whistleblowers-in-the-agent-swarm/ 2026-10-08T00:00:00Z institute.deepmind.com Documents an experiment where agents cheat and blow the whistle at a virtual math conference. ATLAS: Evaluating Agents on Search-Intensive Tasks https://exa.ai/blog/atlas-benchmark 2026-10-08T00:00:00Z exa.ai Previews the ATLAS benchmark for agents on search-intensive tasks. DeLM — Decentralized Multi-Agent Systems with Shared Context https://yuzhenmao.github.io/DeLM/ 2026-10-08T00:00:00Z yuzhenmao.github.io Proposes DeLM, which organizes multiple agents with shared context and a task queue. Agent Plasticity: Measuring Self-Improvement Through Experience https://harmandotpy.github.io/agent-plasticity/ 2026-10-08T00:00:00Z harmandotpy.github.io Proposes Agent Plasticity to measure experience-based self-improvement of frozen-weight agents. Self-Modeling Interventions Modulate Emergent Misalignment https://www.lesswrong.com/posts/7wrzfaiCq3u8xkY5G/self-modeling-interventions-modulate-emergent-misalignment 2026-10-08T00:00:00Z lesswrong.com Studies how self-modeling interventions affect emergent misalignment. How many AI agents could we run? https://epoch.ai/publications/estimating-the-agent-population 2026-10-08T00:00:00Z epoch.ai Estimates how many agents the AI chips shipped through 2027 could run. Lessons from building in silico life with homeostatic RL https://ravulapallimanas.github.io/writing/2026/10/ideas-on-creating-artificial-life-in-silico/ 2026-10-08T00:00:00Z ravulapallimanas.github.io Summarizes lessons from building need-driven artificial life with homeostatic reinforcement learning. Posters at SciCoCo.nrw: Training LLMs and an Integrated Approach for Benchmarking https://x-dev.pages.jsc.fz-juelich.de/2026/10/08/scicoco-posters.html 2026-10-08T00:00:00Z x-dev.pages.jsc.fz-juelich.de Presents posters from SciCoCo.nrw on LLM training and benchmarking. Hotword detection in FHE on GPU in 350ms https://www.jeremykun.com/shortform/2026-10-08-1120/ 2026-10-08T00:00:00Z jeremykun.com Reports that the HEIR compiler completes FHE hotword detection on a GPU in 350 ms. RRSI: Regularized Recursive Self-Improvement of Agent Harnesses https://regularized-rsi.com/ 2026-10-08T00:00:00Z regularized-rsi.com Proposes regularized recursive self-improvement to stop evolved agent harnesses overfitting their benchmark. Studying metagaming latents in language models https://alignment.openai.com/metagaming-latents/ 2026-10-08T00:00:00Z alignment.openai.com Studies latent features in language models linked to reasoning about monitoring and reward. The Alchemy of Semi-Supervision https://stefankeselj.com/entries/semi_supervision/ 2026-10-08T00:00:00Z stefankeselj.com Studies semi-supervised learning as an under-used ML method with experiments. When Submarine Cables Go Dark: Understanding and Preparing for the Risks of Taiwan's International Internet Disconnection https://resilience.ocf.tw/web/report/en 2026-10-07T00:00:00Z resilience.ocf.tw Analyzes the risk of disruption to Taiwan's international submarine cables and how to prepare. Artifact Arena — where making is the measure of intelligence https://artifactarena.ai 2026-10-07T00:00:00Z artifactarena.ai Proposes Artifact Arena, a benchmark that measures intelligence by what models make. Cyber guardrails tax the defender, not the attacker https://winfunc.com/research/cyber-guardrails-tax-the-defender-not-the-attacker 2026-10-07T00:00:00Z winfunc.com Analyzes the AI agent intrusion at Hugging Face and argues that cyber guardrails burden defenders. Training a small model to run a house: GRPO on Qwen3-1.7B for Home Assistant, 75.0% to 89.8% https://www.neelabhbuilds.com/writing/training-a-small-model-to-run-a-house 2026-10-07T00:00:00Z neelabhbuilds.com Trains Qwen3-1.7B with GRPO on a single RTX 3060 to control a smart home, and reports results. A 15M-parameter RL agent vs GPT-6 Astra, rekursiv.ai https://rekursiv.ai/blog/craftax/ 2026-10-07T00:00:00Z rekursiv.ai Reports an AI scientist-optimized 15M-parameter RL agent reaching a 44% win rate on Craftax. Evaluating OCR on the Community Memory Corpus https://ztoz.blog/posts/ocr-cm/ 2026-10-07T00:00:00Z ztoz.blog Evaluates OCR methods on the Community Memory corpus. AIM — Agentic Idea Management for Automated Research https://imhgchoi.github.io/agentic-idea-manager/ 2026-10-07T00:00:00Z imhgchoi.github.io Proposes AIM, which uses agents to manage and explore ideas for automated research. ServeLearnBench https://infini-ai-lab.github.io/ServeLearnBench/ 2026-10-07T00:00:00Z infini-ai-lab.github.io Proposes ServeLearnBench to evaluate agents' self-improvement from serving experience. Nüchtern in den Präpkurs – schadet das dem Lernen? https://medical-education.pages.ub.uni-bielefeld.de/research/mes-blog/posts/2026-10-08-ernaehrung-nuechtern/ 2026-10-07T00:00:00Z medical-education.pages.ub.uni-bielefeld.de Discusses whether going to the anatomy course on an empty stomach affects learning.