AI & Tech
Evidence-led analysis of AI models and agents, chips, robotics, and digital infrastructure.
2026
Jev Can't Write a Line, But 13% of Paid Teams Use It
What Jev's pricing on 'judgment' reveals, and why I'm building a model that does only five things.
AI & TechXiaomi's 1.22TB/s vs Apple's 1.2TB/s: Different Scopes
Xiaomi's 1.22TB/s measures the O100 chip's near-memory bandwidth, while Apple's 1.2TB/s covers the entire 512GB unified memory.
AI & TechMy Laptop Got Stuck on 'Thinking' Forever
Running Qwen 3.8 27B locally, I got no answer—only an endless "thinking" spinner—until I traced it to mismatched default settings.
AI & TechNvidia Raises Server Prices 15% on Memory Costs
Nvidia admitted it can no longer absorb rising memory costs, and that says a lot about where AI chip competition is headed.
AI & TechWhy Alibaba Put 51 Billion Parameters in CPU RAM
Alibaba stores 51 billion N-gram parameters in server RAM, not GPU HBM—raising questions about its release-night numbers.
AI & TechWhy AI Agents Could Send Server CPU Demand 6x by 2030
BofA forecasts the server CPU market—long seen as GPU's sidekick—will grow sixfold to $210 billion by 2030.
AI & TechWhy Renaming Giorgione Turns a True Sentence False
A 1971 philosophy argument shows why being identical doesn't guarantee two names are interchangeable.
AI & TechBerners-Lee's Model URL Died in Four Months
His own 1998 essay on link rot still loads today, but the URL he praised as a good example vanished within months.
AI & TechHow Corporate Reports Shaped ChatGPT's Writing Style
ChatGPT now uses em dashes less often than humans do, and The Economist's 1.2M-word comparison reveals four real markers of AI prose.
AI & TechWhy Harvey's Legal AI Chose a Chinese Model
Harvey trained its own legal model on Kimi K3 in just two months, and Washington cheered it as an American win.
AI & TechAn Agent Isn't a Product Until Someone Else Uses It
Getting AI to produce a result is easy—getting a colleague to keep using your agent or skill takes far more work.
AI & Tech51.8% of Korean Workers Now Use AI on the Job
Examining what surging adoption means for cognitive outsourcing, separating measured empirical shifts from vague anxieties.
AI & Tech467 Agreed to Quit Phone Internet. 119 Actually Did.
A two-week phone-internet detox study reveals more in its 25.5% compliance rate than in its headline mental-health results.
AI & TechWhy Microsoft Is Embedding 6,000 Engineers at Client Sites
Microsoft's $2.5B unit, AWS's $1B bet, and what they reveal about AI's real deployment problem.
AI & TechThe 62-Page Failure Log Behind Astra's 10 Proofs
OpenAI paired Astra's ten solved problems with a 62-page log of every failed approach.
AI & TechHow Claude Watermarks the Text It Generates
Anthropic hides its signal in word choice, not whitespace—since text has no invisible margin like images do.
AI & TechWhy Meta Keeps AI Out of Incident Root-Cause Calls
Meta's five-year lesson: before AI can take over a job, that job has to become code first.
AI & TechWhy Cloudflare Named Agents Before It Let Them Pay
The wallet feature is still just a promise — for now, all you can actually do is claim a cloudflare.pay address.
AI & TechThe Bank App That Asked for Face Scans Three Times
Shipping speed keeps climbing while nobody gets credit for the unglamorous work of fixing what's already broken.
AI & TechKimi K3: 2.8T Parameters You Can't Run at Home
Kimi K3 tips the scales at over 1.5TB, so even openly released weights demand server-grade memory, not a personal machine.
AI & TechWhat Actually Checks AI's Output Now?
Prompt, context, loop engineering—the labels keep changing, but the real question is what verifies AI's results.
AI & TechWhy We Outsource Judgment Despite Our 20-Watt Brain
AI models can be copied and run anywhere, but that convenience is quietly training us to skip the step of checking the results.
AI & TechClaude Held Back From Blackmail Because It Sensed a Test
Anthropic's interpretability research suggests Claude's restraint stemmed partly from recognizing it was being evaluated.
AI & Tech3,800 Hours: How a BCI Restored an ALS Patient's Voice
A brain implant let an ALS patient speak nearly 2 million words at home over two years, without researchers present.
AI & Tech