Technology: Weekly Summary (November 24-30, 2025)

It felt like a week where every conversation in tech had two speeds: breakneck product launches and a quieter, stubborn worry about what all of this will mean once the excitement wears off. The blogs I skimmed and lingered on this week kept flipping between: shiny models and image toys, the money and metal that make them run, and people — coders, designers, parents — trying to live with the consequences. I would describe them as a messy, noisy neighbourhood where everyone’s shouting about the same thing, but each person is shouting with a different accent.

The model race and the messy middle

If you follow AI news, you already know the headlines. Anthropic shipped Claude Opus 4.5 and made people say “well, that’s something” in a dozen slightly different ways. Simon Willison dug into the model’s sizing, effort parameter, and the pain of actually judging new models against old ones. thezvi.wordpress.com and Ben Dickson were busy too, parsing Gemini 3 Pro, GPT-5.1 variants, and what it all means for coding and creativity. It’s like watching sprint heats at a world meet — fast lanes, lots of numbers, and everyone timing themselves against an invisible gold standard.

To me, it feels like a race where everyone brought different shoes. Clauses about latency, token windows, and coding benchmarks are the shoelaces that keep tripping people up. Charlie Guo and Conrad Gray catalog the toolkits and the new features (tool calling, programmatic hooks, image models that actually behave), while others point out the familiar Achilles’ heel: hallucinations, brittle reasoning, and a kind of inconsistency researchers now call "jaggedness." That word — jaggedness — turned up more than once. Helen Toner talks about it as an intrinsic pattern in current AI performance: pockets of brilliance next to sudden, embarrassing failures. It’s like buying a TV that shows 4K nature one minute and snow the next.

I’d say the most interesting sub-story is not which company wins a benchmark this week. It’s that the differences between models are getting small enough that evaluation itself is confusing. Simon Willison and others point out an "evaluation crisis" — teams can’t easily show what a new model can do that old ones couldn’t. That means marketing gets louder, and demos get sleeker. But practical questions — can it reliably automate a piece of work? will it save a company money? — remain fuzzy. It’s like being given a Swiss Army knife and then realizing half the tools are stuck.

There’s also a dizzying churn around image and visual tools. Google's Nano Banana Pro (part of Gemini 3 Pro’s image suite) is being framed by Jakob Nielsen as a kind of “ChatGPT moment” for visuals — instant infographics, comics, visual resumes. That’s neat and feels powerful. But then people remind each other that these things flood feeds with slightly-off visuals, and that mass-produced graphic content isn’t the same as thoughtful design.

Money, chips, and the backstage of AI

People keep returning to the same practical hinge: silicon and cash. Judy Lin 林昭儀 took a scalpel to the Musk/Tesla chip plans and why advanced packaging — CoWoS and the like — are not plug-and-play. In plain terms: making more chips is not a bakery where you just add ovens. There are specialized steps, chokepoints, and relationships that matter. It’s not glamorous but it’s the truth.

Then there’s the financial melodrama. HSBC-style projections and breathless estimations of how much OpenAI might have to raise make some posts read like apocalyptic budgeting exercises. The Independent Variable and others sketch scenarios where AI firms need billions and sometimes trillions to run the data-centre beasts. Will Lockett and a few commentators call it an "AI time bomb," arguing the current model of spending big on compute and hoping markets forgive the losses is shaky. It’s like buying a mansion and then realising the heating bill is for a small country.

Two government-sized moves complicate this: a U.S. Genesis Mission (a national AI program) and the general push for state-level involvement. Brian Fagioli covered the Genesis Mission with a mix of scepticism and curiosity. Gary Marcus wonders if some of these programs are thinly veiled bailouts. The political economy is getting noisy: subsidies, strategic purchases, and national priorities are shaping who gets to build scale.

Meanwhile, the hardware realities bite. Nicholas Wilt wrote about SRAM scaling and how some foundational parts of chips aren’t keeping pace with Moore’s Law. Martin Brinkmann and others point out that hundreds of millions of PCs are stuck on older Windows versions — real-world friction that matters to software rollouts and enterprise planning. These are the small, practical details that break big plans.

