Amazon's Book Scanning Facility Caught on AirTag

For a while, rare book dealers had been noticing something odd: bulk orders, sometimes thousands of books at once, from buyers who didn't seem to care about price. The working theory was AI companies scanning physical books for training data. Now there's proof, according to 404 Media's investigation.

One bookseller agreed to hide an Apple AirTag inside a large order and let 404 Media follow it. The tag led to the VGT3 section of Amazon's LAS8 facility in north-east Las Vegas. The entrance displayed a logo of a dinosaur holding a book. Workers in online forums confirmed that VGT3 destructively scans books in bulk — meaning the books are taken apart in the scanning process and cannot be resold.

Whether this is legal is genuinely unsettled. Copyright law was not written with AI training data in mind, and no court has ruled definitively on whether scanning copyrighted books for this purpose constitutes infringement. Amazon has not commented on the investigation.

For readers: if you sell books online and receive unusually large orders from price-insensitive buyers with minimal details, you are likely looking at this kind of operation. Whether you want to fulfil those orders is your call — but now you know what they are.

Qwen 3.8 27B is a vision-capable language model (an AI that can process both text and images) from Alibaba, licensed under Apache 2.0 — meaning you can use and modify it freely, including for commercial work. At 27 billion parameters (the internal numbers that shape how a model responds), it sits in a sweet spot: capable enough to be useful, small enough to run on a well-specced laptop.

Blogger Simon Willison tested it on two machines and found the results impressive — but complicated. By default, the model is set to its highest reasoning mode, which causes it to overthink even trivial requests. In testing, a simple drawing prompt took 21 minutes and burned through tens of thousands of reasoning tokens. Switching reasoning off brought that same task down to two minutes. Alibaba's own benchmarks look strong, though independent evaluations have not yet reported in.

It is a 17GB download. Free to use.

Honest read: Genuinely good model, but switch the reasoning setting to low or off for anything routine — the default is set for complex tasks and will waste your time on simple ones.

What it is: A scheduling system that decides which AI jobs run on which GPUs, and in what order, to get more useful work out of the same hardware.

What's new: Most GPU clusters use FIFO scheduling (first in, first out — jobs run in the order they arrive, regardless of priority or size). This system replaces that with a constraint-aware allocator that makes each allocation decision based on job priority, duration, and demand shape.

How it works: The allocator treats each scheduling decision as a grid: every GPU, across every time slot, assigned to a job or left empty. It distinguishes between two incompatible job types. Batch jobs — training runs, bulk inference, model compression — need a fixed block of GPUs held continuously until they finish. Real-time inference is elastic: it scales up and down as user traffic changes. FIFO handles this by reserving each real-time application's peak GPU count for the full day, leaving those GPUs idle overnight. The new allocator releases GPUs during low-traffic periods and reclaims them before the next peak, freeing that capacity for batch jobs in the gap.

What it can't do: The gains are measured on seven internal benchmark scenarios. Real-world workloads vary, and the authors do not claim the 33-point utilization improvement generalises to every cluster configuration.

Who should care: Engineering teams running mixed AI workloads on shared GPU infrastructure — the scheduling gains are meaningless on a lightly loaded cluster with plenty of spare capacity.

Read it: Hugging Face Blog

LM Studio is a free desktop app that lets you download and run open-weights models (models whose trained files are publicly available for download) on your own machine, with no data sent to a third party.

Qwen 3.8 27B is available in LM Studio now. Search for the Q4_K_M quantized build — quantization (a compression technique that shrinks the model file while preserving most of its quality) brings it to a 17GB download. The source article notes this specific file size for the build tested.

Once loaded, the first thing to change is the reasoning effort setting. Go to model settings and drop it from "xhigh" to "low" or off entirely for everyday tasks. Then set the context window (the amount of text the model holds in memory at once) to at least 32,768 tokens, or higher if your machine allows.

To get started: download LM Studio, search for Qwen 3.8 27B, and select the Q4_K_M build.

Worth knowing: The 17GB file requires significant free disk space, and performance will depend heavily on your available RAM — check LM Studio's model card for your specific build before downloading.

Anthropic has confirmed that Claude will start embedding invisible watermarks into its text output, according to The Verge. The EU AI Act requires AI-generated text, images, audio and video to carry machine-readable markers identifying them as synthetic. Claude's system uses a version of Google DeepMind's SynthID-Text approach, which encodes a hidden pattern by nudging word choices at low-stakes moments — places where two words are equally valid and the reader wouldn't notice either way. Anthropic says this won't change output quality or cost. Google's Gemini already uses SynthID. OpenAI has not yet detailed its compliance plan.