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 Friday, 11 September 2026

36 articles

AI Models, Agents, and Tooling

  • OpenAI releases a public beta of its Agents API, a managed harness for long-⁠running agents with context management, tool use, subagent coordination, file/environment interactions, and persistent sessions [1] [2].
  • OpenAI’s Habitat storage platform now serves over 1B ChatGPT users and 22M requests/sec, evolved from a Python library into a globally distributed system [3].
  • Google’s ADK for Kotlin 1.0 achieves parity with Python/Java, enabling idiomatic multi-⁠agent development with KMP, zero-⁠reflection KSP, human-⁠in-⁠the-⁠loop workflows, and Android-⁠first extensions (LiteRT-⁠LM, Firebase AI, Room, AppSearch) [4].
  • Google highlights four engineering patterns behind top AI agent submissions: bidirectional MCP for inter-⁠agent comms, async event buses for parallelism, unified validation for fallbacks, and tiered routing to cut inference costs [5].
  • Google demonstrates autonomous LLM post-⁠training with Tunix on TPUs, using autonomous research loops (autofinetune) to optimize LoRA ranks, learning rates, and batch sizes for SFT/GRPO, orchestrated with Antigravity CLI and Gemma on Cloud TPUs [6].
  • Google recommends behavioral evaluations (fast, local, unit-⁠style tests on discrete actions) to complement end-⁠to-⁠end agent benchmarks, enabling safer prompt/system iteration without regressions [7].
  • SWE-⁠2 pushes the Pareto frontier: 50.0% on FrontierCode 1.1 Main1, 64% cheaper than SWE-⁠1.7, matching GPT-⁠5.6 Sol and Fable 5/5.1 at a fraction of their price, and within a few points of GPT-⁠6 Astra at a quarter of the cost [8].
  • GPT-⁠Live-⁠1 launches in the OpenAI API at $0.05/min with full-⁠duplex speech, interruption handling, and 12 voice options, reducing interruptions by 80% vs. turn-⁠based systems [9].
  • Meta’s Muse will introduce Shared Agents, allowing users to create and share customizable agents for workflows like customer support and sales [10].
  • Industry trend: companies increasingly route simpler requests to cheaper open models and invest in model routing to reduce costs, with growing knowledge-⁠sharing on cost-⁠effective AI usage [11].
  • Cognition’s SWE-⁠2 outperforms Grok 4.6 on both score and cost, demonstrating significant efficiency gains in agentic coding [8].
  • Rhoda AI confirms that scaling web-⁠video pre-⁠training improves real robot task performance, with larger models excelling in industrial unpacking and pre-⁠training quality (DINO FD) predicting downstream efficiency [12].
  • Salesforce finds that directly training smaller models on expert agent trajectories can hurt performance after harness optimization, highlighting the need for co-⁠evolution strategies [13].

MLOps, DevOps, and Infrastructure

  • PlanetScale details the lifecycle of a sharded Postgres query, explaining how complex deployments across thousands of servers emulate a single database, with a concrete example across four shards [14].
  • Neki introduces sharded Postgres with a router that plans and fans queries across shard groups while preserving standard wire protocol, drivers, ORMs, and SQL behavior, with online workflows for schema changes and resharding [15].
  • Analysis argues centralized router platforms can deliver better economics than local deployments, with outcomes pricing best suited to the infrastructure layer rather than model or app layers [16].
  • vLLM triples MiniMax M3 inference throughput on AMD’s MI355X (109.1 → 342.4 tokens/s/GPU) via systematic bottleneck chasing, fusing shared experts into grouped GEMMs and reusing sparse-⁠attention block selections [17].
  • Colibri is a pure-⁠C inference engine for Mixture-⁠of-⁠Experts models, treating storage/RAM/VRAM as a single hierarchy to run 744B–2.8T parameter models on consumer hardware without dependencies [18].
  • Hugging Face Kernels now supports Helion, enabling developers to build, autotune, and ship performant, portable kernels via the Hugging Face Hub with first-⁠class autotuning and pre-⁠tuned configs [19].
  • NVIDIA introduces CUDA Rust with two compilation tracks for authoring GPU compute kernels in Rust, bringing memory safety and modern tooling to GPU-⁠accelerated workloads [20].
  • Google commits $15B to Finland data centers and explores nuclear power for long-⁠term AI infrastructure, signing a 22-⁠year power agreement with Fortum [21].
  • TSMC posts record August revenue (+53% YoY) and plans High-⁠NA EUV adoption for 2030 to meet AI-⁠driven chip demand [22].
  • OpenAI may double-⁠source next-⁠gen AI chips at Samsung alongside TSMC, expanding beyond its current Jalapeño ASIC (designed with Broadcom, built at TSMC) [23].
  • Cloudflare reduces HelloRetryRequests from ~52% to 3.7% and cuts p90 handshake latency by >150ms with Automatic Key Exchange, upgrading hundreds of thousands of domains to post-⁠quantum encryption [24].
  • Cloudflare CASB adds automatic remediation policies, enabling event-⁠driven logic to revoke risky file shares and dispatch custom webhooks without manual intervention [25].

