Sean Goedecke chases down the stat underpinning half the AI-bubble case, that inference GPUs last “three years at most” under load. He traces it to an anonymous “GenAI principal architect” quoted via Tegus, a firm that pays insiders by the hour. Goedecke argues the incentive there is to sound confident rather than accurate, and points to evidence that datacentre GPUs depreciate over a longer horizon than the doom case assumes.
Charity Majors pushes back on readers who took her last piece as a licence to skip code review and push slop to prod. Her point is the opposite: now that models write median-quality code faster and more cheaply, the cost shifts to everything around the code. Tests, review, and validation become more load-bearing, not less, the same way immutable infrastructure made discipline mandatory after the era of hand-tended server pets.
Drawing on a year at Imprint, Will Larson lays out how AI tooling rewires engineering management. Complex migrations can now be 95% owned by one person in a tenth of the time, which raises the stakes on individual judgement and makes a single sharp edge costly to colleagues. First-pass code is nearly free, he notes, but working code still depends on your harness: tests, CI, validation, and preview environments.
Rohit Krishnan tested Karpathy’s “LLM Council” idea, where several models critique each other and a chairperson synthesises, against simpler setups. Breaking each answer into atomic idea-cards and scoring them blind, he found the council keeps only a minority of the genuinely good single-model ideas while giving consensus points an extra push. Like a human committee, it smooths out the spiky, idiosyncratic answers that are often the most valuable.
Murphy Trueman starts from Uber’s uSpec, an agent that auto-documents components by reading their structure through a Figma MCP, and draws out the condition that makes it work: the design system is explicit enough to read programmatically. When the structure is a real contract, agents produce accurate output; when it is a loose guideline, they fail or invent. Treating each component as a promise, not a suggestion, is now a build requirement.
A regional court in Munich held Google strictly liable after its AI Overview invented defamatory statements, placing two publishers beside scams atop search results. The hinge: a normal search result points to a third-party page you can sue, but here the AI was the author, so there was no one else to blame. Google’s “everyone knows not to trust AI” defence failed. It is one appealable ruling, not settled law, though the logic travels.
