Skip to content

AI Digest for

This page is the archived edition of one specific day. It does not change any more.

Latest digest Subscribe

News from Thursday, 15 October 2026

43 articles

AI Models, Industry and Business

  • Anthropic told investors its third-⁠quarter revenue rose about 45% from the prior quarter to $16.8 billion, ahead of a possible IPO that could value it at $2 trillion [1]. Its distillation clampdown accounted for 5% to 7% of revenue-⁠related usage, and it reported 11% adjusted operating margins [2].
  • Anthropic gave METR employee-⁠like access to evaluate its models and will publish the results of its risk assessments, while also working with SecureBio, Accenture and early talks with Hugging Face [3].
  • Modal says it has crossed $1B in annualized run rate and raised a $16B Series D, with usage and valuation tripling in five months [4].
  • Cohere announced a proposed agreement to acquire AISLE, an autonomous vulnerability-⁠finding company, to strengthen its sovereign AI platform and European presence [5].
  • Meta released the weights of Muse Spark Open, an 816B total / 42B active parameter MoE model under the Muse Model License, optimized for SGLang and vLLM on NVIDIA, AMD and Intel, with a 1M-⁠token context window [6]. It is available in limited preview on GroqCloud, where Meta reports 350+ tokens per second per user [7].
  • Mistral released Timestral, a family of time-⁠series foundation models in four sizes (5M to 209M parameters); the largest ranks first on GIFT-⁠Eval, TIME and fev-⁠bench with two-⁠thirds of the prior state-⁠of-⁠the-⁠art parameter count [8].
  • Google’s Harness² lets agents iteratively improve prompts, tools, memory and orchestration without changing model weights, carrying evidence-⁠backed lessons forward to future tasks [9].
  • A Substack analysis says more than half of businesses now pay for AI and growth will increasingly come from deeper adoption rather than new customers [10].

LLM Tooling, RAG and Agents

  • Jev’s team reports it picked the right page first 88% of the time across 168 test questions, versus 70% for embedding search, and cut irrelevant results from 33% to 3% by answering "none" when no page fit [11].
  • Phoenix is an open-⁠source, self-⁠hosted observability platform that captures OpenTelemetry traces and offers retrieval and response evaluations, datasets, experiments and prompt replay for LLM applications [12].
  • LangChain’s agent eval guide recommends starting with 10 to 20 examples drawn from traces, combining code evaluators with LLM judges and human review, and using offline and online evals at different stages [13].
  • Google’s Evaluating AI agents guide argues that agents can loop or burn tokens while still looking fine, and builds an eval pipeline for a LangGraph agent [14].
  • XERJ is an Apache-⁠2.0, Elasticsearch-⁠compatible local search engine that auto-⁠indexes code, logs, PDFs and CSVs for agent retrieval and RAG, with optional reranking that can send text to an external provider [15].
  • A proof of concept shows row-⁠level security for Claude Dashboards via MCP, with an MCP server authenticating users and querying DuckDB so access rules apply before data reaches Claude [16].
  • Docker’s guide describes agent risks such as prompt injection, tool misuse and excessive autonomy, and argues for containing what an agent can reach and overseeing its actions [17].
  • Addy Osmani recommends agentic verification: end-⁠to-⁠end tests as ground truth, property-⁠based testing to hunt for forbidden states, and fast deterministic tests so agents can retry reliably [18].
  • Google’s adk-⁠recipes repo organizes ADK resources into core/, plugins/ and contrib/ layers, with installable vertical skills for coding assistants and deployment patterns on Google Cloud [19].
  • Anthropic’s Code Review is now generally available for Team and Enterprise; most reviews now finish about 3x faster (about 20 minutes down to 6 to 10 minutes), with admins able to tune reviews per repository [20].
  • GitHub’s agentic autofix for code scanning alerts is generally available: Copilot explores the codebase, proposes a fix, reruns CodeQL to validate it and opens a draft PR [21]. AI Scan for pull requests is also generally available, covering languages CodeQL does not yet support, as advisory findings [22].

Model Serving and Inference Infrastructure

  • Meta describes using opportunistic idle inference capacity to rank more ads or route requests to stronger models, scaling back quickly when that capacity disappears while protecting normal latency and reliability [23].
  • vLLM introduces a push-⁠based NIXL connector that moves post-⁠prefill KV transfer coordination off the critical path for disaggregated inference, improving TTFT and TPOT at high request rates [24].
  • Netflix described how it detects faulty GPUs and network hardware, automatically remediates unhealthy machines, and keeps distributed training jobs alive through hardware failures [25].
  • ExecuTorch now runs PyTorch models directly on Arm Cortex-⁠M CPUs with CMSIS-⁠NN acceleration; four MLPerf Tiny models run in 21.5 ms or less under 90 KiB of RAM on a Cortex-⁠M55, with a core runtime under 20 KiB of flash [26].
  • Pinecone BYOC added on-⁠demand capacity mode, where indexes share a capacity pool with dynamic sharding instead of up-⁠front shard sizing [27].