Developer Tools and Practices

  • GitHub Copilot adds VS Code Agents usage metrics (daily active users, session counts, user messages) to enterprise and organization reports, enabling adoption tracking [26] [27].
  • Copilot code review now auto-⁠resolves addressed comments, writes smart commit messages, and uses broader shell tools and agent ensembles for more thorough reviews [28].
  • Shopify moves mobile apps from React Native back to native Swift/Kotlin because coding agents made maintaining two native implementations cheaper; its Helix workflow uses small, tested checkpoints with visual/adversarial review [29].
  • Checkly used Claude Code to rewrite a Node.js service (92M daily messages) in Go, guided by a black-⁠box test harness, achieving zero incidents, 70% fewer pods, and 60% fewer DB sessions [30].
  • Google’s modern-⁠web-⁠guidance skill equips LLMs with an on-⁠device index of current frontend practices, compensating for stale training data in React/Vite apps [31].
  • Drydock reviews npm/PyPI/VS Code artifacts before publication, flagging risky changes (install scripts, network access, binaries) without handling publishing credentials [32].
  • llmfit is a terminal tool that recommends open-⁠source LLMs based on a user’s hardware, inspecting CPU/RAM/GPU to score models on memory fit, speed, and quality [33].
  • OpenCodeReview is an AI-⁠powered code review CLI tool from Alibaba, validated at scale, capable of reading full file contents, searching the codebase, and producing deep reviews [34].
  • Python 3.15 soft-⁠deprecates re.match() in favor of the clearer re.prefixmatch(), re.search(), or re.fullmatch() [35].
  • Wrapture is a new monkey-⁠patching library for Python, serving testing and observability (e.g., New Relic-⁠style tracing) simultaneously [36].
Sources
  1. Introducing the Agents API openai.com
  2. Introducing the Agents API openai.com
  3. Rapidly scaling online storage to serve over 1 billion ChatGPT users openai.com
  4. Announcing ADK for Kotlin 1.0: Building Production-⁠Ready AI Agents in Kotlin, Android, and Beyond google ai
  5. 4 engineering patterns behind the strongest AI Agents Challenge submissions google ai
  6. Autonomous LLM post-⁠training with Tunix on TPUs google ai
  7. The Anatomy of Harness Engineering: How to Evaluate, Iterate, and Guard AI Coding Agents google ai
  8. Introducing SWE-⁠2: Pushing the Pareto Frontier cognition.com
  9. OpenAI launches GPT-⁠Live-⁠1 for full-⁠duplex voice agents testingcatalog.com
  10. Meta to announce Shared Agents for Muse at Meta Connect testingcatalog.com
  11. Tech companies move to open AI models blog.pragmaticengineer.com
  12. Does Scaling Web-⁠Video Pre-⁠training Help Real Robots Do Real Work? rhoda.ai
  13. Salesforce Finds Better Ways to Co-⁠Evolve Agents and Their Harnesses arxiv 2609.09134
  14. The lifecycle of a sharded Postgres query planetscale.com
  15. Introducing Neki: sharded Postgres planetscale.com
  16. Router Power hypersoren.xyz
  17. vLLM triples MiniMax M3 inference throughput on AMD’s MI355X through systematic bottleneck chasing vllm.ai
  18. Colibri github.com
  19. Helion x 🤗 HF Kernels: Building and Shipping Out-⁠of-⁠the-⁠box Performant Kernels pytorch.org
  20. Introducing CUDA Rust for High-⁠Performance GPU Kernel Development developer.nvidia.com
  21. Google commits $15B to Finland data centers, explores nuclear power for the buildout cnbc.com
  22. TSMC’s August revenue jumps 53% to a record high, plans High-⁠NA EUV adoption for 2030 cnbc.com
  23. OpenAI may double-⁠source next-⁠gen AI chips at Samsung, alongside TSMC tomshardware.com
  24. Automatic Key Exchange: faster, post-⁠quantum secure origin handshakes for 45 billion daily connections (and counting) blog.cloudflare.com
  25. Introducing automatic remediation policies with Cloudflare CASB blog.cloudflare.com
  26. Add VS Code Agents to Copilot usage metrics github.blog
  27. GitHub Copilot weekly releases: September 7 github.blog
  28. Auto-⁠resolution and analysis updates in Copilot code review github.blog
  29. Native is now the future of mobile at Shopify shopify.engineering
  30. Rewriting a Node.js service in Go with agents checklyhq.com
  31. Testing Google’s "modern-⁠web-⁠guidance" skill on a real React app alfy.blog
  32. Drydock (Website) drydock.org
  33. llmfit github.com
  34. OpenCodeReview github.com
  35. Soft-⁠deprecating re.match() simonwillison.net
  36. Don’t sleep on wrapture simonwillison.net

