Global AI Weekly - Issue Issue 166

In this issue

Highlights

3 stories

blogs.nvidia.com

NVIDIA Buys Hugging Face for $12.9B

NVIDIA has agreed to acquire Hugging Face for $12.93 billion, bringing one of the AI industry’s most important open model platforms under the world’s dominant AI chipmaker. Hugging Face will remain open, with continued support for multiple clouds, accelerators, frameworks and model providers. NVIDIA says its infrastructure and engineering resources will help scale the platform while preserving the open ecosystem used by more than 18 million developers.

cnbc.com

Anthropic Backs Down on Data Retention

Anthropic is reversing its controversial 30-day data retention requirement for enterprise customers after pushback over privacy and compliance. Its replacement, Enterprise Frontier Safeguards, lets businesses keep activity data in their own cloud infrastructure while still running automated misuse detection. Anthropic says the system was developed with more than 100 customers and will roll out this fall. Eligible customers will temporarily get zero data retention on Claude Fable 5 and 5.1 during the transition.

openai.com

OpenAI’s Astra Crosses the Cyber Red Line

In this issue

Research

2 stories

arxiv.org

Coding Agents That Improve for Days

Researchers from Shanghai AI Laboratory introduce a framework that keeps coding agents improving across repeated planning, coding and testing cycles instead of treating development as a single run. Across several benchmarks, it improved results by an average of 52%. In its most ambitious test, the system ran for more than 70 iterations and autonomously built a complete, playable first-person shooter from a product requirements document.

arxiv.org

Randomly Deleting AI Memory Actually Works

Salesforce researchers challenge the assumption that models need sophisticated algorithms to decide which reasoning tokens to keep in their KV cache. Their surprisingly simple approach preserves the prompt, then randomly evicts other cached tokens. Across four models and six reasoning tasks, it matched stronger selection methods while delivering 32 to 43% higher throughput in vLLM. The findings suggest long reasoning traces contain far more redundancy than expected.

In this issue

Video

1 story

youtube.com

From AI Agent Prototype to Production

Microsoft’s new Inside Microsoft Foundry: Quickstart series takes a practical look at what developers actually need to ship AI agents. It starts with building an agent using Microsoft Foundry and the Microsoft Agent Framework, then moves into prompts, tools, tracing and evaluation. Future episodes will cover models, grounding, VS Code development, observability, governance and production deployment. A useful hands-on series for developers moving beyond agent demos.

In this issue

Articles

2 stories

microsoft.com

Help Shape GitHub Copilot Canvas

Microsoft is gathering feedback on GitHub Copilot App Canvas, an experimental way to make agent-driven software visible and interactive while it runs. Instead of treating Canvas as another UI builder, the idea is to use it as a shared runtime where developers and AI agents can observe state, execute actions, validate results, and evolve systems together. If you are experimenting with agentic development, this is your chance to share feedback with the team.

github.blog

Copilot Can Now Approve Your Pull Requests

GitHub Copilot code review can now do more than leave comments. In public preview, Copilot can assess whether a pull request is ready to approve and, when admins enable it, submit an approval that counts toward repository merge requirements. Controls are available at enterprise, organization and repository level, including path-specific approval rules. If new commits arrive, Copilot’s approval is dismissed just like a human reviewer’s.

In this issue

Upcoming Events

1 story

nvidia.com

October 20–22: NVIDIA GTC Berlin

Europe's biggest AI moment is happening this autumn. GTC Berlin brings together developers, researchers, and industry leaders to go deep on the full five-layer AI stack, from energy, chips, and infrastructure to open models and physical AI. The headline keynote by Jensen Huang takes place at the iconic Tempodrom on 21 October, with three days of sessions, hands-on labs, and networking at STATION-Berlin. For developers shipping agentic workflows or navigating open-source AI, GTC Berlin is where it all comes together. Eligible attendees get 25% off.

In this issue

Code

2 stories

aws.amazon.com

Give Strands Agents Memory That Actually Survives

AWS has released strands-dynamodb-storage, an open source DynamoDB backend for the Strands Agents SDK. It gives agents durable storage for session state, transcripts and long-term memory, with support for semantic search, TTL, compression and S3 offload for larger payloads. The package works with both Python and TypeScript and keeps agent state inside your own AWS account, using a single DynamoDB table.

blog.google

Gemini 3.8 Flash Gets a Cybersecurity Twin

Google has launched Gemini 3.8 Flash alongside a security-focused 3.8 Flash Cyber variant. The standard model targets long-horizon coding, agentic workflows and complex reasoning while keeping Flash pricing. Flash Cyber is built for defenders, with frontier-level vulnerability discovery and automated patching. Google says it already helped find a critical vulnerability in under two hours and produced 2.6x more correct Chrome security patches than larger commercial models.

In this issue

Podcast

1 story

open.spotify.com

Stop Prompting, Start Building Agentic Loops

NLW and Nufar Gaspar explore how to move beyond one-shot prompting and build repeatable agentic workflows. They cover how to define verifiable finish lines, decide which tasks benefit from loops, control runaway token costs, and combine multiple agents into larger work graphs. It is a practical look at what changes when AI stops simply answering questions and starts iterating toward an outcome.