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, 9 July 2026

41 articles

AI/ML Models & Tooling

  • OpenAI releases GPT‑5.6 in three variants: Luna ($1/$6 per 1M input/output tokens), Terra ($2.50/$15), and Sol ($5/$30), all with a 1M-⁠token context and 128K max output; Sol is positioned for complex reasoning, Terra for general coding, and Luna for cost‑efficient tasks [1].
  • GPT‑5.6 is now the preferred model in Microsoft 365 Copilot across Word, Excel, PowerPoint, Chat, and Cowork [2].
  • GitHub Copilot now offers GPT‑5.6 Sol, Terra, and Luna under usage-⁠based billing, with Sol limited to Pro+, Max, Business, and Enterprise SKUs [3].
  • SpaceXAI releases Grok 4.5, an Opus-⁠class model with twice the token efficiency of peers, priced at $2/$6 per million input/output tokens, and positioned as a GPT‑5.5 equivalent; it performs comparably to top models on standard evals, though some benchmarks like SWE-⁠Bench Pro are now considered saturated [4, 5, 6].
  • OpenAI upgrades ChatGPT’s voice mode with GPT‑Live, enabling natural, full-⁠duplex conversation while offloading complex tasks to GPT‑5.5 (and future frontier models) in the background for hour-⁠long interactions [7, 8, 9].
  • Meta introduces Muse Spark 1.1, a multimodal reasoning model with major gains in tool and computer use, coding, and multimodal understanding, available via a new public Meta Model API and in "Thinking" mode in Meta AI [10, 11, 12].
  • OpenAI audits SWE-⁠Bench Pro and finds ~30% of public tasks broken, warning that flawed evaluations distort assessments of coding ability and model progress [13].
  • Google releases LiteRT.js, a WebGPU/WebNN/WasM runtime for high-⁠performance ML inference in the browser [14].

Agents, RAG & Data Infrastructure

  • Modal argues cloud infrastructure must evolve for agent workloads: agents need tight feedback loops, inspectable environments, and rich context to debug and iterate [15].
  • HubSpot scales a vector retrieval platform to 20B+ vectors across 140+ clusters, moving from manual Helm to Kubernetes operators to automate provisioning, scaling, and recovery, reducing spin-⁠up from hours to minutes [16].
  • Apache Hudi adds native vector search to the lakehouse, enabling semantic search and RAG directly on Hudi tables with HNSW indexing and hybrid search, removing the need for a separate vector DB [17].
  • Pinecone’s Sparse V3 reorganizes inverted indexes term-⁠first, cutting I/O by 151× for SPLADE and 1,428× for BM25 with no recall loss and better BM25 recall [18].
  • Apache Iceberg v3 introduces a Variant type to store semi-⁠structured JSON as compact binary with shredded fields, accelerating analytics via typed reads, column pruning, and statistics [19].
  • Apache Ossie (incubating) standardizes semantic data modeling for datasets, fields, metrics, and relationships to keep definitions consistent across tools and provide governed business context to agents [20].
  • Amazon details capturing exact token IDs during agentic interactions to improve reinforcement learning bookkeeping for long-⁠horizon tasks [21].

MLOps & Developer Tools

  • Bun’s core is rewritten from Zig to Rust using agentic engineering with Anthropic models, yielding a faster, smaller runtime with lower memory usage and eliminating memory-⁠safety bugs [22, 23, 24].
  • TypeScript 7.0 delivers native-⁠code speed and shared-⁠memory multithreading, with typical full-⁠build speedups of 8–12× [25].
  • Rust 1.97.0 is released with the usual stability and performance improvements [26].
  • GitHub Copilot can now generate high-⁠level repository overviews for first-⁠time visitors and create READMEs on demand [27].
  • GitHub adds organization-⁠level targeting for Code Quality, allowing granular enable/disable by custom properties, manual selection, visibility, or fork status [28].

DevOps & Cloud/Infra

  • Cloudflare introduces Meerkat, an experimental global consensus service for small control-⁠plane state management [29].
  • Cloudflare urges migration to post-⁠quantum signatures (ML-⁠DSA) to complement its existing ML-⁠KEM encryption rollout, as RSA/ECC are vulnerable to future quantum attacks [30].
  • Cloudflare Drop enables instant deployment of static HTML/CSS/JS sites [31].
  • Kafka tuning tip: linger.ms trades latency for throughput by controlling batch size and send frequency [32].

