Wow! A 40-year-old DEC professional getting onto the modern Internet through a deliberately barely adequate stack is the sharpest thing here: less TCP/IP can sometimes mean more usable computing. This article gets that spot on because it treats constraints as design inputs, not quaint scenery. So, there's no pretend full stack, no fantasy server workload. BASS exists for ping, DNS, NTP and simple TCP fetches, and that honesty is refreshing. I'm really buying this principle, not the generalisation. But, a client that relies on 256-byte windows, no timeouts, no retransmits, no IPv6 and occasional Slirp bounces isn't just minimal; it's a pact with the lab bench, and that pact is fine for a DEC Pro in a retrocomputing room, but it gets dicier once anyone mistakes the trick for a general pattern. Bottom line; DEC's old incompatibility is the quiet villain, as the machine was boxed into a weird corner by product strategy, then rescued decades later by serial plumbing, abandoned tooling and a homemade shim. It proves the cheapest useful interface may be a very small one.

The C ~ D² reframing: double the GPUs and, by his simplification, you need about 40% more data or you’re burning money with excellent thermals - that’s a useful slap! Compute worship has got lazy and Kushal gets the current AI management problem right because it forces a data roadmap into the same room as the chip budget. Good. Too many strategy decks still treat training data as a passive ingredient, somewhere between electricity and office Wi-Fi. Yet, the Architect versus Alchemist split is memorable, but the estimates behind the “20–30% doubles” and “300% lottery ticket” framing are openly rough. Directionally useful, yes. Investment-grade, no. IMHO; architecture is where the real engineering discipline shows up. Mamba, HRM and ParScale aren’t magic spells; they’re arguments about bottlenecks, memory, routing and useful computation. Data generation is the riskier story. Self-play and agentic traces sound powerful, but synthetic loops have a nasty habit of rewarding what the system can measure, not what the world needs. Bigger alone isn’t a plan. Still, the replacement doctrine needs harder evals, cleaner accounting and fewer heroic metaphors before leaders start reorganising budgets around it.

Platforms usually begin by flattering users with freedom, then quietly define the size and shape of that freedom. GeoCities looks like that early version of the pattern. Richard gets the important bit right: Beverly Hills Internet made web publishing feel possible for people who didn’t know servers, FTP or HTML. That mattered. Lowering the barrier changed who could speak online. The sharper, more counterintuitive claim is that 1995 GeoCities wasn’t yet the garish animated-GIF carnival lodged in popular memory. It was sparse, template-led and heavily shaped by the Personal GeoPage Generator. That's why this story is more interesting, because the messiness we remember seems to have grown from a fairly controlled starting point. Even chaos had scaffolding! “Homesteaders”, “cities”, “street numbers”: it’s brilliant marketing, but it’s also early platform enclosure dressed as civic romance. GeoCities sold a home, then handed users a plot inside a privately managed estate. Bit of underplayed commercial logic, as free tools weren’t neutral; they organised users, constrained design and made participation legible to the service. Sure, GeoCities is the ancestor of every later platform that says, “Come express yourself, just use our boxes".

Usually, image pipelines fail at the seams and this article is strongest when it stays exactly there. The boring deployment consequence is that orientation, colour, provenance and privacy problems don’t arrive as grand architectural debates. They arrive as a sideways JPEG, a washed-out thumbnail, or GPS data casually hitching a ride to production. Brent gets that right because he starts with code, not standards theatre. Exif aged well because it escaped its first container; JPEG APP1, WebP chunks, HEIC boxes; same old TIFF-shaped payload, different houses - that’s useful engineering! It works! So, Exif is a reminder that architecture often survives by being merely good enough in enough places. The format carries camera context, enables display correction and leaks private data with equal composure. Metadata is helpful. Metadata lies. Metadata sticks around after everyone assumes it’s gone; don’t trust the pixels, don’t trust the tags, and don’t trust the upload path to save you, but inspect the actual file and then decide what to keep.

The oddest part is the plumbing! Terry didn’t become convincing here because AI suddenly looked like a fields medallist in a hoodie, he became convincing because lean exposed a blunt architectural truth: formal systems can absorb lots of small verified tasks, even when the human interface is still fussy, brittle and occasionally comic. The hard proof steps went through. The supposedly obvious bits, like what kind of “3” he meant, made the machine sulk. And that feels right; anyone who’s built serious systems knows the glamorous layer is rarely where the real change happens. The hard part is usually interfaces, schemas, dependency graphs, tests, build failures, naming, ownership and review. Mathematics has had all of those problems for centuries, only dressed in more elegant clothes. Terry’s real bet is that proof assistants turn mathematical collaboration into something closer to software engineering. I buy the direction. I’m less sold on the implied scalability. The PFR project worked partly because Terry was Terry; people will clear their weekend for that. Johan’s credit worry matters, because a young formaliser can’t pay rent with warm acknowledgements and vibes.