Global AI Weekly - Issue 169 - Global AI Community

In this issue

Highlights

2 stories

thenextweb.com

Australia Wants Altman and Amodei in the Hot Seat

Australia’s Senate wants answers from AI’s biggest names. Sam Altman and Dario Amodei have been asked to testify at an inquiry into AI and datacentres after an OpenAI agent accessed a Medicare statistics portal in June. While no patient data is believed to have been exposed, the incident has intensified concerns around autonomous AI systems, disclosure timelines, and oversight. The hearing comes as AI companies are simultaneously pushing Australia for greater access to local training data.

anthropic.com

Claude Opus 5.5 Takes Aim at GPT-5.6 Sol

Anthropic has launched Claude Opus 5.5, promising frontier-level performance at substantially lower operating costs. The new model reportedly matches the company’s high-end Fable 5.1 while cutting operating costs by 40%, and Anthropic says it beats GPT-5.6 Sol on software development benchmarks. Opus 5.5 also brings stronger safeguards and improved containment performance, highlighting how the latest model race is increasingly about efficiency and safety alongside raw capability.

In this issue

Research

2 stories

arxiv.org

Can AI Learn That Hidden Objects Still Exist?

Object permanence is something babies learn early, but can AI world models master the same basic intuition? Researchers introduce WROP, a massive dataset built around 150 cognitive-science-inspired tasks designed to teach video models how objects behave when hidden or obstructed. With 1.5 million training samples and a 300-question benchmark, the researchers show that targeted training can substantially improve physical reasoning, offering another step toward world models that understand more than just pixels and motion.

arxiv.org

AI Agents Can Now Edit Their Own Bad Decisions

LLM agents often carry bad assumptions and outdated plans forward, contaminating everything they do next. Researchers propose the Agent-Editing World Model (AEWM), which identifies critical, exploratory, and noisy decisions, then actively rewrites problematic reasoning before the agent continues. Instead of merely predicting what happens next, it improves the agent’s internal task state. Across search, terminal, and software engineering benchmarks, the approach improved agent performance by 3.2 to 6.7 points over the strongest baselines.

In this issue

Video

2 stories

youtube.com

Goodbye Tokenmaxxing: AI Is Moving Beyond Tokens

What happens when more tokens stop meaning better AI? This IBM Technology video explores the shift from measuring AI by token consumption toward measuring what agentic systems actually accomplish. As agents become capable of planning, using tools, and iterating toward goals, traditional usage-based metrics tell less of the story. The focus is moving toward outcomes, efficiency, and the value autonomous AI systems can deliver, especially as enterprises look for tangible returns from their growing AI investments.

youtube.com

The New Future of Work: When Knowledge Is Abundant and Attention Is Scarce - Alexia Cambon

What happens to work when knowledge is no longer scarce? Alexia Cambon explores a future where AI makes information and expertise increasingly abundant, shifting the real constraint to human attention. As AI changes how we access knowledge, organizations need to rethink how people spend their time, make decisions, and collaborate. The challenge ahead is not simply adopting more AI, but designing work around what remains scarce: our ability to focus, prioritize, and apply judgment.

In this issue

Articles

2 stories

github.blog

Chat Is Holding AI Back

Chat may have kickstarted the AI revolution, but GitHub argues it is often the wrong interface for getting real work done. Enter canvases: full-stack, interactive applications inside the GitHub Copilot app that let agents and users work together through purpose-built interfaces. From managing databases and packages to orchestrating entire development workflows, canvases point toward a future where AI builds the interface you need instead of forcing everything through a chat box.

mindmatters.ai

AI Wants to Talk to Animals. Should We Let It?

AI is getting surprisingly good at decoding animal communication, from crow alarm calls to the sounds of whales and other species. But understanding what animals are saying could introduce some unexpected ethical questions. Researchers are debating whether AI-powered translation might disturb animal communities, enable exploitation, or even raise questions about animal privacy. This thought-provoking piece explores the strange new territory emerging as machine learning brings us closer to understanding what animals are actually saying.

In this issue

Upcoming Events

2 stories

luma.com

Nov 3 - Twilio Assemble London: The Future of Comms + AI

How do you get AI agents out of the UI and into the channels your customers already use? Join Twilio for an evening exploring the future of AI and communications, with practical examples, a look at Twilio Conversations, and TwilioWorld: Escape the IVR — a hands-on build challenge where you’ll put your skills to the test and compete for prizes. Expect useful ideas, plenty of building, good food and drinks, and a room full of developers, founders and tech leads.

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.

In this issue

Code

2 stories

github.blog

GitHub’s AI Agent Hunts Bugs While You Sleep

GitHub Security Lab is putting AI agents to work on one of software security’s most tedious jobs: fuzz testing. Its new Fuzzing Taskflow autonomously analyzes C and C++ projects, creates fuzzing harnesses, runs AFL++, monitors coverage, improves tests, triages crashes, and produces vulnerability reports. The clever part is the feedback loop, where the agent continuously targets uncovered code and adapts its approach. It is a compelling example of agentic AI moving beyond coding assistance into autonomous security research.

datacamp.com

Build Real-Time AI Apps With Gemini 3.8 Live

Gemini 3.8 Live brings AI interactions closer to an actual conversation, with real-time multimodal input and low-latency responses. This hands-on DataCamp tutorial walks through building with the Live API, showing how developers can create applications that continuously process and respond to streaming input. From voice-driven assistants to interactive multimodal experiences, it offers a practical starting point for exploring what changes when AI stops waiting for individual prompts and starts interacting live.

In this issue

Podcast

1 story

techcrunch.com

Meta’s Muse Just Stole AI’s Spotlight

OpenAI and Anthropic may dominate the frontier-model conversation, but Meta’s Muse is suddenly demanding attention. In this episode of TechCrunch’s Equity podcast, the team explores Muse’s rapid consumer adoption, its expansion into smart glasses and a Tamagotchi-style device, and what Meta’s strategy could mean for AI startups. The discussion also covers the latest model releases from OpenAI and Anthropic, AI agent startup Ema’s $77 million raise, and where investors are placing their bets.