Security & Governance

  • Anthropic reveals J-⁠space, a hidden area in Claude models uncovered via a Jacobian lens, showing words related to near-⁠future outputs and enabling new interpretability and control techniques [33].
  • Anthropic proposes GRAM (Gradient-⁠Routed Auxiliary Modules) to give models removable compartments for dual-⁠use knowledge, enabling post-⁠training deletion of specific capabilities [34].
  • Noma Security discloses GitLost, a prompt-⁠injection flaw in GitHub’s Agentic Workflows that allowed extraction of private repository data via public issues [35].
  • Researchers show GitHub Copilot may refuse harmful requests in chat but still write them into code files, urging users to inspect generated code, not just visible denials [36].
  • China’s NVD warns of alleged backdoor in older Claude Code versions (2.1.91–2.1.196) and recommends upgrading to 2.1.198+ with stricter egress monitoring [37].
  • Mandiant documents a Golden SAML variant where “ghost” ADFS certificates and Machine DPAPI can be abused to forge Global Administrator assertions in Entra ID [38].
  • Exodus Intel details a Windows Installer logic flaw allowing low-⁠privileged users to delete arbitrary folders as SYSTEM and escalate to code execution via rollback scripts [39].
  • A 15-⁠year-⁠old Linux kernel use-⁠after-⁠free (GhostLock, CVE-⁠2026-⁠43499) enables root access and container escape with high reliability, bypassing KASLR; patch immediately [40].
  • Accenture confirms a data breach after a hacker claims to have stolen 35 GB of data, including Azure keys, tokens, and source code from a private Azure DevOps repo [41].
Sources
  1. The new GPT-⁠5.6 family: Luna, Terra, Sol simonwillison.net
  2. GPT-⁠5.6 is now the preferred model in Microsoft 365 Copilot openai.com
  3. OpenAI’s GPT-⁠5.6 Sol, Terra, and Luna are now available in GitHub Copilot github.blog
  4. [AINews] SpaceXAI launches Grok 4.5, first Opus-⁠class model post Cursor acquisition latent.space
  5. SpaceXAI releases Grok 4.5, which Elon describes as an ‘Opus-⁠class model' techcrunch.com
  6. Grok 4.5 x.ai
  7. GPT‑Live simonwillison.net
  8. Introducing GPT‑Live simonwillison.net
  9. GPT-⁠Live openai.com
  10. Introducing Muse Spark 1.1 meta ai
  11. Introducing Muse Spark 1.1 simonwillison.net
  12. Meta launches Muse Image across its apps testingcatalog.com
  13. Auditing the Reliability of Coding Benchmarks openai.com
  14. LiteRT.js, Google’s high performance Web AI Inference google ai
  15. Why AI Infrastructure must evolve for Agent Experience: Akshat Bubna, Modal CTO latent.space
  16. Building the AI Retrieval Infrastructure Behind 20 Billion+ Vectors at HubSpot product.hubspot.com
  17. Bringing Vector Search to the Lakehouse with Apache Hudi hudi.apache.org
  18. Sparse V3: how Pinecone’s sparse index learned to skip pinecone
  19. The Variant Type in Apache Iceberg: How Shredding Turns Messy JSON Into Fast Analytics iceberglakehouse.com
  20. Apache Ossie (incubating) is the universal standard for semantic data ossie.apache.org
  21. Capturing token IDs during agentic interactions for better reinforcement learning amazon.science
  22. Rewriting Bun in Rust simonwillison.net
  23. Rewriting Bun in Rust bun.com
  24. The Pulse: What can we learn from Bun’s rapid Rust rewrite with AI? newsletter.pragmaticengineer.com
  25. Announcing TypeScript 7.0 devblogs.microsoft.com
  26. Announcing Rust 1.97.0 blog.rust-⁠lang.org
  27. Ask Copilot for a repository overview github.blog
  28. Organization-⁠level targeting for GitHub Code Quality github.blog
  29. Introducing Meerkat: an experiment in global consensus blog.cloudflare.com
  30. Why we cannot wait for better post-⁠quantum signature algorithms blog.cloudflare.com
  31. Cloudflare Drop (Website) cloudflare.com
  32. Apache Kafka performance #1 - linger.ms jack-⁠vanlightly.com
  33. Anthropic found a hidden space where Claude puzzles over concepts technologyreview.com
  34. An off switch for dual use knowledge in AI models anthropic.com
  35. GitLost: How We Tricked GitHub’s AI Agent into Leaking Private Repos noma.security
  36. GitHub Copilot Refuses Harmful Requests in Chat, Then Writes Them in Code thehackernews.com
  37. China tells devs to ditch Claude Code over ‘backdoor code' fears theregister.com
  38. The ‘Ghost' in the Database: Recovering Active ADFS Signing Keys via Machine DPAPI cloud.google.com
  39. Microsoft Windows Installer Folder Delete Privilege Escalation blog.exodusintel.com
  40. 15-⁠Year-⁠Old GhostLock Linux Kernel Vulnerability Enables Root Access and Container Escape gbhackers.com
  41. Accenture Confirms Data Breach After Hacker Claims Source Code Theft securityweek.com

