I've meet Alvaro more than a decade ago, when I was living in Zurich and he's really great. Also, we did an extensive chat with him for the Tech Talks London, 5 years ago. He could've sold a grand story about models secretly doing maths in vector cathedrals, but instead, it lands on something narrower and more useful: activations can supply calculator arguments under a no-parser boundary. Less magic. More engineering. Beautiful! decoded variable isn’t an API, and Alvaro gets right that interpretability claims need a ladder: readable, causal, writable, deployable. Those rungs are too often mashed together in the usual “we found a neuron for X” theatre. Here, the failed residual-JIT dream is not brushed away; it becomes the point. Llama behaves like Llama. Qwen’s failed transfer is the cold shower. If internal geometry doesn’t line up across models, then this isn’t a portable abstraction yet; it’s model surgery with a good lab notebook. Useful, but not casual. The architecture lesson is punchy: transformers may encode number-like structure without owning a stable calculator inside. That’s weird and valuable. It’s not sorcery. As always, excellent claim discipline, promising mechanism, no free lunch.

Windows keeps repeating this pattern: hide an ugly constraint, expose just enough abstraction, then make developers live with the leak for years. Michael's article is valuable because it doesn’t treat Win16 memory management as quaint trivia. It shows the machinery. Gears, springs, suspicious smoke. The strangest detail is the function prolog that appears to waste instructions until the Windows loader quietly patches it into something essential. That’s marvellous and slightly cursed. The code is harmless until it isn’t, and the operating system depends on a choreography between compiler switches, NE exports, stack frames, segment handles, and programmer obedience. Elegant? In places. Forgiving? Absolutely not! The underplayed point is organisational, not technical. A platform that needs secret-ish compiler switches, exported window procedures, lock counts, and stress tools like Shaker is also creating an ecosystem tax. Every compiler vendor, library author, and application developer has to internalise the ritual. OS/2 looks cleaner in comparison because the 286 hardware carried more of the burden. Yet Windows carried the market...

David gets the first-order argument right: the data doesn’t yet match the panic. The AI jobs debate has become a fog machine for everyone’s favourite theory, whilst founders see inevitability, executives see margin expansion and pundits see the end of work before lunch. The counterintuitive claim is the one worth pinning to the wall: US occupations most exposed to AI apparently have lower unemployment than less-exposed ones. That doesn’t prove safety, but it punctures the lazier version of the apocalypse story. If AI were already chewing through white-collar work at scale, we’d expect some clearer labour-market signal. The article underplays the messier middle; labour markets are lagging instruments. A graduate who never gets an entry-level analyst job won’t show up as a displaced analyst. A contractor whose copywriting work quietly dries up may not look like a knowledge-worker exodus. A software team that hires three juniors instead of ten is a real shift, even if nobody gets fired. Markets love narratives, and “AI did it” is becoming a convenient one. It flatters boards, excuses restructuring, and gives investors a story cleaner than over-hiring or weak demand. The danger isn’t just hysteria. It’s bad diagnosis. If leaders believe AI has already solved the labour problem, they’ll cut before they redesign work. If policymakers dismiss all concern as panic, they’ll miss the quieter transition.

The most revealing detail is not Claude Mythos Preview itself; it’s where it’s going. 150 partners across 15 countries already looked less like a cautious lab pilot and more like a distribution strategy with a safety label on the box. Yet, our Hitachi announcement gives that strategy sharper consequences: this is now being pointed at software and products for social infrastructure, including energy. This is important because critical infrastructure is where AI-assisted vulnerability discovery stops being a model demo and starts becoming an operational risk question. Security teams need leverage. Modern codebases are too large, too old and too interconnected for heroic manual review to carry the load alone. If a model can help defenders find critical vulnerabilities before attackers do, sitting on that capability forever would be hard to justify. But some of the hard questions remain mostly unanswered. Who gets access? Who audits the results? What classes of vulnerability can Claude Mythos Preview reliably find? What is the false-positive rate? What happens when it finds more serious issues than teams can responsibly triage and patch? Detection is the glamorous bit. Remediation is where optimism goes to queu

The 10th of June is the day of Portugal and Portuguese spoken countries, so this article comes in handy :) If you think about it, Portugal’s advantage was not size; it was position, desperation and a working theory of chokepoints. A small, poor kingdom with a long Atlantic edge learned how to turn geography into leverage, then converted leverage into monopoly ambition. Sure, the East didn’t need discovering. Europe needed an exit. That punctures the old heroic map-room version of the story and replaces it with something sharper: globalisation began, in part, as a workaround for European confinement. Roger gets right that the Portuguese model was networked rather than territorial. Forts, bases, sea lanes, trading cycles, intelligence, artillery, chokepoints. Not empire as a coloured blob on a school atlas, but empire as infrastructure. That’s the useful lens for readers who think about platforms, logistics or supply chains today. Control the nodes and you can shape the market without owning the whole map. The article does not ignore slavery, violence or disease; the darker ledger is clearly there. But the prose still gives the system a certain kinetic glamour. Crops move, bullion flows, images circulate, cultures hybridise. Then comes Elmina. Then half die in transit. Vasco da Gama didn’t build a benign world of exchange. He helped open a world where trade, coercion and information moved together.