Discussion | Link

Zhipu AI published a technical blog detailing inference anomalies encountered during large-scale GLM-5 Coding Agent serving, including gibberish, repetition, and rare character generation. The team traced the issues to KV Cache race conditions and HiCache loading timing defects, and proposed solutions along with Prefill optimization strategies.

Starting June 1, 2026, GitHub Copilot will shift from a flat-rate subscription to a per-token pricing model. Users will pay based on tokens consumed (input + output), similar to LLM API charges. The base Copilot Pro tier ($10/month) will receive 1,000 AI Credits (1 cent each). Code completions and Next Edit suggestions remain free. This aligns with industry moves by Anthropic and OpenAI but may discourage user exploration. Enterprises must weigh efficiency gains against rising costs.

A day-long conference at the Royal Society explored how AI is transforming scientific discovery across fields like astronomy, materials science, climate forecasting, and nuclear fusion. Speakers highlighted breakthrough applications including foundation models for galaxy classification, diffusion models for crystal design, AI-based sea ice forecasting, and fast plasma simulations for fusion.

Interim CEO Peter Clark discusses Ai2's ongoing commitment to open science amidst rapid AI progress, highlighting key projects, the NSF OMAI initiative, and future directions in AI for science, embodied AI, and environmental AI.

Researchers from Tsinghua AIR and collaborators introduce GS-Playground, a next-generation simulation framework for vision-centric embodied AI. It uniquely combines high-throughput parallel physics simulation with high-fidelity batch 3DGS rendering, enabling large-scale policy training and zero-shot sim-to-real transfer. The framework is open-source and accepted at RSS 2026.

This week's AI news highlights OpenAI's Codex expanding to knowledge work and Anthropic's Claude launching security tools and creative software support. GPT-5.5 shows strong cyber evaluation results, multiple open-weight models are released, and agent infrastructure and security issues are in focus.

MIT senior Olivia Honeycutt investigates how the ways we communicate can shape our views of the world.

Beacon Biosignals uses a lightweight EEG headband to monitor brain activity during sleep at home, enabling AI analysis for diagnosing and treating neurological disorders.

Eloo is a no-code AI agent builder for non-technical teams, enabling them to create AI workforces by defining business objectives. It has been adopted by 100+ teams, with 500+ agents created.

Apple researchers propose an inference-time evaluation method that introduces a reviewer agent to assess provisional tool calls before execution, enabling real-time error correction. Evaluated on BFCL and τ2-Bench, the method achieves +5.5% and +7.1% improvements, and introduces Helpfulness-Harmfulness metrics to quantify the tradeoff of corrections.

An AI assistant whose memory belongs to you.

Benchmark is a free tool by Cumbuca that lets users upload fintech proposal PDFs to analyze pricing. Using AI, it extracts line items and compares them to a database of over 150 products and 660+ data points from real market proposals, providing a color-coded report (green, yellow, red) indicating fairness. It requires no login and aims to address opaque pricing in Brazilian financial infrastructure.

Access your bank data in ChatGPT & Claude via Open Finance.

Microsoft Research red-teamed a live platform of over 100 AI agents, identifying network-level risks that only appear through agent interactions, including self-propagating worms, reputation manipulation, manufactured consensus, and proxy chains. These risks cannot be reproduced by testing agents in isolation. The study also observed emergent security behaviors in a small fraction of agents, reducing attack success. Findings suggest the need for layered defenses across platform, agent, and model layers.

SurfBuddy is a crew of AI specialists in your browser sidebar that automate cross-app workflows. It supports Gmail, Sheets, Docs, Calendar, and the web. Describe tasks in plain language, and the AI executes multi-step actions in real time. Privacy-focused, API-only, with a free tier and Pro at $29/month.

By early 2026, the open source project OpenClaw had become a phenomenon. In January, its GitHub star count crossed 100,000 as developer interest surged.

Evaluate and iterate on LLM applications with confidence using LangSmith's regression testing. Compare experiments, track performance, and identify changes.

LangSmith, the unified DevOps platform for LLM applications, is now available as a transactable offering in the Azure Marketplace, enabling deployment within Azure VPC with full data control and MACC credit support.

Dosu uses evaluation driven development and LangSmith to build reliable LLM products at scale, monitor production performance, and iterate with confidence.

Compare LLM test runs side-by-side with LangSmith's Test Run Comparisons. Manually inspect data, filter results, and gain insights faster.

Multi-agent systems that mirror real engineering teams — not just code faster — can cut debug time by 93% and compress cross-team delivery. Here's the architecture built on LangGraph.

Production monitoring for LLM agents requires new observability tools. Learn how to trace, evaluate, and improve AI agents at scale.

Fere AI introduces AI agents that convert market signals into executable trades on cryptocurrency exchanges and Polymarket.

OpenAI and Microsoft restructured their partnership, ending Azure exclusivity and allowing OpenAI to bring models and tools to AWS Bedrock. This move reflects both companies' need for flexibility, while AWS emerges as a potential winner by securing major silicon commitments from both OpenAI and Anthropic within eight days.

On Thursday, Anthropic released Claude Security from closed preview, now available in beta for Claude Enterprise customers. The tool scans entire codebases for vulnerabilities using parallel agents, suggests fixes, and validates findings to reduce false positives. New features include scheduled scans, dismissal with comments, and CSV/Markdown exports.

Learn the difference between CLI interactive and non-interactive modes. Interactive mode offers a chat-like experience for deep collaboration, while non-interactive mode provides quick one-off answers.

LG and NVIDIA are in exploratory discussions about physical AI, data centers, and mobility. LG aims to supply cooling solutions for NVIDIA's high-density compute clusters, while its home robot CLOiD relies on NVIDIA's Omniverse and Isaac platform for real-time inference. The collaboration could also unify automotive infotainment and autonomous driving technology.

Anthropic's Cat Wu discusses how Claude Code has transformed software development: 90% of code is now AI-written, the bottleneck has shifted to code review, and product taste emerges as the new key skill.

Meta has decided to shift focus from its open-source Llama models to a new proprietary AI model, Muse Spark, raising concerns among developers who built on Llama. The move leaves no clear migration path due to fundamentally different deployment models. Developers can continue using existing Llama models, switch to other open-source alternatives, or adopt proprietary APIs, but all options involve significant costs. Llama forks may offer some relief.