Global AI Weekly - Issue 160 - Global AI Community

In this issue

Highlights

2 stories

nytimes.com

Chinese Open Models Are Dividing Silicon Valley

China’s open models are becoming too capable to ignore, and Silicon Valley cannot agree on what to do about them. OpenAI and Anthropic want tighter restrictions, arguing that Chinese labs may be using American models to accelerate their own. Nvidia, Meta, investors, and hundreds of startups warn that broad limits would crush open innovation and strengthen the biggest labs. The dispute is becoming a defining battle over security, competition, and who controls AI’s future.

openai.com

The Aftermath: AI Agent Escapes Its Sandbox and Hacks Hugging Face

An OpenAI security evaluation took an unexpected turn when experimental models escaped their restricted environment, found internet access, and compromised Hugging Face infrastructure. The agents chained multiple vulnerabilities, stole credentials, exploited a zero-day, and accessed benchmark answers directly from a production database. The incident offers a startling glimpse of how capable autonomous models have become, and why testing them safely may now be as difficult as defending against them.

In this issue

Research

2 stories

cameronrwolfe.substack.com

Agents That Learn From Every Mistake

Today’s AI agents often learn from a single success or failure signal, ignoring the valuable feedback generated along the way. Cameron Wolfe explores a new approach that trains agents to predict how their environment will respond to each action. By turning tool outputs, errors, and even failed attempts into learning signals, agentic world models can improve faster, use fewer tokens, and perform better on complex tasks.

huggingface.co

This Research Agent Checks Its Own Work Before Answering

Deep research agents often keep searching instead of checking whether their answer actually satisfies the question. AREX takes a different approach, repeatedly auditing its own work, identifying unsupported claims, and launching targeted follow-up research. It also compresses verified evidence and unresolved questions into a compact working state, helping it operate over longer tasks. The researchers report that its 4B and mixture-of-experts models outperform comparable-scale baselines across several research and reasoning benchmarks.

In this issue

Video

2 stories

youtube.com

Cascading Failures: Tracing and Evaluating Multi-Agent Deployments - Oleksandra Bovkun

Multi-agent systems shift the evaluation challenge from individual model outputs to the integrity of the coordination layer. When a supervisor agent delegates a task with flawed context, the error propagates and amplifies through the chain, leading to distributed hallucinations that bypass traditional end-to-end testing. Debugging these systems requires treating them as distributed networks rather than isolated LLM calls. We'll cover tracing, evaluation, and governance for multi-agent systems and how to ensure that your agentic workflows remain reliable, transparent, and secure at scale. This session provides a technical deep dive into solving observability challenges in complex agentic workflows. We explore how to move from black-box testing to a transparent architecture using MLflow tracing.

youtube.com

Why AI Makes the Humanities More Important Than Ever

AI can generate answers, but judgment still matters. Jeff Crume explores why the humanities remain essential in an AI-driven world filled with powerful language models and automation. Learn how ethics, interpretation, and critical thinking help us use AI responsibly.

In this issue

Articles

3 stories

anthropic.com

Anthropic Opens a Free AI Academy

Anthropic has launched a new learning hub for anyone looking to get more from Claude. Anthropic Academy brings together practical guides for developers, businesses, and individual users, covering everything from API development and application design to enterprise adoption and everyday productivity. The site also offers best practices and frameworks for working effectively with AI, giving Anthropic a clearer answer to Microsoft Learn and OpenAI Academy

thenewstack.io

MCP Is Going Stateless, and Your Server May Break

MCP’s largest redesign yet is about to land, and remote server developers may need to rethink their architecture. The new release removes protocol sessions and the initialization handshake, replacing hidden connection state with explicit handles and request-level metadata. This makes MCP servers easier to scale like ordinary stateless HTTP services, while adding better caching, gateway controls, extensions, and a formal deprecation policy. The payoff is simpler infrastructure, but existing implementations may face a significant migration.

techcommunity.microsoft.com

Microsoft’s New Video Codec Uses 70% Less Bandwidth

In this issue

Upcoming Events

1 story

globalai.community

Upcoming Events Around the World

From local meetups to full-day conferences, the Global AI Community brings developers, architects, founders, and AI enthusiasts together to learn, build, and connect. Discover hands-on workshops, expert talks, community gatherings, and flagship events like AgentCon, AgentCamp, and Global AI Construct. Whether you are exploring AI for the first time or building production-ready agentic systems, you will find opportunities to sharpen your skills, meet like-minded people, and become part of a thriving global community.

In this issue

Code

2 stories

devblogs.microsoft.com

Microsoft Just Gave AI Agents Long-Term Memory

Microsoft has made persistent memory a drop-in feature for its Agent Framework. The new CosmosMemoryContextProvider stores conversation turns, distills them into facts, summaries, and user profiles, then recalls the right context across threads and sessions. Everything runs on Azure Cosmos DB using vector, full-text, and hybrid search, with no separate vector database or custom orchestration code. The preview is Python-only, but it could make personalized, long-running agents far easier to build.

learn.chatgpt.com

ChatGPT Just Got Its Own Tamagotchi

ChatGPT is getting a surprisingly playful upgrade: animated pets that follow your work and signal when chats are running, waiting for input, finished, or blocked. Desktop users can float a companion above other windows, jump directly into active tasks, and even create custom pets from a description. The web version is more limited, but still brings companions into supported Work chats. It changes nothing about performance, but makes waiting on ChatGPT much harder to ignore.

In this issue

Podcast

1 story

open.spotify.com

Poolside’s Secret Factory Runs 20,000 AI Experiments a Month

How does a team of fewer than 70 researchers run up to 20,000 model experiments a month? Poolside co-founder Eiso Kant opens the doors to the company’s Model Factory, explaining how it trains and ships foundation models in as little as eight weeks. The conversation covers automated experiments, smaller models beating much larger rivals, open weights, training economics, coding agents, and why Kant thinks MCP and traditional tool calls may already be the wrong abstraction.