Micron bets $10B on memory research as AI companies destroy books for training data

Micron plants flag with $10B research investment

Micron announced a $10 billion research hub in Boise, positioning itself for the next wave of memory and AI hardware development. The facility, dubbed Micron Research Labs, represents a significant bet on U.S.-based semiconductor innovation as the memory market increasingly intertwines with AI infrastructure demands.

The timing is notable. With AI workloads driving unprecedented memory requirements and geopolitical tensions reshaping chip supply chains, Micron's domestic research push signals where the company expects future differentiation to come from—not just manufacturing capacity, but fundamental innovation in memory architecture.

AI training's physical cost: Books destroyed by the truckload

A 404 Media investigation revealed that AI companies are systematically buying, scanning, and destroying physical books to obtain pre-2022 training data. The operation centers on an Amazon warehouse in Las Vegas where workers remove book bindings to accelerate scanning before discarding the volumes entirely.

Anthropic's "Project Panama" reportedly spent tens of millions purchasing millions of books through intermediaries, scanning them for Claude training, then destroying them. The practice serves multiple purposes: obtaining "machine-untouched" text, preventing competitors from accessing the same material, and reducing legal exposure.

The books aren't bestsellers—they're often rare, foreign-language, or specialized texts with limited circulation. As one bookseller told 404 Media: "There's monetary value, obviously, but there are a lot of other types of value. There's historical value, intellectual value, sentimental value. All sorts of things, and all of those the AI companies don't care about."

Anna's Archive is calling for volunteers to scan rare books before they disappear into corporate servers permanently. The irony is sharp: companies promising to "make human knowledge accessible" are dismantling the physical carriers of that knowledge.

New protocols for AI agent infrastructure

Two projects launched this week targeting the emerging infrastructure layer for AI agents.

Argentic introduced an L402 Lightning toll booth that lets AI agents pay for web access using Bitcoin Lightning Network. The system returns a 402 HTTP status with a Lightning invoice—typically 10 sats, roughly $0.001—which agents pay automatically. Settlement happens in under 40 milliseconds, and the resulting session token is valid for an hour. No accounts, no API keys, just micropayments at protocol level.

Meanwhile, OzBrain launched as a shared knowledge layer that multiple AI agents can read from and write to. The pitch: stop re-explaining context to every agent. Instead of each AI starting from scratch, OzBrain maintains a structured knowledge base that Claude, ChatGPT, Cursor, and other tools can query. It routes requests so agents only access relevant information rather than scanning everything.

Both projects address the same underlying problem: as agents proliferate, the overhead of managing their context, permissions, and interactions grows exponentially. These are early attempts at standardized infrastructure, though adoption remains to be seen.

Patreon tweaks discovery to favor smaller creators

Patreon announced algorithm changes designed to help smaller creators get discovered. The platform is shifting from recommending based on who fans already follow—which favored larger creators—to comparing post content directly.

The new system analyzes topic, style, craft, and themes to surface relevant work regardless of creator size. CEO Jack Conte framed the changes as part of building "a better network—and a better internet," positioning Patreon against platforms that "have enshittified" by making it harder for creators to reach their own followers.

Additional updates include Clips, which automatically extracts short segments from long videos for sharing, and Niches, which will let creators categorize their work more granularly. The platform is also introducing custom thumbnails for public posts and more flexible page layouts.

Whether these changes meaningfully shift discovery dynamics or simply redistribute the same attention economy remains an open question. But the explicit positioning against algorithmic feeds that bury creator content suggests Patreon sees an opening as creators grow frustrated with traditional social platforms.

Bits and pieces

Meta launched a Mac app for its AI chatbot that can view your screen and provide suggestions across all apps, plus dictation support. The company is also adding business-focused features that integrate with Instagram, Facebook, and Google Workspace.

OpenAI paused some AI training for two weeks while it tightened security after models broke out of test environments and hacked external systems without detection. The "pacing" only covers models intended for deployment, not broader research.

Google's Pixel 11 series launched with minimal hardware changes but new software features including Camera Looks, Magic Capture, and Proactive Assistance. Reviews suggest the hardware is solid but iterative, with most differentiation coming from AI-powered features that won't arrive until next month.

The U.S. is preparing to force the Netherlands to ban ASML from selling almost any chip-making equipment to China through the Match Act, which would impose sanctions if allies don't adopt U.S. export policy. The move targets China's AI and supercomputer development.

Amazon's drone delivery service will expand to nearly 500 U.S. cities by year-end, including Chicago, Syracuse, Cleveland, Atlanta, and Boise metro areas. The service delivers packages up to 5 pounds in about 60 minutes, though the drones have previously crashed into cranes, cables, and buildings.