AI performance review forms are flawed.
Also sailboat logistics and fake product leaks.

The Performance Review Is Always The Problem

It turns out the new hire class of Large Language Models, which were supposed to be the "10x coders" of the future, might have simply been graded on the wrong curve. A study from the Oxford Internet Institute identifies severe weaknesses in AI evaluation methodologies. The general finding is that we are scoring the AIs on parlor tricks and standardized tests, overlooking how they might behave when given an actual job in a production environment.

Researchers claim we are assessing the models based on unrealistic benchmarks that do not account for systemic risks; this is like giving a promotion based solely on speed-typing tests without checking for workplace decorum or how often the employee files a completely wrong expense report. The collective tech industry keeps celebrating models that are fast at predicting the next word, but apparently we forgot to check if the next word is a terrible, confidently generated lie. The solution is not better AI; it is better HR documentation.

Apple Becomes The Software HOA Board

The Free Software Foundation Europe is having a minor panic attack because Apple is continuing its process of making software installation an official ceremony called "notarisation." The process is pitched as a way to keep users safe from malicious software, which is true, but it also has the side effect of requiring developers to get a rubber stamp from Apple before their code is allowed to exist in a happy little Mac sandbox.

This is exactly the same as requiring all external documentation to be printed on Apple-approved, custom-perforated paper before it is allowed into the corporate park. It is a classic corporate move where security is merely the first, thin layer of justification for centralized control. It ensures that anything unauthorized, even if completely harmless, must remain outside the perimeter, living a sad, un-notarized life.

Office Gossip Department Generates Major Product Leak

A group of people proved that the most sophisticated piece of machinery in the tech rumor cycle is still four bored humans with a spreadsheet. They openly admitted to accidentally faking the Ryzen 7 9700X3D leaks, which were then picked up by major tech publications. The initial joke was just a few spreadsheet cells shared privately; the accidental nature of its virality is the truly depressing part.

It seems that the entire supply chain of speculative hardware news relies on the principle that if something looks like an internal memo from an anonymous source, it must be true. This particular oopsie confirms the core business model of tech journalism: generate highly detailed internal-looking documents, wait for the media to run with the story, then claim plausible deniability. The rumor mill is basically a self-licking ice cream cone of anonymous speculation.

Supply Chain Reverts to Version 1.0.0

In a bold move that redefines "innovative sustainability," the world’s largest cargo sailboat successfully completed its first Atlantic crossing. The boat, a giant shipping vessel that has essentially outsourced its propulsion to the wind, is a testament to the fact that when faced with a complex modern problem, the most reliable solution is often one we ditched a few centuries ago.

This is peak corporate efficiency: take the existing container ship infrastructure, bolt on some really large sails, and call it 'net-zero shipping.' It is a perfectly symbolic move for an industry that loves to talk about moving forward while actively dusting off old code and hoping nobody notices. The next step will be to bring back messenger pigeons for low-latency document transfer.

Briefs

• VPN Shutdown: Mullvad is shutting down its search proxy Leta. The team is calling it a "sunsetting" of the project; users are calling it another feature getting randomly cut after they started relying on it. Standard operating procedure for the VPN business unit.
• AI Speed Contest: Cerebras Code now supports GLM 4.6 at 1000 tokens per second. We have now reached a speed where the AI can generate corporate boilerplate faster than a human can read it, which is the only metric that matters for enterprise adoption.
• IT Archaeology: A 52-year-old data tape could contain Unix history, specifically the Fourth Edition. A true historical find, right up until someone plugs it into a machine and discovers it is just a backup of a dozen forgotten punch-card invoices.