LLM Six Month Review Includes Pelicans.
Also, Floppy Disks and Bad Lunch Tracking.

The Quarterly LLM Retrospective: Apparently, Nobody Saw The Pelicans

Six months of unprecedented, frantic LLM innovation have been perfectly summarized by the prompt: "Generate an SVG of a pelican riding a bicycle." This seemingly absurd phrase was adopted as a personal LLM benchmark by engineer Simon Willison, who found the impossible task of rendering the wrong-shaped bird on a bicycle frame was a better gauge of model reasoning than any official leaderboard full of vanity numbers.

The benchmark worked because text-based LLMs should not be able to draw anything; yet, they can output SVG code, which is technically text. The models try their absolute best to complete the uncompletable task, providing a glimpse into their internal logic, which is the exact moment they stop being an all-powerful oracle and become a very confused intern. The irony has now achieved critical mass, as the author has noted that even Google appears to have started optimizing its models to pass the Pelican on a Bicycle test, accidentally turning a witty joke into the industry standard.

IT Procurement Approves 'Democracy' Server Racks For Autocracy

In what can only be described as a classic corporate compliance oversight, American AI firms are moving forward with selling world-class AI supercomputing clusters to autocratic governments. Former AI safety official Helen Toner pointed out this week that the key to national power is now access to these massive clusters of advanced Nvidia chips, essentially making the US the global supplier of future geopolitical leverage.

The argument from the companies is that this is a noble effort to spread "democratic AI rails" globally, which is the kind of mission statement that sounds great on a PowerPoint slide until it clashes with the fact that these are autocratic regimes. When pressed, the sales teams default to the universal IT excuse; "if we didn't sell it to them, someone else would," ignoring the fact that the 'someone else' (like China) cannot match the American supply in the first place. This is not democracy promotion; this is just hitting the quarterly sales number with extra steps.

AI Calorie Apps Produce Data Hallucinations About Lunch

The latest attempt to replace human effort with an opaque algorithm has failed exactly as expected, this time in the domain of tedious self-tracking. New AI-powered calorie counting apps promised users they could simply snap a photo of a meal and receive an accurate nutritional breakdown, eliminating the need for manual logging. Instead, they delivered what analysts called results even worse than expected, because a 2D photo cannot accurately measure the volume of a 3D meal.

One application even had the audacity to suggest a weight loss goal that would have placed the user into an underweight BMI category. Another app required users to manually confirm and input the portion size anyway, completely nullifying the core feature and reducing the AI to a glorified photo album for your regrettable life choices. The apps, like Cal AI and Calorie Mama, turn out to be the perfect symbol of the modern tech product: a vast, expensive, powerful system that only confirms your eyeball estimate was just as inaccurate.

Briefs

• Tech Debt Audit: The Federal Aviation Administration (FAA) is finally, formally planning to eliminate floppy disks from its air traffic control systems. This multi-billion dollar project, which will also phase out Windows 95, comes after a recent evaluation found a third of systems unsupportable, because the department waited until the physical magnetic media had essentially turned to dust.
• Cloudflare AI-Code: An analysis emerged of a new Cloudflare OAuth library that was written entirely by an AI. This marks the moment we transitioned from "will AI write our code" to "did an AI write this security vulnerability."
• Nostalgia Corner: A deep dive on the long-deceased HTML tags <Blink> and <Marquee> reminds us that user experience peaked in 1999 when everything on the screen moved and flashed constantly.