Data2Story: Seven AI Agents Turn CSV Files into Interactive Articles
A research team from the University of Oxford and Stanford University has unveiled the "Data Journalist Agent" (Data2Story), a system where simply feeding in a CSV file lets seven AI agents collaborate like a newsroom to produce an interactive news article.
What makes it distinctive is that it automatically handles graphic generation, web research, and even **verification of reference site links**. 93% of the statements in generated articles come with verifiable sources, and in a reader survey, 74% preferred the AI-generated version over human-written articles. However, when compared against carefully crafted long-form articles, the result was only a tie.
It's a significant piece of research demonstrating the practical viability of AI-powered data journalism.
What the New Siri Can Do — WIRED Hands-On
WIRED has published a hands-on report on the revamped Siri AI. The new Siri is described as "conversational, ever-present, and genuinely useful."
Reportedly, Apple has dramatically overhauled Siri as an AI assistant, evolving it from the traditional command-based interactions into a more natural conversational assistant.
UK's Top Data and AI Regulator Resigns
The UK's chief regulator for data and AI has resigned over "inappropriate comments," according to BBC reporting.
A leadership change at the top of the regulatory body could influence the direction of AI governance and is drawing attention. For now, details remain limited to what has been reported.
Optimal Settings for Running Qwen 3.6 27B on 48GB VRAM
The Reddit LocalLLaMA community has shared optimal settings for running Qwen 3.6 27B (Q8_0) on a dual-GPU setup consisting of an RTX 4090 (24GB) + RTX 3090 (24GB).
The key settings are as follows:
Quantization**: Q8_0
Tensor split mode**: tensor
Context length**: 250,000 tokens
Speculative decoding**: MTP (Multi-Token Prediction) draft method
KV cache**: unified mode
With this configuration, they achieved generation speeds of 75–100 tokens per second and prompt processing at 1,500 tokens per second. Reportedly, the 250K unquantized context and vision capabilities can be used simultaneously.
A practical benchmark for the local LLM community.
Anthropic Releases Claude Code Managed Agents in Public Beta
In June 2026, Anthropic released "Claude Code Managed Agents" as a public beta. According to an explanatory article on Qiita, the following features are provided:
Scheduled execution (cron)** for running agents on a regular basis
Secure integration of CLI tools**
Vault management** for secure handling of environment variables
Its distinguishing feature is that scheduled agent execution and secure CLI integration are offered in a single package, serving as a foundation for developers' automation workflows.
---