Papers from Friday, 11 September 2026

20 papers

Language Models and Training Dynamics

  • MoE models degrade more rapidly under data repetition than dense Transformers, with degradation scaling with total (not active) parameters; dense 80M models tolerate 8× repetition with minimal loss, while MoEs suffer from 4× and underperform dense baselines after 32× [1].
  • Distance generalization (changing inter-⁠token distances at fixed context length) is poorly understood: positional encodings like RoPE/ALiBi show limited gains over NoPE, and performance depends heavily on training data diversity and transfer learning conditions [2].
  • Layerwise interventions reveal that query-⁠routing and target-⁠knowledge dependencies evolve differently during question answering, with a causal window where pair-⁠conditioned request directions grow before interventions alter later knowledge [3].

Multimodal and Specialized Models

  • Edge-⁠deployable VLMs (2–8B) retain taxonomic knowledge but degrade sharply (9.6–26.6 pp) on camera-⁠trap imagery vs. clean photos, suggesting domain gaps stem from image legibility rather than fine-⁠grained discrimination limits; BioCLIP (300M) shows similar trends [4].
  • MindTopo benchmarks topological reasoning (continuity, separation, order, enclosure, knots) across 11k instances, evaluating both direct reasoning and closed-⁠loop planning with MLLMs; results highlight gaps in current models’ topological intuition [5].
  • Nuha-⁠Speech introduces a 1.5M-⁠sample Arabic Speech-⁠QA corpus, supervised fine-⁠tuning recipes, and an evaluation suite to address Arabic underrepresentation in speech-⁠LLMs [6].

Safety, Alignment, and Evaluation

  • SpecGuard repurposes speculative decoding for zero-⁠cost inference-⁠time backdoor detection in LLMs, leveraging verification signals to flag triggered generations without extra computation [7].
  • A multi-⁠signal hallucination detection pipeline (DeBERTa-⁠v3 + MC Dropout + temperature scaling) achieves F1=0.915 and AUROC=0.977 on HaluEval, with MC Dropout boosting accuracy to 93.2%; DPO fine-⁠tuning reduces hallucinations in a Qwen2.5-⁠0.5B generator [8].
  • RAG-⁠Safety-⁠Bench isolates RAG’s safety impact via four controlled conditions (non-⁠RAG, oracle harmful, related-⁠but-⁠no-⁠answer, random safe documents), enabling reliable measurement of unintended side effects in harmful prompts [9].
  • Medical LLM research shows a widening evaluation lag: from 2023–2026, studies’ newest model age grew from 1.33 to 6.08 quarters, with randomized trials evaluating older models (median 4.6 quarters) than other designs [10].

Optimization and Training

  • GPU-⁠CFR compiles game trees into static dataflow (flat arrays, precomputed indices) and uses CUDA Graph Replay to achieve 80× speedups in Counterfactual Regret Minimization, overcoming GPU kernel launch overheads [11].
  • AdamX integrates cosine similarity into gradient descent for adaptive update magnitudes, with variance rectification for smoother early training; empirical results show competitive convergence across benchmarks [12].
  • CoRA-⁠NAS combines static zero-⁠cost proxy consensus (CoRA-⁠Rank) with anchor-⁠residual learning-⁠curve refinement (CoRA-⁠Refine), achieving high Spearman correlations (0.715–0.946) across NAS benchmarks at ~1% of full training cost [13].

