Friday, May 29, 2026

11 articles — 🟡 4 important , 🟢 7 interesting

🤖 Models (1)

🟢 🤖 Models May 29, 2026 · 2 min read

Google: Nine Demo Videos Showcase Gemini Omni and 3.5 Capabilities

Google released nine demo videos showcasing the capabilities of the Gemini Omni and Gemini 3.5 models, unveiled at the I/O 2026 conference. The demos cover conversational video editing, agentic task execution, generative user interface in search, and the personal AI agent Gemini Spark.

📦 Open Source (1)

🟡 📦 Open Source May 29, 2026 · 2 min read

arXiv:2605.31463: PithTrain, a Compact MoE Training System Tailored to Agents

PithTrain is a new training system for Mixture-of-Experts models optimized for AI coding agents. The paper introduces the agent-task efficiency metric and the accompanying ATE-Bench, on which PithTrain achieves up to 62 percent fewer agent steps and 64 percent less active GPU time at comparable throughput.

⚖️ Regulation (1)

🟢 ⚖️ Regulation May 29, 2026 · 2 min read

OpenAI: Guidelines for Trustworthy Third-Party Evaluations of AI Models

OpenAI published a shared playbook for external evaluations of AI models. The document describes how independent evaluators can reliably measure model capabilities, test safeguards and confirm the validity of results for advanced frontier systems.

🤝 Agents (3)

🟡 🤝 Agents May 29, 2026 · 2 min read

Anthropic: Claude Managed Agents Arrive on Claude Platform on AWS

Anthropic announced that key Claude Managed Agents features — webhooks, multi-agent orchestration and self-hosted sandboxes — are now available on Claude Platform on AWS, along with new IAM actions and the managed AnthropicSelfHostedEnvironmentAccess policy for authorizing access.

🟡 🤝 Agents May 29, 2026 · 2 min read

arXiv:2605.31584: LongTraceRL Learns Long-Context Reasoning from Search-Agent Traces

LongTraceRL is a new reinforcement learning approach for long-context reasoning. It builds training data from search-agent traces with tiered distractors and uses rubric rewards with entity-level process supervision, achieving consistent improvements across five benchmarks for models from 4 to 30 billion parameters.

🟢 🤝 Agents May 29, 2026 · 2 min read

Anthropic: Claude Code 2.1.157 Brings Plugin Auto-Loading and an Opus 4.8 Hotfix

Anthropic released Claude Code versions 2.1.157 and 2.1.156. The newer version brings plugin auto-loading from .claude/skills without a marketplace, scaffolding of new plugins, agent override and mid-session worktree switching, while 2.1.156 is a hotfix for API errors on Opus 4.8 caused by modified thinking blocks.

🔧 Hardware (1)

🟢 🔧 Hardware May 29, 2026 · 2 min read

AMD: Speculative Decoding on MI300X Speeds Up Inference by 4.3x

On the ROCm blog, AMD presented an implementation of speculative decoding on Instinct MI300X GPUs. The technique, in which a smaller draft model predicts tokens and a larger one verifies them, achieved up to 4.3 times higher throughput compared to classic autoregressive generation.

🏥 In Practice (1)

🟢 🏥 In Practice May 29, 2026 · 2 min read

OpenAI: Braintrust Speeds Up Experiments with Codex and GPT-5.5

OpenAI published a customer story about the company Braintrust. Its engineers use Codex powered by the GPT-5.5 model to launch experiments and write code faster, turning user requirements directly into functional software solutions.

💬 Community (1)

🟢 💬 Community May 29, 2026 · 2 min read

Google: Waterloo Students Build AI Tutors in the Futures Lab

Google showcased the work of University of Waterloo students who develop AI prototypes for education and work as part of the Futures Lab. The projects include a tutor for American Sign Language, an app for learning Japanese, and a tool for tracking exercise with a camera.

🛡️ Security (1)

🟡 🛡️ Security May 29, 2026 · 2 min read

OpenAI: Expanding Trusted Access to the GPT-Rosalind Model for Biosecurity

OpenAI announced an expansion of trusted access to the GPT-Rosalind model for vetted developers and U.S. government partners, framing the step in the context of biosecurity and strengthening societal resilience against biological threats, with controlled access intended for defensive and research applications.

✨ Curiosities (1)

🟢 ✨ Curiosities May 29, 2026 · 2 min read

arXiv:2605.30963: AMix-2 Introduces Proteins as a Native Modality in LLMs

AMix-2 is a protein-text foundation model that unifies protein understanding and sequence design in a shared token space. It uses a block-wise diffusion language backbone, introduces the ProteinArena benchmark, and outperforms frontier LLMs while competing with specialized protein models.