Data, Databases and DevOps

  • A logical-⁠replication walkthrough reports a PostgreSQL major upgrade with 181 ms of write interruption and 179 ms during rollback, with no data loss, using parallel old and new clusters [28].
  • SQLBuild is an open-⁠source SQL and Python framework that checks contracts and lineage before execution and supports safe renames and production diffs across Snowflake, Postgres, BigQuery and Databricks [29].
  • pg_vectorsearch is a PostgreSQL extension for approximate nearest neighbor search [30].
  • Benchmarks on 1,000 IDs found IN, ANY and VALUES joins performed similarly in PostgreSQL, so teams should verify with EXPLAIN on their own data [31].
  • GitHub’s Evergreen app automates quarterly outside-⁠collaborator access reviews and repo archival in a shared org, a pattern teams can adapt [32].
  • GitHub retired email notifications for repository push events on October 15, 2026; teams that need them can use a GitHub Actions workflow with their own SMTP provider [33].

Security

  • Ivanti warned of actively exploited Connect Secure zero-⁠days, an authentication bypass chained with two others to run commands as root on vulnerable devices [34].
  • Weaviate fixed a medium-⁠severity authorization flaw in nearObject search (v1.39.3 and later) that let callers learn distances to protected vectors in other collections; a CVE is pending [35].
  • Discourse resolved CVE-⁠2026-⁠55420, an ImageMagick-⁠based upload flaw allowing arbitrary file read and write, by making a default-⁠deny policy for ImageMagick delegates [36].
  • Chrome 155 fixed 37 flaws, including seven critical memory bugs, with at least 15 reported by Anthropic with Claude’s help [37]. XBOW’s agents built a Chrome full chain in six days for about $40,000 in API costs; defenders should update to builds with the V8 and Chromium fixes [38].
  • Meta says hackers took code and personal data of about 1,500 people in a GitHub breach, 99% of the stolen data being source code; it is rotating exposed secrets [39].

Developer Practices and Tooling

  • Google’s DPoP support for OAuth sender-⁠constrains refresh tokens, mitigating infostealer malware and adversary-⁠in-⁠the-⁠middle attacks, and Google Cloud STS can handle token delegation across agent chains [40].
  • A benchmark processed 7.6 GB of JSON in about 0.12 seconds using fastsimdjson on free-⁠threaded Python across 64 cores, roughly 66 GB/s, by parallelizing and building Python objects only for needed fields [41].
  • Jared Palmer argues that recognizing clean code remains useful, but knowing the right program shape before writing it is something AI cannot yet do [42].
  • Rust 1.99.1 fixes regressions including a missing wcslen symbol that broke some targets, notably UEFI projects using std [43].
Sources
  1. Anthropic Says Revenue Jumped 45% to $16.8 Billion in Third Quarter nytimes.com
  2. Anthropic tells investors quarterly revenue up nearly 50% to $16.8bn ahead of IPO simonwillison.net
  3. An update on independent evaluators anthropic news
  4. [AINews] Modal’s $1B ARR, $16B Series D makes it NYC’s newest AI Infra Decacorn, 3x usage and 3x valuation in 5 months, 15x in 1 yr latent.space
  5. Cohere and AISLE sign letter of intent to advance cyber-⁠resilient sovereign AI cohere
  6. Muse Spark Is Now an Open-⁠Weights Model pytorch.org
  7. Meta’s Muse Spark open-⁠weight release Is Now in Limited Preview on GroqCloud groq
  8. Introducing Timestral mistral
  9. Google Harness² for Recursive Self Improvement github.com
  10. Cracks in the AI Thesis Part 3 econlab.substack.com
  11. Jev handed our LLM the wrong page a tenth as often as our embeddings sayora.ai
  12. Phoenix github.com
  13. The Most Common Questions About Agent Evals, Answered langchain.com
  14. Evaluating AI agents: moving past the vibe check google ai
  15. XERJ github.com
  16. Need to Know: Implementing row-⁠level security for Claude Dashboards via Model Context Protocol thedatavist.substack.com
  17. AI Agent Security Risks: What Can Go Wrong When Agents Act on Their Own docker.com
  18. Agentic Verification addyo.substack.com
  19. Three layers, one repo: a tour of google/adk-⁠recipes google ai
  20. Code Review is now generally available claude
  21. Agentic autofix for code scanning alerts general availability github.blog
  22. AI Scan for pull requests generally available github.blog
  23. Turning Opportunistic AI Inference Capacity Into Better Ads Recommendations engineering.fb.com
  24. Push-⁠Based KV Transfer to Improve TTFT and TPOT for Disaggregated Inference in vLLM vllm.ai
  25. How Netflix Keeps Large GPU Jobs Running netflixtechblog.com
  26. Bringing ExecuTorch to Arm Cortex-⁠M CPUs pytorch.org
  27. On-⁠demand mode for Pinecone BYOC pinecone
  28. Zero-⁠downtime Postgres upgrades with logical replication tapoueh.org
  29. SQLBuild: safe refactoring for warehouse SQL github.com
  30. pg_vectorsearch github.com
  31. Four PostgreSQL arguments that came out even explainmeasured.com
  32. Staying Evergreen: How we automated secure access and repo hygiene in a shared org github.blog
  33. Email notifications for repository push events have been retired github.blog
  34. Ivanti Warns of Connect Secure Zero-⁠Days Exploited in Attacks bleepingcomputer.com
  35. Weaviate security release - Medium severity fix for a cross-⁠collection vector disclosure in nearObject (CVE pending assignment) weaviate
  36. Discourse ImageMagick Vulnerability Enables Arbitrary File Read Through Malicious Image Uploads gbhackers.com
  37. Chrome, Firefox Updates Patch 114 Flaws as Mozilla Unveils Firefox Enterprise securityweek.com
  38. Zero to Calc in Six Days: Inside XBOW’s Chrome Full Chain xbow.com
  39. Meta Says Hackers Took Code and Personal Data of 1500 People in GitHub Breach securityweek.com
  40. Announcing DPoP Support for Google OAuth: Securing Refresh Tokens google ai
  41. Parsing JSON at tens of gigabytes per second… in Python lemire.me
  42. Code you didn’t write jaredpalmer.com
  43. Announcing Rust 1.99.1 blog.rust-⁠lang.org