Causal and Theoretical Advances

  • A unified framework for distribution shift introduces γ-concept shifts via entropic optimal transport, with estimable bounds and a DataShifts algorithm for quantifying covariate/concept shifts across applications [14].
  • CausalArena standardizes causal discovery evaluation with synthetic, semantic, and formula-⁠grounded SCMs, addressing inconsistencies in prior benchmarks and isolating pretraining-⁠test overlap effects [15].

Applications and Domain-⁠Specific Insights

  • TART is a modular audio-⁠to-⁠tablature pipeline (MIDI transcription → technique classification → string-⁠fret assignment → tablature generation) that achieves +6.67% audio-⁠to-⁠MIDI F50 and +8.5% string-⁠fret Tab F1 over baselines in zero-⁠shot settings [16].
  • AssayBench-⁠Loop and AssayLoop introduce large-⁠scale adaptive hit discovery for CRISPR screens, combining amortized acquisition policies (AssayFormer) with LLM-⁠derived priors for efficient sequential experimental design [17].
  • Target leakage (not model class) explains high AUROC (≈0.89) in cardiovascular screening models; removing two post-⁠diagnostic features collapses performance across all tested models (linear to tabular foundation) into a 0.0045-⁠wide band [18].

Efficiency and Deployment

  • 3D Point Splatting (3DPS) is the first differentiable radar renderer for mmWave novel view synthesis, combining ITU-⁠R P.2040 material models with closed-⁠form phasor splatting for physically faithful, complex-⁠valued outputs [19].
  • Component-⁠Aware DP for federated speech-⁠LLMs addresses cross-⁠component budget collapse via α-⁠split, a two-⁠pool allocation normalizing encoder/LLM parameters independently to preserve gradient fidelity [20].
Sources
  1. Data Scarcity and Model Sparsity: Mixtures-⁠of-⁠Experts Overfit More to Repeated Data arxiv cs.CL
  2. Distance generalization in transformers: why bother with positional encoding? arxiv cs.CL
  3. From Parameters to Answers: How LLMs Retrieve and Use Their Internal Knowledge arxiv cs.AI
  4. Can Edge-⁠Deployable Vision-⁠Language Models Identify Species? arxiv cs.AI
  5. MindTopo: Can Foundation Models Reason in Topological Space? arxiv cs.CL
  6. Nuha-⁠Speech: Building General-⁠Purpose Arabic Speech-⁠LLMs arxiv cs.CL
  7. SpecGuard: Inference-⁠Time Backdoor Detection For Free arxiv cs.CL
  8. Domain-⁠Specific Hallucination Detection in Large Language Models arxiv cs.CL
  9. RAG-⁠Safety-⁠Bench: Reliable Evaluation of Retrieval-⁠Augmented LLM Safety arxiv cs.CL
  10. The widening evaluation gap in medical large language model research 2023 to 2026 arxiv cs.CL
  11. GPU-⁠CFR: 80x Faster Counterfactual Regret Minimization by Compiling the Game to Static Dataflow and CUDA Graph Replay arxiv cs.AI
  12. AdamX: Cosine similarity meets gradient descent arxiv cs.LG
  13. CoRA-⁠NAS: Coarse Ranking and Anchor-⁠Residual Refinement for Neural Architecture Search arxiv cs.LG
  14. General Quantification of Covariate and Concept Shifts arxiv cs.LG
  15. CausalArena: Benchmarking Causal Discovery in the Foundation Model Era arxiv cs.LG
  16. TART: A Modular Tool for Technique-⁠Aware Audio-⁠to-⁠Tablature Guitar Transcription arxiv cs.LG
  17. Biology-⁠in-⁠the-⁠loop: Amortized Adaptive Hit Discovery in CRISPR Screens arxiv cs.CL
  18. Target leakage, not model class, explains reported accuracy in survey-⁠based cardiovascular screening: a leakage-⁠tiered audit of glass-⁠box and tabular foundation models arxiv cs.CL
  19. 3D Point Splatting for mmWave Radar Novel View Synthesis arxiv cs.LG
  20. Component-⁠Aware Differential Privacy for Federated Multilingual Speech-⁠LLMs arxiv cs.CL