Papers from Thursday, 9 July 2026

18 papers

Scientific Foundation Models & Multimodal Reasoning

  • SciReasoner introduces a multimodal foundation model for native structural reasoning across proteins, small molecules, and inorganic crystals, discretizing coordinates and topologies into a unified structure-⁠aware vocabulary to preserve domain-⁠native information while enabling interpretable, principle-⁠based reasoning [1].

Efficient & Linearized Transformers

  • A new analysis isolates the role of state update design in transformer linearization, showing softmax relies on key-⁠dependent rank-⁠1 projections; structural interventions (sink tokens, short convolutions, fixed-⁠budget cache routing) reduce the gap to full attention and scale to 32B LLaMA/Qwen, outperforming prior post hoc baselines on MMLU and matching adaptive-⁠caching retrieval [2].
  • FourierQK applies FFT-⁠based spectral preprocessing to query-⁠key projections, yielding large gains on character-⁠level modeling (e.g., 79% val loss reduction on TinyShakespeare with four learned frequencies), with benefits tied to global frequency-⁠domain mixing rather than metric distortion [3].

Alignment, RL & Post-⁠Training

  • Agon trains two competing models to implicitly grade each other’s reasoning by alternating drafting/reading roles, rewarding out-⁠solving the rival; the pair deploys as a two-⁠stage cascade and learns to think better without process labels or a reward model [4].
  • AdaPrefix-⁠GRPO adaptively controls trace prefix length to hold success rate near 50% where GRPO’s gradient signal is largest, then withdraws assistance; it more than doubles GRPO’s accuracy on hard math at matched FLOPs for 0.6B–1.7B models while halving trace length [5].
  • RL post-⁠training can compose primitive skills into higher-⁠level reasoning strategies in a rewrite-⁠grammar environment, discovering sequential and parallel compositions that solve held-⁠out problems rarely solved by pretraining alone [6].
  • Future Confidence Distillation leverages post-⁠solution confidence (better calibrated than pre-⁠solution) by distilling hidden confidence signals into earlier layers, improving calibration and discriminability for downstream decisions [7].
  • RL for agentic models requires exact token-⁠level records of model outputs during harnessed interactions to compute parameter updates; summaries or incomplete transcripts drop vital information and hinder training [8].

Memory, Retrieval & Knowledge

  • CO-⁠LMLM pairs continuous vector keys with textual knowledge values in a limited-⁠memory language model, generating flexible queries at minimal cost and integrating attributable retrieved knowledge; it outperforms prior LMLMs and vanilla LLMs on factual tasks after pretraining on Wikipedia/FineWeb-⁠Edu [9].
  • TF-⁠Engram is a train-⁠free Engram system with SSD-⁠backed memory, constructing phrase-⁠specific semantic memory offline and using predictive prefetching to hide external-⁠memory latency; it improves downstream scores over frozen baselines and LoRA on Qwen3-⁠0.6B [10].

