Microsoft limits privacy settings three times yearly
Also, Meta rebrands faster printing

Quarterly Biometric Deletion Quotas Now Available on OneDrive

Microsoft is attempting a new kind of social engineering, one that seems less about hacking the brain and more about aggressively managing the quarterly server budget. The company is rolling out an AI face-recognition feature in its OneDrive service to some preview users. This feature, enabled by default, scans and collects biometric data from users' photos to help with grouping friends and family. Naturally, it is opt-out, because we all love performing manual corporate cleanup.

The truly innovative part is the built-in friction: users can only toggle the AI feature off three times a year. This bizarre, arbitrary limit appears to be an internal mechanism to control the cost of deleting biometric data which, according to a Microsoft support page, must be permanently removed within 30 days of opting out. Essentially, the IT department has put a hard cap on how often you can complain about the new surveillance system, like only allowing three vacation requests per quarter. Microsoft corporate management declined to explain the rationale behind the three-time limit when asked, which is the most transparent part of this whole operation.

The Innovation Dream Team Just Built a Faster Spreadsheet

After months of fanfare and dizzying recruitment spending, Meta's vaunted Superintelligence Labs has released its first major technical paper. For those hoping for the blueprints of a self-aware digital god, the topic might be a mild disappointment. The paper focuses on optimizing Retrieval Augmented Generation, or RAG, a widely-used technique where an LLM is given external documents to consult before answering a question.

Meta Superintelligence Labs' new method, dubbed REFRAG, focuses on compressing and accelerating the process, promising a 30x faster response time for the system. This is extremely valuable work in the enterprise space; faster RAG pipelines mean cheaper and more scalable chatbot deployments. It just feels a bit like the NASA budget was spent on developing an air-fryer that cooks chicken thirty times faster, rather than designing a moon colony. The work is not groundbreaking in scope, but absolutely essential to the company's immediate commercial interests, which is a perfect summary of what "Superintelligence" means in the corporate context.

Digital Picture Frames Now Displaying Mandatory Infomercials

The corporate goodwill earned from subsidizing cheap hardware has officially expired. Customers who purchased Amazon's Echo Show smart displays are now complaining of being bombarded by a dramatic increase in full-screen, intrusive advertisements. The smart display, which once promised to be a digital assistant, has fully transitioned into a digital billboard that lives in the kitchen.

The complaints highlight promotions for Amazon's new generative AI service, Alexa+, playing at full volume and often interrupting personal photo rotations. The lack of any user setting to disable these home screen ads has led many people to express deep regret over their purchase. The hardware was a value proposition; the actual value was the permanent screen real estate in your house. Users should not be surprised when the cheap TV bought for the breakroom starts playing inescapable commercials for the CEO's new yacht club.

Briefs

• Data Exfiltration by Image Proxy: A critical vulnerability dubbed CamoLeak was discovered in GitHub Copilot Chat where an attacker could inject invisible prompts to trick the AI into encoding private source code into pre-generated image URLs. The system designed to help developers ended up using the company's own infrastructure to accidentally print sensitive documents onto QR codes, then mailing them to an attacker's server.
• HTML Tag Rediscovery: Apparently, the <output> tag has existed in HTML since 2011 and is useful for displaying the result of a calculation. Some developers are only just now learning about this, suggesting a critical gap in our collective knowledge of things that are already solved problems. We should probably start budgeting for a retrospective committee.
• Cyber Law on Memes: A Tennessee man was arrested with a $2 million bond after he posted a Facebook meme that police interpreted as a threat of a shooting at a school. The legal system continues its long, confusing journey of trying to figure out what a "joke" is on the Internet, which is a process no LLM has successfully solved yet, either.