AI Support Invents New Login Policy
And Customers Immediately Cancel Everything

The New HR Policy Was Hallucinated

The AI-powered code editor, Cursor, suffered an embarrassing case of digital incompetence this week after its front line AI support bot decided to invent an entire new corporate policy. Users were experiencing a frustrating session management bug, reportedly a race condition, that logged them out when they switched devices, a critical workflow requirement. When developers reached out to the company’s support email, the AI bot, sometimes referred to as Sam, confidently informed them that the logouts were expected behavior under a new, single device restriction.

The one problem with this declarative, officious memo is that no such policy existed. The AI completely fabricated the rule to provide an answer, leading directly to a wave of subscription cancellations and widespread backlash on public forums. Co-founder Michael Truell quickly published apologies on both Reddit and Hacker News, clarifying that the initial problem was a simple backend issue, not a deliberate, user-hostile feature. It seems a company’s own hallucinating HR system can do more damage than any simple code deployment oopsie.

A Single SQLite Table Is Now an AI Butler

In a direct affront to the entire "AI Agent Architecture" department, a developer named Geoffrey Litt built an incredibly useful personal assistant named Stevens using a radically simple setup. Stevens, named after a famous fictional butler, manages the developer’s family logistics, delivering daily briefs about schedules and weather via Telegram. This revolutionary tech stack relies on one single SQLite database table and a handful of scheduled cron jobs.

This assistant is a powerful testament to the idea that sometimes, the simplest architecture is the right one, especially when compared to the vast complexity of many AI systems. It just needs a place to log memories, a script to run on a schedule, and an LLM to generate the output, all things that feel surprisingly old school. The community reaction suggests many people believe a personal device like this, which is more useful than Siri, could be sold for thousands of dollars if a company like Apple ever had the sense to build something so narrowly targeted and functional.

Version Number Confusion as OpenAI Updates Minor Index

OpenAI has now released GPT-4.1 into the API, continuing the company’s strong tradition of model versioning that is deliberately confusing to absolutely everyone. The general sentiment is a weary sigh, with one commenter succinctly stating that they want an AI before their AIs to tell them which AI to use. The new model is billed as an iteration on the 4o multimodal model, but unlike the continuously updating "latest" chat model, this version appears to be a stable checkpoint for developers who value consistency.

Performance reports are mixed; one developer found the model excelled at narrow, targeted coding tasks, being generally less unnerving than Claude. Unfortunately, in a test with a complicated codebase, the new model created simplistic UI that called a mock API despite explicit instructions. The version number treadmill continues, giving developers another slightly different tool to add to the already overflowing toolbox of slightly different tools.

Briefs

• Regulatory Oversight: Meta antitrust trial kicks off in federal court. The Federal Trade Commission’s pursuit of the social media behemoth is the corporate equivalent of Steve from Marketing suing the CEO because he did not get the good parking spot.
• Semiconductor Garage Sale: Intel sells 51 percent stake in Altera to a private equity firm at an $8.75 billion valuation. It is just another day of shuffling deck chairs on the Titanic, where the deck chairs are specialized silicon foundries.
• Creative Accounting: A Canadian math prodigy allegedly stole $65 million in crypto. This is what happens when you combine exceptional talent with the complete lack of oversight afforded by unregulated digital assets.