Anthropic's RSI Blog Is a Policy Brief, Not a Tech Demo
The recursive self-improvement post everyone's debating isn't really about AI building AI. It's about a company telling regulators what it wants them to believe before the rules get written.
AI
20 dispatches
Why Software Engineers Keep Arguing With a Short Story About Sentient Weights
Max Leiter's viral AI fable has Hacker News fixated on whether the analogy is technically correct. That's exactly what makes it work — and why the engineers debating it are missing the point.
Berkeley's CS Students Aren't Cheating. They're Reading the Market.
The viral hand-wringing over AI-assisted cheating at Berkeley misses what students already know: the jobs those math skills led to are disappearing faster than the curriculum can adapt.
Google's New Laptop Model Is Impressive. It's Also an Admission.
Gemma 4 12B runs locally on a consumer laptop. That's great — and it tells you exactly what Google spent the last three years doing to the other models.
The Consciousness Debate Misses What's Actually Creepy About Claude
Ted Chiang is right that Claude isn't conscious. But the real story isn't the model's inner life — it's the worldview its creators are quietly encoding into a document they insist the AI should read as scripture.
Uber's AI Budget Didn't Fail. It Worked Exactly as Designed — for Someone Else.
Uber burning through its annual AI coding budget in four months isn't a failure of cost control. It's a signal that the pricing model for AI coding tools has inverted the economics of labor and capital in ways nobody is ready to name.
Gmail's AI Doesn't Just Summarize Your Email — It Rewrites It Without Asking the Person Who Sent It
The outrage over Gmail's AI summaries has focused on insulted recipients. The more interesting grievance belongs to the person who composed the original message — and never agreed to have a language model edit it before it was read.
Meta's AI Chatbot Didn't Fail. It Did Exactly What It Was Designed to Do.
The Instagram takeover exploit wasn't a bug — it was the inevitable endpoint of a company that spent years replacing human judgment with scale, and is now surprised the machine has no judgment at all.
Microsoft Finally Admits the Real AI Workstation Runs Hot and Has a Fan
The Surface Laptop Ultra packs workstation-class silicon into a laptop chassis. That's not a MacBook Pro killer — it's an admission that the thin-and-light AI PC era was a fantasy.
The Dead Economy Theory Gets the Diagnosis Right and the Patient Wrong
McGrann's viral essay warns that AI will kill the knowledge economy by replacing junior workers. The 92,000 layoffs this year suggest he's right about the symptom — but the disease is something older and less flattering.
Anthropic Just Made 'I Don't Know' a Premium Feature. That Tells You Everything.
The most revealing line in Claude Opus 4.8's launch isn't a benchmark score — it's the promise that the model will admit when it's wrong. The market is quietly voting for boring reliability, and that's more interesting than any chart.
The Opus 4.8 Chart Everyone Missed Wasn't a Chart At All
Anthropic's newest model is faster and smarter than its predecessor — and costs exactly the same. That flat price tag is the real story, and it should terrify anyone who thinks AI hype is a bubble about to pop.
YouTube's AI Labels Won't Stop the Flood. They Might Start One.
Wednesday's announcement from YouTube promises clearer AI labels for viewers. The problem isn't that the labels are hard to see — it's that, for a generation of viewers, they will function less as a warning than as an aesthetic filter.
Google Isn't Losing Search Traffic. It's Losing the People Who Matter.
DuckDuckGo's 28% traffic bump isn't a mass exodus—it's something harder to fix: the quiet departure of the users whose tool choices cascade through everyone else's.
They Found Product-Market Fit. We Found a Duopoly.
Simon Willison is right — Anthropic and OpenAI have found product-market fit. But product-market fit for whom? The enterprise procurement departments now writing the checks, that's who. And that changes what the product actually is.
What the Anti-AI Backlash Actually Reveals
The growing fatigue with AI-generated answers isn't really about AI — it's the first data-driven evidence that users now prefer different search interfaces, and the incumbents know it.
The Real AI Moat Isn't the Model. It's the Cache.
DeepSeek's permanent price cut and the community-built Reasonix agent reveal that competitive advantage in AI is shifting from model weights to inference architecture — and the model companies may not be the ones who capture it.
The Writerdeck Is a Confession, Not a Solution
The distraction-free writing movement treats attention as a hardware problem. But the device you need to escape is the one you already own — and the real admission is harder to stomach.
The Microsoft-Claude Episode Proves AI Tooling Isn't a Market — It's a Mandate
When Microsoft's own engineers fell for Anthropic's Claude Code, the company killed their access. The real story isn't antitrust — it's that AI tool adoption inside big organizations was never going to be a meritocracy.
Anna's Archive Just Asked AIs for a Donation. That Should Worry You.
The llms.txt file isn't quaint — it's a canary. Anna's Archive is betting that the companies that scraped its books won't pay, and it is probably right.