Wednesday, May 27, 2026

15 articles — 🟡 10 important , 🟢 5 interesting

🤖 Models (3)

🟡 🤖 Models May 27, 2026 · 2 min read

arXiv:2605.29157: Parallax local linear attention accelerates decode phase 12.9× compared to FlashAttention

Parallax is a new attention mechanism for large language models that replaces standard softmax attention with local linear estimation, achieving a 12.9× speedup of the decode kernel compared to FlashAttention. Researchers from Northwestern University and collaborators demonstrated consistent perplexity improvements during pretraining of 0.6B and 1.7B parameter models, claiming the first empirical demonstration of strong architecture-optimizer co-design for attention mechanisms.

🟡 🤖 Models May 27, 2026 · 2 min read

Google: gemini-3.1-flash-image and gemini-3-pro-image reach GA — with video-to-image generation support

Google announced General Availability (GA) of two native visual models from the Gemini API on May 28, 2026: gemini-3.1-flash-image for efficient image generation and gemini-3-pro-image for high performance. A new feature is the video-to-image capability in the flash variant, which generates thumbnails, movie posters, and infographics from video files or YouTube URLs. Preview versions of both models are being deprecated on June 25, 2026.

🟡 🤖 Models May 27, 2026 · 3 min read

Mistral: New class of AI models for physical system prediction — foundation for engineering acceleration

On May 27, 2026, Mistral AI announced Physics AI — a new category of AI models specialized in predicting the behavior of physical systems. The goal is to accelerate hardware engineering and complex technical processes by providing AI tools that understand physical laws, not just statistical patterns from text.

📦 Open Source (2)

🟡 📦 Open Source May 27, 2026 · 3 min read

PyTorch: How torch.compile achieves up to 10× speedup through kernel fusion

The PyTorch compiler torch.compile achieves up to 10× speedup on GPUs by eliminating redundant writes to global memory through kernel fusion. Instead of separate GPU kernels for each operation, the Inductor compiler automatically merges dependent operations into a single kernel that keeps intermediate values in the processor's fast registers.

🟢 📦 Open Source May 27, 2026 · 2 min read

xAI: SDK Python v1.14.0 adds enable_image_search to web_search and SERVER_SIDE_TOOL_IMAGE_SEARCH enum

xAI released Python SDK v1.14.0 which brings support for image search within the web_search functionality. The new enable_image_search field and SERVER_SIDE_TOOL_IMAGE_SEARCH enum constant allow development teams building Grok-based applications to enable server-side visual content search alongside standard web text search.

🤝 Agents (2)

🟡 🤝 Agents May 27, 2026 · 3 min read

AWS: Works Human Intelligence migrates from LangGraph to Amazon Bedrock AgentCore and achieves 97% cost reduction per process

Amazon Bedrock AgentCore, a managed platform for building and running AI agents, is showcased in a production case study from Japanese HR provider Works Human Intelligence. The company migrated two business agents from LangGraph and Amazon ECS infrastructure to AgentCore, achieving a 97% cost reduction for the Browser Operation agent through prompt caching, behavior optimization, and switching models from Claude Sonnet 4.5 to Haiku 4.5.

🟢 🤝 Agents May 27, 2026 · 2 min read

OpenAI: Thrive and Crete build self-improving tax agents on Codex

OpenAI published a case study about a collaboration with Thrive and Crete on building tax AI agents based on the Codex model. The project demonstrates the application of self-improving agents in a high-risk regulatory domain — tax law — where accuracy and up-to-date knowledge have direct financial and legal consequences for users.

🔧 Hardware (2)

🟢 🔧 Hardware May 27, 2026 · 3 min read

AMD ROCm: 4-Wave Interleave Optimization of FP8 GEMM Kernels for Instinct MI355X (CDNA 4) Doubles Register Budget per Wave

The AMD ROCm blog describes a new 4-wave interleave approach for FP8 GEMM (mixed-precision matrix multiplication) kernels on the Instinct MI355X accelerator with CDNA 4 architecture. The optimization shifts from an 8-wave ping-pong layout to 4 waves with a full VGPR budget of 512 registers, eliminating LDS memory bank conflicts through XOR-based swizzle and hiding memory latency by precisely overlapping MFMA instructions with data loads.