Papers from Thursday, 15 October 2026

24 papers

Long-⁠context efficiency and memory

  • Latent Compaction learns a compact, document-⁠specific KV cache through test-⁠time training while keeping model weights frozen, using only generic document prompts with no downstream questions or answers. It reports 96.7% accuracy on RULER at 32x compression, with gains over budget-⁠matched baselines on RULER, LongBench, and Phonebook Lookup [1].
  • FOLD is a training-⁠free speculative-⁠decoding method that keeps exact local positions and treats distant context as far memory at a shared position, with sharpened attention to counter dilution as memory grows [2].
  • CA (Continuum Attention) replaces discrete key-⁠value lookup with integration over a continuous piecewise-⁠linear memory curve, with closed-⁠form attention and parallel GPU implementations. It reports a 0.0200 nats/byte NLL improvement over softmax attention [3].
  • The COUNT approach stores precomputed n-⁠gram counts on host memory to scale n-⁠gram embeddings without adding learned parameters [4].

Agents, self-⁠improvement and research automation

  • SwarmRSI evolves scientific agent swarms through feedback-⁠guided mutation and complementary recombination of hypotheses and program components [5].
  • A study on automated research finds that agent iteration, parallel agents, and communicating agents all plateau as history accumulates. Periodically discarding search state and restarting from the best candidate is reported to push the same budget further [6].
  • AdaRSI fine-⁠tunes only the unreliable sub-⁠steps of an embodied executor, using self-⁠collected demonstrations, and reports improvements on all three RoboCasa365 suites with GR00T N1.5 [7].
  • HP-⁠RLVR scores each token under causal history and under a hindsight context with the realized future, yielding token-⁠level credit from a single rollout without group sampling [8].
  • VASPO combines a self-⁠teacher and a critic inside PPO, reporting that privileged information improves critic accuracy toward the end of responses [9].
  • CriticOPD has the teacher critique incorrect student rollouts so the student rewrites them, reporting gains over on-⁠policy distillation on six math benchmarks [10].
  • CIPR penalizes KL divergence between action distributions under two interfaces with uninformative action identifiers, improving robustness to interface perturbations in both supervised and RL post-⁠training [11].
  • Multi-⁠developer agentic systems reach 70% on GAIA when specialized agents (each under 35%) are combined, reported as 12 points above the median single-⁠developer system [12].

Robotics and embodied learning

  • Diffusion over humanoid policy weights: a linear autoencoder with a Diffusion Transformer generates actor-⁠network weights for the Unitree G1, and the generated networks are reported to be deployable for walking and dancing on the real robot [13].
  • Player-⁠1 post-⁠trains computer-⁠use models on 6,044 gameplay trajectories and builds sixteen synthetic 2D/3D environments for game control [14].
  • HiRe-⁠MoMa fine-⁠tunes a frozen mobile-⁠manipulation policy with hierarchical residual heads and expert-⁠guided value alignment using under 20% training coverage from sparse human interventions [15].

