So, an 83% reduction in context tokens is impressive, but the bigger story - I think - is that code structure has become an operating expense. Every oversized file now taxes every future interaction with an AI agent. That’s a different economic model from the one most engineering organisations still use. Yet, the savings didn’t come from deleting code, they came from making the architecture legible enough that the agent could ignore most of it. That’s a subtle distinction, and one that matters far beyond token pricing, as it suggests that modularity has become an optimisation for machine attention as much as human comprehension. But, the experiment also exposes something many AI coding demonstrations quietly gloss over. Claude wasn’t discovering better architecture. It was executing a carefully designed engineering plan. The intelligence still sat with the human deciding which refactoring should happen, in what order, and how to verify that behaviour stayed unchanged. That’s much closer to today’s reality than the vision of autonomous software engineering. I’d like to see the same methodology repeated across mature enterprise codebases before drawing broad conclusions, but the experiment introduces a metric worth tracking. Measuring AI context cost alongside build time and test duration feels like a practical next step...

Every major communications platform eventually goes looking for schools! That’s where habits are formed, future customers are created, and large institutional budgets exist. The technology changes. The commercial logic barely moves. It’s easy to think AI in education is unprecedented until you see a 1939 advertisement making remarkably familiar promises about modernising classrooms, expanding access and helping teachers do more. Swap radio for large language models and much of the sales pitch survives intact. What the article leaves less explored is why these cycles keep recurring. Educational institutions are unusually sticky markets. Once a technology becomes embedded in lesson plans, teacher training and procurement, it can remain there for decades. Winning education isn’t simply about today’s revenue. It’s about shaping expectations for an entire generation. That’s why hardware vendors, broadcasters, internet platforms and now AI companies all arrive with remarkably similar ambitions.

I think the gap between the headline and the evidence is narrower than I expected, but it still matters. Whilst Google undeniably removed or sidelined a long list of RSS-related features over the years - that’s documented -, the bigger leap is concluding that Google “destroyed” RSS adoption. The article builds a plausible case, yet it stops short of proving that Google was the dominant cause rather than the most visible participant in a much broader shift towards algorithmic feeds and social platforms. But, products shape perceptions of standards and most people never chose RSS because they loved XML, they used it because Google Reader made it effortless. Once the flagship experience disappeared, the protocol suddenly looked obsolete even though nothing fundamental had changed. That’s a lesson we’ve seen repeatedly with open technologies. The article occasionally assumes intent where commercial incentives already explain the outcome. I’d have preferred more evidence there. Even so, the historical record it assembles is useful because it reminds us that open standards rarely disappear through technical failure. They fade when the largest distribution channels stop making them easy to discover. The protocol survived. The default user experience changed. That’s an important distinction.

History has a habit of making unfamiliar ideas look irrational when they’re often solving recognisable problems. This Ars Notoria struck me less as a spellbook than an ambitious attempt to engineer expertise. I think the extraordinary claims deserve scepticism, yet the underlying ambition feels surprisingly modern: compress the time needed to acquire difficult knowledge. Reading this with modern assumptions switched off makes the work much easier to understand, even if its promised outcomes remain unconvincing. What receives less attention is the infrastructure surrounding knowledge. Medieval universities had scarce books, slow copying and expensive education. Any technique claiming to accelerate learning naturally carried enormous economic value. Whether the method worked almost becomes a secondary question. The demand certainly existed.

There's a detail buried in the paper struck me more than the headline result; the mice could become equally habitual in their decision-making while differing substantially in how much they actually performed the habit. That’s an elegant distinction that feels obvious once stated, yet most discussions of habits quietly treat those two things as interchangeable. I think the authors earn credibility by resisting that shortcut, as they don’t simply identify another brain region involved in habit formation, they separate two questions that neuroscience has often bundled together: how an action stops depending on conscious evaluation, and how vigorously that action continues once it has become automatic. Those turn out to rely on different plasticity mechanisms in different prefrontal circuits. So, this work has an appealing systems quality. Instead of searching for a single “habit centre”, it describes specialised modules with distinct responsibilities. One circuit governs behavioural policy by deciding whether actions remain goal-directed. Another maintains execution intensity once that policy has shifted. That’s closer to how engineered control systems are designed than many earlier accounts of behaviour. If those principles translate beyond mice, they could reshape how researchers think about compulsive disorders. Interventions might eventually target excessive execution without necessarily disrupting the underlying behavioural repertoire.