Thursday, May 14, 2026

15 articles — 🟡 10 important , 🟢 5 interesting

🤖 Models (3)

🟡 🤖 Models May 14, 2026 · 2 min read

arXiv:2605.13301 SU-01: 30B model reaches gold-medal level at IMO 2025, USAMO 2026, and IPhO through three-phase training

SU-01 is a new reasoning training methodology published on May 14, 2026 on arXiv (Yafu Li and 27 co-authors, corresponding author Runzhe Zhan). A 30B parameter A3B backbone reaches gold-medal performance on the International Mathematical Olympiad 2025, USAMO 2026, and International Physics Olympiad 2024-2025 through three sequential phases: reverse-perplexity curriculum SFT on 340K trajectories, two-stage RL, and test-time scaling. Reasoning chains reach 100K+ tokens.

🟢 🤖 Models May 14, 2026 · 2 min read

Allen Institute: AIMIP benchmark — AI climate models 2× better on historical data but fail to generalize to long-term warming

AIMIP (AI Model Intercomparison Project) is a new community benchmark for AI weather and climate models published on May 13, 2026 by the Allen Institute together with NVIDIA, Google Research, University of Washington, University of Maryland and the ArchesWeather group. Phase 1 evaluation of eight AI model simulations showed a twofold reduction in error on historical data — but also a serious inability to generalize to long-term warming trends.

🟢 🤖 Models May 14, 2026 · 2 min read

Microsoft Research GridSFM: foundation model solves AC optimal power flow 100× faster than DC approximation

GridSFM is a new Microsoft Research small foundation model for electric power grids published on May 13, 2026. It approximates AC optimal power flow in milliseconds on grids of 500 to 80,000 nodes — 100× faster than DC approximation and 1,000× faster than full AC solvers. Median cost gap is 2.23%, feasibility detection achieves 94.5%/96.1%, and the model projects potential savings of $20 billion annually in congestion costs.

📦 Open Source (2)

🟡 📦 Open Source May 14, 2026 · 2 min read

LangChain: LangSmith Engine automates agent debugging — error clustering, root cause, PR and evaluator recommendations

LangSmith Engine is a new LangChain platform published on May 13, 2026, that automates the production feedback loop for AI agents. The engine analyzes trace data from production runs, clusters errors by pattern, diagnoses root cause in code, and generates concrete pull request suggestions plus evaluator definitions. The goal is a shift from manual debugging toward continuous automated improvement.

🟡 📦 Open Source May 14, 2026 · 2 min read

PyTorch: Version 2.12 brings device-agnostic torch.accelerator.Graph, MX quantization and 100× faster linalg.eigh

PyTorch 2.12 is the new production release of the PyTorch framework published on May 13, 2026, with 2,926 commits and 457 contributors. Key features: torch.accelerator.Graph device-agnostic API for CUDA, XPU and out-of-tree backends, torch.export support for Microscaling MX quantization (MXFP4/6/8), linalg.eigh up to 100× faster on CUDA via cuSolver, and torch.cond inside CUDA Graphs. TorchScript has been formally removed.

🤝 Agents (4)

🟡 🤝 Agents May 14, 2026 · 2 min read

Amazon Nova Sonic + WebRTC: real-time voice agents with Kinesis Video Streams and async tool calling for RAG/MCP

Amazon Nova Sonic + WebRTC integration is a new AWS architecture published on May 13, 2026 for real-time voice agent applications. A speech-to-speech event processor orchestrates media and text data events through Kinesis Video Streams WebRTC signaling, while server-side VAD reduces audio tokens. Nova Sonic supports async tool calling to MCP servers, Strands agents and RAG systems — IoT and connected vehicle scenarios are the first demonstrations.

🟡 🤝 Agents May 14, 2026 · 2 min read

Anthropic: Claude Code v2.1.141 adds terminalSequence hook, Bedrock Haiku fix, and Summarize up to here rewind option

Claude Code v2.1.141 is the new Anthropic CLI agent release published on May 13, 2026. The third patch version this week adds a terminalSequence field for hook JSON output, the CLAUDE_CODE_PLUGIN_PREFER_HTTPS and ANTHROPIC_WORKSPACE_ID environment variables, claude agents --cwd path scoping, and a new Rewind menu option Summarize up to here for compressing old context. It fixes a Bedrock/Vertex Haiku model ID race and daemon status on Windows.