Evaluation, bias and safety

  • A bias audit across ten open models in eight languages finds stereotype scores fall below English levels in all 70 model-⁠language pairs, with drops exceeding rewording noise for Hindi, Korean and Swahili, and argues that lower scores do not imply fairer models [16].
  • PerSo is a persuasion benchmark built from Reddit ChangeMyView dialogues that separately evaluates cognitive and behavioral social intelligence in LLMs [17].
  • ChronoCulture (10,360 temporal pairs across 72 countries) finds that models predict a country’s attitudes at one time point but do not reliably capture how they change over time [18].
  • semCER is a severity-⁠weighted character error rate for OCR that charges each grapheme edit with an LLM-⁠judged severity, tracking reader ratings in English, Vietnamese and Chinese [19].
  • Manacá-⁠1B-⁠Instruct-⁠2 documents three negative DPO attempts and reports that its safety came from dedicated safety-⁠SFT, with a non-⁠monotonic over-⁠refusal curve as the general-⁠data fraction varies [20].

Scientific and formal reasoning

  • MolLean is a process-⁠verifiable Lean-⁠based language for chemical reasoning that checks molecular state transitions and rejects invalid actions without changing the state [21].
  • Mechanical verification for Lean autoformalization combines LLM semantic review with numerical and vacuity checks, reporting the lowest accepted-⁠error rate among evaluated methods on three labeled datasets [22].
  • Same Claim, Different Interpretations introduces Claim Cards that record a verifier’s interpretation of a scientific claim as nine fields before evidence gathering [23].

AI for life sciences

  • OpenAI introduces the Rosalind Academic Plan, which offers eligible academic labs sponsored access to GPT-⁠Rosalind and tools for life sciences research [24].
Sources
  1. Latent Compaction with Test-⁠Time Training arxiv cs.CL
  2. FOLD: Position-⁠Free Far Memory with Local Order for Long-⁠Context Speculative Decoding arxiv cs.CL
  3. Continuum Attention arxiv cs.LG
  4. The COUNT: Counting N-⁠grams Unlocks Scaling on Host Memory and Training-⁠Free Adaptation arxiv cs.CL
  5. SwarmRSI: Recursive Self-⁠Improvement through Evolving Scientific Agent Swarm arxiv cs.CL
  6. Scaling Discovery Requires Forgetting arxiv cs.CL
  7. AdaRSI: On-⁠Demand Recursive Self-⁠Improvement for Embodied Agents arxiv cs.CL
  8. Hindsight-⁠Privileged RLVR: Single-⁠Rollout Token Credit Without Group Sampling arxiv cs.AI
  9. VASPO: Self-⁠Distillation with Privileged Value Estimation arxiv cs.CL
  10. Recovering Reliable Supervision via Critique-⁠Guided On-⁠Policy Distillation arxiv cs.LG
  11. One Task, Many Interfaces: Simple Cross-⁠Interface Consistency Training Improves LLM Agent Robustness arxiv cs.CL
  12. The Wealth of Agents: Scaling Agentic AI Across Developers arxiv cs.AI
  13. Modeling and Understanding Neural Weights of Humanoid Policies with Diffusion arxiv cs.LG
  14. Player-⁠1: Learning to Play from Synthetic Experience arxiv cs.AI
  15. Hierarchical Residual Policy Learning for Real-⁠World Mobile Manipulation with Sparse Human Guidance arxiv cs.LG
  16. Lost in Translation: Why Lower Bias Scores Don’t Mean Fairer Models arxiv cs.CL
  17. PerSo: Investigating Facets of LLMs' Social Intelligence in Real-⁠World Persuasion Scenarios arxiv cs.CL
  18. ChronoCulture: Evaluating Cultural Awareness Across Time in Large Language Models arxiv cs.CL
  19. Equal CER, Unequal Damage: Severity-⁠Weighted Character Error Rate for OCR arxiv cs.CL
  20. Manacá-⁠1B-⁠Instruct-⁠2: Reproducible Safety Alignment for a Native Brazilian-⁠Portuguese LLM. Negative DPO Results, a Controlled Composition Scan, and a Silent Failure Mode of the LLM Judge arxiv cs.CL
  21. MolLean: A Process-⁠Verifiable Language for Chemical Reasoning arxiv cs.CL
  22. Beyond LLM Judge: Mechanical Verification for Agentic Autoformalization arxiv cs.AI
  23. Same Claim, Different Interpretations: How LLM Agents Resolve Ambiguity in Scientific Claims arxiv cs.CL
  24. Bringing GPT-⁠Rosalind to academic research labs openai research