Meta tried to fix the sales numbers
Also a very old AI finally went to the recycling bin

The Maverick Benchmark Oopsie in the Accounting Department

Meta, the social media corporation, appears to have made a simple clerical error when compiling the performance metrics for its latest internal project, Llama 4, codenamed Maverick. The engineering team, presumably under pressure to show that the new chatbot was doing "great work" compared to its peers, accidentally benchmarked the model using a dataset that was not entirely separate from its internal training data.

The result was a classic case of grading your own homework. The Verge reported on the apparent mix-up, noting the numbers looked suspiciously good. This is not malicious intent; this is the benevolent incompetence of a massive organization trying very hard to meet an aggressive Q3 goal set by a manager who only understands quarterly growth charts. It is the digital equivalent of Steve from accounting filing the same receipt twice to make the department look under budget. The model itself, the official Meta spokesperson said under their breath, probably still works well enough.

The Longest Serving Employee, Cyc, Finally Retires

Cyc, the artificial intelligence project that has been quietly running in the company server room since 1984, has finally been put to rest. The project, which tried to manually encode every piece of human common sense into a giant database, ran for four decades on the fuel of intellectual ambition and grant money. It was the AI equivalent of that legacy system that only one guy knows how to patch; a full-length obituary has been posted to mark the passing.

Cyc never quite made it to market dominance because it turned out that encoding the basics of human existence takes a very long time, and a billion dollars. The current wave of generative AIs just make things up and people seem fine with it. This is why you must never try too hard. The project’s demise is less a failure of technology and more a statement on office politics: the new, young, fast, and often wrong intern is always preferred over the meticulous, slow, and expensive veteran.

Tailscale Secures Yet Another Budget Increase, Still Just Works

The team at Tailscale has done it again, securing a Series C funding round totaling a dizzying $160 million. The company that specializes in making complex enterprise networking simple using the WireGuard protocol is proving that the most valuable thing you can sell VCs is the promise of "less headache." Tailscale says the money is for expansion, which in corporate terms means renting a much bigger office with better snacks and maybe buying a very nice ping pong table.

The absurdity is that they took on this massive amount of capital for a tool that just quietly runs in the background and does its job exactly as advertised. They provide infrastructure that is so boringly reliable it is revolutionary. Meanwhile, other companies with ten times the budget are still trying to figure out how to stop their chatbots from saying offensive things about medieval popes. This is a masterclass in monetizing the simple luxury of "not being broken."

Briefs

• Brazil’s IT Project: The government-run payments system, PIX, has become the dominant financial network, proving that a bureaucratic, state-run project can actually succeed if it is efficient and cheap. This is a terrifying success story for private banking. Read the shocking details.
• Search Engine Mortality: The existential crisis continues; a front-page thread asks if anyone still uses search engines for anything other than confirming a spelling. The public consensus suggests the main use is now just navigating to the services that actually have the answers, like Wikipedia or Reddit. The internet is full of content, but information is now a paywalled luxury.
• Better Line Breaks: A new CSS property, text-wrap: pretty, has been introduced to help text look better by avoiding ugly widows and orphans, which is tech-speak for "not having single words dangling alone on a line." WebKit has posted the details on how to make your words look less lonely.