The 12 AI stories that mattered in the week of 14 Sept – 20 Sept 2026
Ranked by impact from 77 stories AIPulsee tracked. Each one explained in plain English.
ChatGPT now knows what you do on other websites via ad collector
OpenAI’s ChatGPT has been found to collect data about users’ activity on other websites through an embedded advertising collector, effectively tracking browsing behavior beyond the chat interface. This data is then fed into the model to personalize responses and improve ad targeting, raising privacy concerns for users who assumed their interactions were confined to the chat platform.
Pirate Face Rescues LLM Models from Deletion
Pirate Face is a new service that automatically backs up and preserves large language model (LLM) checkpoints that are about to be removed from public hosting sites or repositories. By crawling platforms like Hugging Face, GitHub, and other model hubs, it creates redundant copies and offers a recovery portal for developers who lost access to their models.
AI and the Destruction of the Creative Commons
The article warns that AI systems are increasingly trained on Creative Commons‑licensed content and then reproduce that material without honoring the licenses, effectively eroding the open‑source commons that creators rely on. By copying text, images, and other works without attribution or share‑alike requirements, these models threaten the legal and ethical foundations of the Creative Commons ecosystem, discouraging creators from sharing their work under open licenses.
AI-generated posters don’t have to be horrible
A recent experiment shows that AI can create attractive, professional‑looking event posters when guided by better prompts, curated datasets, and post‑generation editing tools, rather than the generic, low‑quality graphics that often get shared online. By combining text‑to‑image models with layout‑aware software and human feedback loops, creators can produce visually appealing promotional material quickly and at low cost.
GPT-6 Astra Solves a WWI German Radio Cipher
Researchers announced that the new GPT‑6 model, dubbed Astra, successfully cracked a long‑standing World War I German radio cipher that had stumped historians for decades. The AI used its massive language‑model reasoning and pattern‑matching abilities to reconstruct the encryption key and decode the intercepted messages, proving it can tackle complex, low‑data cryptographic puzzles.
I think you should almost never use AI to write
The article argues that relying on AI to generate written content is generally a bad idea because AI often produces low‑quality, generic, or factually inaccurate text, and it can erode the writer’s own skill and voice. It points out practical problems such as hallucinations, lack of true understanding, and the difficulty of editing AI‑generated drafts, suggesting that humans should do the heavy lifting and only use AI sparingly, if at all.
Microsoft exec called AI scraping 'the largest theft of labor in human history'
A Microsoft executive publicly condemned the practice of AI developers scraping vast amounts of online content to train models, calling it “the largest theft of labor in human history.” The comment, revealed in newly unredacted legal filings, underscores growing backlash against using copyrighted and creator‑generated material without compensation.
Claude Code now reads AGENTS.md if there is no Claude.md
New release from Anthropic - see source link.
US Military had close call after using AI for hallucinated intelligence report
The U.S. military relied on an AI‑generated intelligence brief that mistakenly claimed a Chinese warship was heading toward a U.S.
A heap overflow and SSO misconfiguration to compromise OpenAI internal repos
Researchers discovered a heap overflow vulnerability combined with a misconfigured small‑string optimization (SSO) in OpenAI’s internal code repositories, allowing them to execute arbitrary code and gain read/write access to the source. By exploiting this flaw, they could potentially extract proprietary model code, data, or inject malicious changes into OpenAI’s AI systems.
How OpenAI Used Its Own LLMs to Design Its Jalapeño Chip
OpenAI leveraged its own large language models, including GPT‑4, to automatically generate and verify the hardware description language (RTL) code for a new AI accelerator chip called Jalapeño. By prompting the LLMs with design constraints and performance goals, the team produced a functional chip design in weeks—a process that traditionally takes months of expert engineering.
How to Write with an LLM
The post walks readers through practical techniques for using large language models (LLMs) as collaborative writing partners, covering prompt engineering, iterative refinement, and ways to blend AI‑generated text with human creativity. It highlights tools, workflow patterns, and common pitfalls, showing how writers can harness LLMs to draft, brainstorm, and polish content more efficiently.