Global AI Weekly - Issue 162 - Global AI Community

In this issue

Highlights

2 stories

techcrunch.com

OpenAI’s Green Card Hiring Is Now Under DOJ Watch

OpenAI and Statsig have agreed to pay $3.2 million and accept three years of Justice Department oversight over hiring tied to green card sponsorships. The DOJ alleged the companies made it harder for U.S. workers to apply for certain roles while sponsoring foreign employees for permanent residency. Neither company admitted wrongdoing. The settlement covers fewer than ten positions, but adds new reporting requirements and government scrutiny to OpenAI’s hiring practices.

reuters.com

Google Just Rewired the Top of Its AI Empire

Google is reshuffling the leadership of its AI operation. DeepMind co-founder Demis Hassabis is stepping away from day-to-day management to become Chair of Google DeepMind and Chief Scientist of Alphabet, focusing more on AGI and scientific research. CTO Koray Kavukcuoglu will take operational control, while longtime Google AI heavyweight Jeff Dean is leaving after 27 years to launch a new AI venture. It is a major changing of the guard inside one of AI’s biggest players.

In this issue

Research

2 stories

arxiv.org

AI Agents Have Skills. They’re Just Bad at Using Them.

Giving an agent a library of reusable skills sounds great, but can it actually recognize when to use them and follow the instructions correctly? Skill-Use tests eight models across 177 executable tasks using 79 real skills. The strongest setup scored just 0.613. The researchers found separate failures in triggering the right skill and following its procedure, with performance also changing significantly depending on the agent harness around the model.

arxiv.org

Can AI Agents Actually Get Better With Experience?

Persistent agents promise to learn from previous tasks, but does keeping memories really make them better? PAST-Bench tests seven models and four agent frameworks across sequences of tasks involving memory, procedural reuse, information gathering, and updates. Agents did improve when retaining experience, but the gains were inconsistent and did not always come from genuine reuse of what they had learned. The research highlights how far persistent agents still are from reliable self-improvement.

In this issue

Video

2 stories

youtube.com

LLM Quantization Explained

LLM quantization is how a 70B model that needs 140GB of memory gets small enough to run on a normal GPU. Every model you chat with is really a giant file of numbers, and those numbers have to fit in memory somewhere. Quantization stores each number in fewer bits so the file shrinks. In this video, KodeCloud untangles the three things people constantly mix up: number formats, quantization methods, and file formats. Then shows why dropping precision is never free.

youtube.com

Lessons from a No-Code Library

The allure of Spec-Driven Development (SDD) is that you bring the specs and tests and the code comes for free. In reality, SDD presents a new challenge: keeping the specs, tests, and code in sync as you move forward. This talk distills lessons from shipping a no-code library and studying the first wave of SDD projects, which largely leaned on existing software as sources of truth. That approach shines in a specific shape of problem: tightly-defined surfaces like APIs, plugin systems, and conformance suites, where behavior is testable and the spec can credibly stand in for the code. It's also a promising model for open source maintainers who want to seed community contribution: ship the spec and tests, let contributors bring their own agents and intent. But the second wave of SDD is pushing past that genre into territory where no external validator exists, and there the sync problem compounds fast. We've been here before: the 1968 Software Crisis was our inability to manage codebases larger than one engineer's head, and the answer then was process. Agents have triggered a second crisis (more code than humans can review) and the answer now is also process. Thankfully, we have AI to help this time

In this issue

Articles

3 stories

sciencedaily.com

Claude Just Broke an 87-Year-Old Math Conjecture

Anthropic mathematician Levent Alpöge used Claude Fable 5 to uncover a surprisingly simple counterexample to the famous Jacobian conjecture, a problem mathematicians have wrestled with for decades. The discovery shows the conjecture is false in three dimensions and higher, while the two-dimensional case remains unsolved. More strikingly, the counterexample is tiny. It suggests AI could become especially powerful at searching enormous mathematical spaces for solutions humans have overlooked.

thenewstack.io

AWS Wants to Set Your Coding Agent Free

AWS is rebuilding Kiro around the Agent Client Protocol, an open standard designed to separate AI coding agents from the editors and interfaces they run inside. Think of it as an LSP-style layer for agents: developers could connect the same agent to different IDEs, terminals, or custom tools without rebuilding the integration. If ACP catches on, the real competition may shift from who owns the editor to who builds the best agents, governance, and developer experience.

aws.amazon.com

DynamoDB Just Became a Vector Database

In this issue

Upcoming Events

2 stories

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.

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

github.blog

GitHub’s Lawyers Are Building Their Own AI Tools

GitHub’s legal team is using Copilot CLI to build its own tools, without waiting for engineers. Lawyers created workflows for contract drafting, DMCA analysis, NDA triage, compliance checks, and risk assessments using mostly plain-language instructions and Markdown. One lawyer says his contract review and drafting time has been cut roughly in half. It is a useful glimpse at what happens when coding agents move beyond developers and into everyday business teams.

github.com

AgriciDaniel/skill-forge: Ultimate Claude Code skill creator

Skill Forge is an open-source toolkit for building production-ready AI agent skills. It goes beyond generating a SKILL.md, supporting planning, scaffolding, quality reviews, evaluations, benchmarking, and packaging. Skills can also be converted across platforms including Claude Code, OpenAI Codex, Gemini CLI, Cursor, and Antigravity. The project essentially treats agent skills like software: structured, testable, reusable, and portable across different AI coding environments.

In this issue

Podcast

1 story

latent.space

The AI Bottleneck Has Moved From Training to Inference

Getting a powerful model is only half the battle. In this Latent Space episode, Baseten’s Philip Kiely and Ali Taha unpack the growing discipline of inference engineering, covering quantization, KV caching, speculative decoding, GPU routing, and model parallelism. Some optimizations can make models several times faster without retraining them. It is a deeply technical look at the infrastructure work increasingly determining how fast, reliable, and affordable AI products actually become.