Here, one of the most useful instructions in this playbook, involves a rather unglamorous example: a medical billing application that understands the peculiarities of the 340B drug programme. Anthropic suggests turning precisely this kind of industry-specific edge case into a repeatable test, building a collection of scenarios that generic competitors would struggle to handle. I think that points towards a more convincing competitive advantage than the much-publicised ten-person unicorn. When AI can generate competent applications cheaply, knowing exactly where those applications fail becomes extraordinarily valuable. A founder who understands the obscure exceptions, regulatory quirks and operational compromises within an industry has something that a general-purpose coding agent cannot simply conjure from a prompt. More interestingly, that expertise can become executable knowledge. Each customer-reported exception becomes another scenario to test, another assumption made explicit and another opportunity to distinguish a genuinely useful product from a convincing demonstration. A growing suite of real-world tests also gives engineers something concrete against which to evaluate new models, prompts and architectural changes. There is a danger of collecting an impressive library of historical failures while missing emerging problems. The feedback loop must remain connected to actual customer behaviour.

We are living in an age that television now inserts an advert before its we can open Netflix, and this is making a remarkably candid statement about whose attention it thinks it owns. That's the detail we get from Sarah Perez's report; it explains why people are prepared to hand an AI agent fairly invasive access to a perfectly functional appliance just to get back a screen with a handful of apps on it. Also, this weekend I've started doing the same to mine... Anyway... It's a curious reversal here; where manufacturers have packed more software into televisions to extend their commercial usefulness after purchase, and we, customers, are now using a new generation of software to remove the extras. The impressive trick isn't building a launcher; plenty already exist, as the article acknowledges. It's that a request in ordinary language can be translated into model-specific actions over Wi-Fi. The barrier between annoyance and modification has fallen far enough that a family TV becomes an afternoon project. Features that survive because removing them is tedious are exposed when the tedium can be delegated, of course, replacing a launcher should not require gambling with an operating system, and the reported bricked set is a useful reminder of what enthusiastic demonstrations leave out. I'd like this to end with a less heroic outcome: televisions shipping with a proper 'just show my inputs and apps' setting. If customers are writing agents to delete the user experience, perhaps the experience needs deleting...

A proof running to hundreds of thousands of lines of Lean is a strange form of scientific success.... It may be formally correct while remaining almost impossible for a human to absorb. Imagine receiving a fully verified answer that nobody on the team can explain. For a famous conjecture, that still settles a question. For the rest of science, it could be a rather expensive dead end. The result has to travel. Someone must recognise which ideas generalise, which assumptions matter, and which experiments become possible. This is why Maggi's notion of mathematization deserves more attention than the headline battle between human and machine mathematicians. Physics did not advance merely because someone proved increasingly difficult statements inside an existing vocabulary. It advanced when people found new ways to describe the world, creating questions that hadn't previously been expressible. AI might make that work easier by clearing away proof obligations. Equally, a flood of formally certified results could crowd out the slower work of interpretation. There's something faintly absurd about imagining a future where we've solved the hardest problems in mathematics but need another AI to explain what the answers mean. A library full of certificates is a remarkable achievement. Whether it remains a living mathematical culture depends on what humans can do with them.

De Batz sitting off Dammam while sections of Meta's 2Africa Pearls cable lie on the seabed, still unconnected - that's the image that stays... Sure, the technology exists, the investment has been made, and the vessel is there, but a war has turned the last few kilometres into an indefinite scheduling problem. It's a particularly uncomfortable illustration of how reassuring a network diagram can look while concealing the practical dependencies underneath. Most reliability planning treats repair time as an engineering variable: source a spare, dispatch a crew, restore service. Bagley's examples expose a nastier category of failure. If a ship cannot safely enter the water, mean time to repair stops being a statistic you can extrapolate from historical incidents. It becomes a political question with no service-level agreement. Planners can buy another route, but a second path through the same diplomatic choke point may simply double the sunk cost. This suggests a more demanding way to evaluate international connectivity. Alongside latency, capacity and route diversity, ask who can authorise repair work, which ports a cable ship can use, and what happens if access is denied for six months. That doesn't make global cooperation unnecessary; it makes the operational assumptions of every proposed alternative painfully concrete. When a cable is broken, the map tells you where it is. It cannot tell you when anyone will be allowed to reach it.

The unlikely hero of printed books in Joseph's account is TikTok... whilst short videos compete with reading for attention, BookTok helped drive a reported 70% surge in book sales. That's a useful warning against treating every new content technology as a straightforward substitute for what came before. Sometimes the channel that appears to be eating an industry turns out to be its best shop window. So, sure, AI-generated media will make producing a competent image, song or video cheaper, but that sounds wonderful until everyone can do it - literaly! The economics then drift away from making the thing and towards convincing someone to spend time with it. For creators, the harder task may be building a recognisable voice, a community and a reason for an audience to return. The article's music analogy points in the same direction: when recordings became cheap, live performances mattered more. This catch in the BookTok story is that Joseph also describes a shrinking pool of dedicated readers; a booming market can coexist with a narrower audience and rewards concentrated among a few successes and increased sales don't automatically imply healthier creative careers. I'd be wary of any forecast that counts the number of AI-made works and calls it a prediction for employment. Yet, we need to understand who earns the next pound of audience spending, and how many people can make a living from it. BookTok helped sell books by making them a social experience. Generating another million novels won't necessarily achieve the same thing.