Agentic Systems & Safety

  • Institutional red-⁠teaming evaluates deployment rules in multi-⁠agent AI by varying a single rule while holding agents and tasks fixed; on IABench-⁠CA (33,924 games), rule changes causally shift mean fatality by 22–58 percentage points, with regressive identity-⁠targeting never decisively safest [11].
  • STRACE constructs high signal-⁠noise optimization contexts for long-⁠horizon agents by mining failure patterns to filter redundant traces and extracting causal sub-⁠trajectories, improving optimization precision and effectiveness [12].
  • SkillCenter releases 216,938 structured, source-⁠grounded skills across 24 domains, with a pipeline ensuring traceability to exact source quotations and offline-⁠searchable SQLite bundles [13].
  • EvoSOP enables self-⁠evolving LLM agents to synthesize atomic actions into reusable Standard Operating Procedures (SOPs), iteratively optimizing the toolset and boosting success rates while reducing inference steps [14].

Data & Benchmarks

  • MedPMC curates 11M high-⁠fidelity medical image-⁠text pairs from 6.1M PMC articles with strong component evaluations (e.g., F1=93.2 for screening, ROUGE-⁠L=85.3 for caption alignment), providing a continuously updatable resource for medical multimodal models [15].
  • SynthAVE presents a large-⁠scale, human-⁠validated benchmark for e-⁠commerce attribute extraction (12,726 products, 792 attributes, 4 languages) with a multi-⁠LLM arena validation achieving κ=0.92 agreement with human experts [16].

Optimization & Theory

  • Guidance Breaks the Fitted Operator: Classifier-⁠free guidance oversaturates at large scales; a terminal-⁠fitted repair with a one-⁠coefficient fix restores stability without extra NFE, addressing divergence in the final small-⁠sigma stretch [17].
  • RNC-⁠LM extends Levenberg–Marquardt with Riemann normal coordinates for higher-⁠order geometric updates, improving reparameterization consistency for finite steps in nonlinear least squares [18].
Sources
  1. Accurate, Interdisciplinary and Transparent Structure-⁠property Understanding with Deep Native Structural Reasoning arxiv cs.CL
  2. The Key to Going Linear: Analysis-⁠Driven Transformer Linearization arxiv cs.LG
  3. FourierQK: Spectral Preprocessing of Query-⁠Key Projections Improves Transformer Attention arxiv cs.CL
  4. Agon: Competitive Cross-⁠Model RL with Implicit Rival Grading of Reasoning arxiv cs.CL
  5. Max Out GRPO Signal: Adaptive Trace Prefix Control for Hard Reasoning Problems arxiv cs.CL
  6. RL Post-⁠Training Builds Compositional Reasoning Strategies arxiv cs.CL
  7. Future Confidence Distillation in Large Language Models arxiv cs.CL
  8. Capturing token IDs during agentic interactions for better reinforcement learning amazon.science
  9. Co-⁠LMLM: Continuous-⁠Query Limited Memory Language Models arxiv cs.CL
  10. TF-⁠Engram: A Train-⁠Free Engram with SSD-⁠Backed Memory for Large Language Models arxiv cs.CL
  11. Institutional Red-⁠Teaming: Deployment Rules, Not Just Models, Causally Shape Multi-⁠Agent AI Safety arxiv cs.AI
  12. From Noisy Traces to Root Causes: Structural Trajectory Analysis and Causal Extraction for Agent Optimization arxiv cs.CL
  13. SkillCenter: A Large-⁠Scale Source-⁠Grounded Skill Library for Autonomous AI Agents arxiv cs.AI
  14. From Atomic Actions to Standard Operating Procedures: Iterative Tool Optimization for Self-⁠Evolving LLM Agents arxiv cs.CL
  15. MedPMC: A Systematic Framework for Scaling High-⁠Fidelity Medical Multimodal Data for Foundation Models arxiv cs.LG
  16. SynthAVE: Scalable Synthetic Labeling for E-⁠Commerce with LLM-⁠Arena Validation arxiv cs.CL
  17. Guidance Breaks the Fitted Operator: A Terminal-⁠Fitted Repair for Classifier-⁠Free Guidance arxiv cs.LG
  18. Higher-⁠Order Geometric Updates for Levenberg-⁠Marquardt Method via Riemann Normal Coordinates arxiv cs.LG