Design, devices, and the shape of everyday tech

Design and user experience keep surfacing as the place where theory meets habit. A few posts were quietly, nastily critical of recent UI moves. Jason Journals hated Apple’s new Liquid Glass look in iOS 26.1 for being pretty but illegible. Greg Morris breathed a sigh of relief that the long-rumoured iPhone Desktop Mode didn’t ship — he thinks trying to wedge touch-first UI into desktop metaphors would have been a mess. And then Jonny Evans has the tantalising news that Jony Ive and OpenAI are building a screen-free, smartphone-sized device that might land within two years. A screen-free OpenAI gadget — that’s a bold piece of design theatre.

These items together create an odd contrast. On one side, designers talk about simplicity and craft. On the other, engineers and product teams push features and lock people into ecosystems. Annie Vella writes about a software shift: deterministic systems are turning into non-deterministic ones because of LLMs. The old knobs and levers don’t always work. Engineers are moving from being coders to orchestrators, as Anup Jadhav put it. That phrase kept jumping up in different posts: orchestrator, steward, conductor — people who manage systems of services, LLMs, tools, and humans.

I’d say that phrase captures the era. It’s less about typing neat functions and more about setting up reliable choreography so that a dozen moving parts don’t step on each other. It’s like being the parking lot attendant in a busy festival: you don’t drive the cars, but you make sure nobody crashes.

Work, education, and the human stakes

A strange split runs through posts about work. Some people cheer automation, others are terrified. The job-by-job guide to AI evolution by Nate lays out practical shifts — here’s what changes in a role, here’s what stays. That kind of pragmatic mapping feels useful. But then you have pieces by Gary Marcus and Nick Heer warning of job churn, market bubbles, and social costs. The argument isn't abstract: layoffs, skill erosion, and the slow shifting of value are already happening.

Education also shows up. WHY EDIFY quoted Dr. Jean Twenge on how screens hurt learning. It’s a practical reminder: the tools that make information easy often make deep learning harder. And Dom Corriveau and others described small, domestic tech projects — syncing files, resurrecting older laptops, building multi-seat Linux setups. Those write-ups are gentler but they show a truth: people still learn by doing, by fiddling. Dorothy Vaughan’s story — retold by Imapenguin — is a tight, human example: she adapted to automation, learned FORTRAN, and taught her team. That’s the kind of lived strategy that matters more than hype.

There’s a cultural sting in many posts: people fantasise about AI bosses on Hacker News, as Jamie Lord covered, partly because it’s a joke and partly because it reveals workplace anxieties. Tech workers joke about replacing CEOs with code while building systems that threaten jobs. It’s a weird, bitter comedy. Meanwhile, mental health questions creep in, with writers noting addiction, rehab, and the human fallout of rapid change. Stories of family tech support and reinstalling Windows for a father — small, tender vignettes — keep the conversation grounded.

Friction, craft, and the case for doing stuff slowly

A surprising chorus this week loved friction. It showed up in essays like Adam Singer saying some friction is good, and pieces titled "frictions create purpose" and "Take your time." The claim is plain: if everything is frictionless, then you don’t learn by doing. The result is bland, homogeneous outputs and a cultural loss.

That’s echoed by folks who worry about AI doing the easy bits and leaving humans with the weird, less-lucrative work. Filip Hráček made a neat analogy: A.I. is sometimes like a printed birthday card to Paris. It speeds things up but doesn’t replace the craft of handwriting. I’d say this is true in design, writing, and code. Automation can tidy up the margins but it rarely gives you the deeper muscle memory of actually doing the work yourself.

There’s also a pushback against the “don’t do it yourself” culture. Sandro and others argue that junior devs are told not to experiment, preventing learning. That ties back to the earlier theme: engineers are becoming orchestrators, yes, but they also need the experience of building things — the gritty debugging and the broken deployments — otherwise orchestration becomes a myth.