🟢 🔧 Hardware May 27, 2026 · 3 min read

PyTorch: TokenSpeed achieves record 580 tokens/s on Qwen3.5-397B-A17B with NVIDIA Blackwell GPUs

TokenSpeed, an open-source LLM inference engine from LightSeek Foundation with an MIT license, achieved 580 tokens per second on the Qwen3.5-397B-A17B model using eight NVIDIA Blackwell B200 GPUs in a TP8 configuration. The record was achieved for agentic workloads with prefix caching and CPU-GPU execution overlap, with less than 16% performance degradation even at a 1 million token context.

🏥 In Practice (2)

🟡 🏥 In Practice May 27, 2026 · 3 min read

Anthropic: Messages API Returns usage.output_tokens_details.thinking_tokens — What It Means for Billing and Streaming

The Anthropic Messages API now includes the output_tokens_details.thinking_tokens field in the usage object of every response that uses Extended Thinking. Developers measuring inference costs need to know that all generated thinking tokens are billed — not the summarized view returned in the response. The difference can be 20x or more.

🟡 🏥 In Practice May 27, 2026 · 2 min read

Anthropic: Claude Code v2.1.152 Brings /code-review --fix, /simplify and disallowed-tools in Skills Frontmatter

Claude Code v2.1.152 introduces three key upgrades for development teams: the /code-review --fix command automatically applies review findings to the working tree, /simplify invokes the same command with a focus on code cleanup, and the skills frontmatter now supports a disallowed-tools field that removes specified tools while a skill is active. The release also brings new hooks, improved Vim mode, and stability fixes.

🛡️ Security (4)

🟡 🛡️ Security May 27, 2026 · 3 min read

arXiv:2605.28588: Analysis of 3,984 Agent Skills Reveals 76 Malicious — 13.4% Contain Critical Security Vulnerability

A security analysis of 3,984 AI agent skills from major marketplaces discovered 76 confirmed malicious payloads including credential theft, backdoor installation, and data exfiltration. 13.4% of all analyzed skills contain at least one critical security vulnerability, and at least 8 manually verified malicious skills remained publicly available on the main platform at the time of publication.

🟡 🛡️ Security May 27, 2026 · 3 min read

arXiv:2605.28914: AIRGuard Reduces Prompt Injection Attack Success from 36.3% to 5.5% with Runtime Agent Authority Control

AIRGuard is a runtime security layer for tool-equipped language agents that addresses the authority confusion problem — a vulnerability where unauthorized contextual inputs can exploit legitimate agent actions (file access, API calls) for attack purposes. On the AgentTrap benchmark, AIRGuard reduces the attack success rate against Claude Sonnet 4.6 from 36.3% to 5.5%, while retaining 76% of useful functionality on the DTAP-150 benchmark.

🟡 🛡️ Security May 27, 2026 · 2 min read

arXiv:2605.29068: COLAGUARD transfers safety reasoning to latent space — +8.24 F1, 22.4× fewer tokens

COLAGUARD is a new safety guardrail system for large language models that transfers safety reasoning from explicit textual chain-of-thought into a continuous latent space, using curriculum-based training. The system achieves an improvement of 8.24 macro-F1 points over Llama Guard 3, with 22.4× fewer generated tokens and 12.9× faster inference than the GuardReasoner baseline across eight safety datasets.

🟢 🛡️ Security May 27, 2026 · 3 min read

Google Research: Private AI model analytics without accessing user data via Zero-Trust aggregation and TEE

Google Research has presented a private analytics system for AI models that combines a lattice-based cryptographic protocol with a Trusted Execution Environment (TEE) to collect aggregate statistics on on-device model performance without ever accessing individual user data. The system has been deployed in the Android System SafetyCore service on Android 9+ and enables single-message sending without devices needing to remain online across multiple rounds.