A friend of mine, let’s call her Jane, runs a marketing consultancy. Fifteen years in, small team, strong client list, had used Claude for chat only. I offered to spend an hour showing her what’s possible now we’re well beyond that. Wednesday night I realised winging a 75-minute demo next morning would burn too much time on navigation and too little on value. I opened Claude Cowork, uploaded a brief I’d quickly hacked together from emails and went on with my evening.
Aaron Haspel published britannica11.org last week. It is a complete, searchable, cross-referenced digital edition of the Encyclopædia Britannica’s celebrated 11th edition (1910 to 1911). One person built it. Project Gutenberg and Wikisource have been chipping away at the same artefact for over twenty years and are still not done.
The credits page is the giveaway. “Thanks most of all to Anthropic and Claude Code Opus, which did nearly all the heavy lifting, and to OpenAI and GPT Codex, which drafted the specification.” Haspel is a New York based aphorist and essayist who also writes code. That blend is the new shape of solo scholarship.
We’ve spent decades framing distractibility as a deficit. A disorder. Something to medicate and manage. Anne-Laure Le Cunff, a neuroscientist at King’s College London, thinks we have it backwards.
In a compelling new essay for Aeon, Le Cunff makes the case that what we now label ADHD was once an evolutionary advantage. The hypercurious mind, the one that can’t stop scanning the horizon, that gets bored with routine, wasn’t broken. It was built for exploration.
“Human attention did not evolve in an environment saturated with infinite information and algorithmically optimised distraction. For most of our history, novelty was relatively rare and often meaningful; today, exposure to novelty is constant and difficult to escape.”
Daniel Vaughan ran Gemma 4 as a local model inside OpenAI’s Codex CLI on both a MacBook Pro and a Dell GB10 with NVIDIA Blackwell. The results are worth your time. The headline number is striking. Google’s Gemma 4 hit 86.4% tool-calling accuracy versus Gemma 3’s 6.6%. That’s not incremental improvement. That’s a generational leap in what a local model can do inside an agentic coding workflow. But the details tell a more familiar story.
tl;dr: Arthur Soares ran a 14-agent household AI fleet on Claude, got forced off it overnight by an Anthropic policy change, spent a full day migrating to GPT-5.4, and wrote up everything he learned. The short version: Claude and GPT-5.4 require fundamentally different prompting styles. A rigorous 8-model benchmark shows Claude Sonnet 4.6 still leads at 92% overall, but three Chinese open-weight models (GLM 5.1, Qwen 3.5, Kimi K2.5) are right behind GPT-5.
They’re home.
At 8:07 p.m. EDT on Friday, the Orion capsule carrying four astronauts splashed down in the Pacific Ocean off San Diego, completing a ten-day journey of nearly 1.1 million kilometres. Artemis II is the first crewed flight beyond low Earth orbit since Apollo 17 in 1972. More than half a century between drinks.
Caleb Fenton spent last month writing 20,000 lines of code in a language he doesn’t know. He used AI agents to write it, and his verdict: it was easier than using the language he’d relied on for a decade.
That’s a sentence worth pondering.
Many developers have gravitated toward Python because it is forgiving. You can be loose, approximate and casual. Python mostly lets things slide and figures it out. That philosophy made sense when humans were doing the coding. Humans get frustrated. Humans lose patience. Humans quit and find a different approach when a tool keeps telling them they’re wrong.
Something wasn’t right with my iOS app. I couldn’t pinpoint it. The screens worked, the features functioned, but the whole thing felt slightly off. Like a suit that fits everywhere except the shoulders.
So I did what any reasonable person does in 2026. I assembled a panel of AI judges.
I took screenshots of the current state, pulled up the original product spec and fed everything independently to three models: Claude AI (Sonnet 4.6), ChatGPT 5.3 and Google Gemini 3. Each got the same brief. Review the spec. Review the current design. Tell me what’s wrong and what you’d do differently.
There is now a button. A literal one. Two ex-Apple engineers who worked on the Vision Pro have built a brushed-aluminium puck that looks like an iPod Shuffle, costs US$179, ships in December, and exists for one reason: you press it, and a chatbot answers.
It is called Button, it is in Y Combinator’s Winter 2026 batch, and Wired’s Boone Ashworth got the demo.
The founders are Chris Nolet and Ryan Burgoyne. The pitch has two pillars. Privacy: it only listens when you push it, unlike the always-on Friend pendant or the Bee. Speed: replies inside a second, unlike the Humane AI Pin, which took so long to answer it died of embarrassment.
Saw this thread from @paoloanzn doing the rounds and it nails something I’ve been chewing on. vibecoder asks claude code to build a chat app, gets a working prototype in 20 minutes, immediately tweets “just killed slack and discord”… brother you don’t even know what a distributed system is. you don’t know what database replication means. you have no idea how websocket connections behave at scale or what happens when 50k people are online at once and someone’s message needs to show up in 200ms across 3 continents