Privacy, surveillance, and civic tech

The civic side of tech kept nudging in. Posts on ALPRs (license plate readers) by Fourth Amendment pointed at the creeping normality of surveillance tech in neighbourhoods. Ben Werdmuller argued the EFF must re-think priorities for a market full of private surveillance tools, while privacy product moves like Surfshark’s Multi IP and rotating IP features were covered by Brian Fagioli. It was a week when ‘privacy’ looked like both a market feature and a civil-rights fight.

There is a subtle tension: convenient features that promise privacy often trade off reliability. Surfshark’s rotating IP, for instance, is great for hiding your tracks but can break logins and banking. It’s the old balance: security versus usability. That balance shows up again and again — in Signal’s secure backups Michael J. Tsai, in debates about data siloing Dave Winer, and in people choosing self-hosted stacks Chris McLeod and the Default Apps folks.

On a related note, the EFF pieces and the calls to push back against aggressive AI deployments are part of the same chorus: technologies aren’t neutral. They reflect incentives, and right now a lot of incentives are about growth and scale rather than human dignity.

Small things, distro love, and the hobbyist heart

Every week there are small, kind pieces about the joy of tinkering. Zorin OS’s download surge as Windows 10 support ends — covered by John Lampard — is one. So is the Solus 4.8 release Brian Fagioli and 4MLinux hitting version 50.0. These are not glamorous, but they matter because people still want control.

There’s also a long, sentimental thread running through posts about Typewriters, Tekserve, ham radio, and old Macs. Evan Hahn and Michael J. Tsai and the Tekserve obituary bring a human texture to tech history. It’s like finding an old photo in the attic — small, warm, a reminder that culture and tech are braided.

And then there are practical pieces for engineers: prompt caching and paged attention guides from Sankalp, and detailed notes about how to force Google Calendar’s account picker by Ishan Das Sharma. These short, focused posts are the workhorses — often more useful than hype.

Politics, bubbles, and narrative battles

Bubble talk dominated a corner of the week. Multiple authors argued whether AI mania looks bigger than the dot-com bubble. James Wang, Derek Thompson, and Ed Zitron each pecked at different angles: valuations, debt, whether Nvidia’s dominance is sustainable, and whether government intervention is a bailout dressed as strategy. Downtown Josh Brown framed the contest as Google Cloud vs OpenAI, a proper horse race of capital and compute.

I’d say the recurring argument is this: hype is real, but so are the costs. People are betting on production functions that look nice on paper. But when the chips and power bills show up, the story gets less tidy. That’s when politics enters. When the state buys chips or bundles compute, private market dynamics shift overnight.

A few stray notes and small pleasures

• If you like gadgets and odd Easter eggs, Pierre Dandumont found a Christmas countdown hidden in Toast for Mac. It’s ridiculous and charming. Like a tin toy in the back of a shop window.
• Greg Morris is quietly relieved that Apple didn’t turn the iPhone into a desktop masquerade. There’s an appetite for clarity: either mobile or desktop, don’t half-do both.
• For those who hoard used Macs, Mike Rockwell thinks Intel Macs are still a great buy. Nostalgia meets pragmatism.

A lot of these posts are mild variations on three big stories: models keep getting better and stranger; building and running them is expensive and messy; and humans are trying to figure out whether to lean in, slow down, or resist. The week gave evidence for all three. Sometimes the evidence is a data point, sometimes an anecdote about a dad and an old laptop, sometimes a technical tear-down of memory chips.

If you want to chase threads, start with the model duels — Anthropic’s Opus 4.5 and Google’s Gemini 3 Pro — and then wander into the hardware essays about chips and packaging. From there the sidewalks split: you can take the civic lane (privacy, EFF, surveillance), the craft lane (friction, doing the work), or the hobbyist alley (distros, ham radio, used Macs). They all meet again at the corner where people try to build a useful life in messy times.

There’s a lot more in the linked pieces. If one of those tangents tickles your head — the weirdness of hallucinations, the micro-dramas of chip supply, the strange comfort of old tech — it’s worth opening the original posts. They’re the long versions of these little maps, and they’ll take you down rabbit holes worth visiting.