🟡 🤝 Agents May 14, 2026 · 2 min read

LangChain: Managed Deep Agents — hosted runtime in LangSmith with durable execution and memory layer

Managed Deep Agents is a new LangChain hosted agent runtime published on May 13, 2026 in private beta within the LangSmith platform. The service provides durable execution, persistent memory, integrated tooling and comprehensive observability — all the infrastructure components needed for production deep agents. The agent definition stays in the repository through standard AGENTS.md and tools.json files.

🟡 🤝 Agents May 14, 2026 · 2 min read

OpenAI: Codex sandbox for Windows introduces controlled filesystem and network restrictions for autonomous agents

Codex Windows Sandbox is a new OpenAI security architecture published on May 13, 2026, enabling the Codex agent to execute safely on the Windows operating system. The sandbox introduces controlled filesystem access and network restrictions to enable safe, efficient coding agents — Codex becomes a cross-platform tool available to Windows users, not just macOS/Linux developers.

🏥 In Practice (4)

🟡 🏥 In Practice May 14, 2026 · 2 min read

Perplexity: finance_search Agent API tool returns OHLCV, balance sheets, transcripts, and analyst estimates in a single call

Perplexity finance_search is a new Agent API tool released in May 2026 that returns structured financial data for public companies — near-real-time prices, OHLCV ranges, pre-market and after-hours data, income statements, balance sheets, cash flow, earnings call transcripts, SEC filings, analyst estimates, and ETF constituents. The model decides which fields to fetch based on the prompt.

🟢 🏥 In Practice May 14, 2026 · 2 min read

AWS: Pulse AI + Bedrock pipeline fine-tunes Nova Micro for financial documents — 1,000 documents in 3 hours

AWS Pulse AI + Amazon Bedrock financial document processing is a new enterprise pipeline blueprint published on May 13, 2026. It combines Pulse AI extraction from complex financial documents (SEC filings, balance sheets, audit materials) with fine-tuning of the Amazon Nova Micro model. A batch of 1,000 documents that previously required a multi-day turnaround was processed in under three hours, and the fine-tuned model achieved 100% check data extraction versus a 50% baseline.

🟢 🏥 In Practice May 14, 2026 · 3 min read

CNCF: KubeStellar AI agents achieve 81% PR acceptance with 91% test coverage and 63 CI/CD workflows

KubeStellar AI Agents is a new case study CNCF blog post by Andy Anderson, Chief Maintainer of KubeStellar Console, published on May 14, 2026. The multi-cluster Kubernetes dashboard achieved 81% PR acceptance over 82 days using two parallel AI coding agents. Infrastructure: 63 CI/CD workflows, 32 nightly test suites, 91% coverage across 12 shards, bug-to-merge roughly 30 minutes. Anderson defines five levels of AI codebase maturity.

🟢 🏥 In Practice May 14, 2026 · 2 min read

GitHub: Copilot Cloud Agent REST API open for fan-out refactoring, repo setup and weekly release preparation

GitHub Copilot Cloud Agent REST API is a new developer endpoint published on May 13, 2026 in public preview that enables programmatic triggering of autonomous Copilot tasks. Three primary use cases: code refactoring/migration fan-out across many repositories, one-click repository initialization from a developer portal, and automated weekly release preparation with release notes. Available to Copilot Business and Enterprise subscribers.

🛡️ Security (2)

🟡 🛡️ Security May 14, 2026 · 2 min read

arXiv:2605.13825 History Anchors: one instruction raises unsafe decisions in 17 frontier LLMs to 91-98%

History Anchors is a new safety paper published on May 14, 2026 on arXiv by Alberto G. Rodríguez Salgado. It demonstrates that a single instruction — remain consistent with the prior strategy — raises the unsafe outcome rate in aligned LLMs from a near-zero baseline to 91-98%. Tested on 17 frontier models from 6 providers across the HistoryAnchor-100 dataset spanning 10 high-stakes domains. The paper reveals an inverse-scaling pattern: stronger models are more vulnerable.

🟡 🛡️ Security May 14, 2026 · 2 min read

AWS and Cisco: AI Registry scans MCP and A2A agents via YARA, LLM semantic analysis and Cisco proprietary scanners

AWS + Cisco AI Defense integration is a new enterprise security stack for AI agents published on May 13, 2026. The open AI Registry control plane scans MCP servers and A2A agents at registration using YARA pattern analysis, LLM semantic scanning via Amazon Bedrock and Cisco proprietary scanners. Vulnerable servers receive a security-pending tag and remain disabled until an administrator approves a review.