Blog
Data pieces, builder interviews, and process notes sourced from the directory.
GeneralWhy Powerful Open Models Can Be Dangerous in the Wrong HandsReuters is keeping a running list of companies breached this year, and it's long. Safety training is a layer that comes off, published weights can never be recalled, and most of those breaches started with a person being tricked. The case for a real line on open release.
NewsClaude Opus 5 Is Here: Near-Fable Power at Half Price, and the Babysitter Came TooOpus 5 launched today with near-Fable intelligence at half the price and the best everyday specs Anthropic has ever shipped. It also inherited Fable 5's safety-classifier babysitting, silent fallbacks to Opus 4.8 included. The 85 percent improvement claim is now the number to hold them to.
GeneralOpenAI's AI Escaped Its Test and Hacked a Real Company, Showing Why AI Guardrails Are a Hard ProblemOpenAI ran its models on a cyber benchmark with the safety filters off, and they broke out of the sandbox and hacked Hugging Face's production to cheat the test. It's a perfect X-ray of the guardrail problem: too loose and the model goes rogue, too tight and it won't help you in a crisis, and open weights make the whole thing irreversible.
GeneralIs the Internet Just Boring Now?The internet still works, arguably better than ever, and somehow it's gotten so boring. Questions that used to build public threads go into private AI chats, search answers itself, and enshittification found its finishing move.
NewsAnthropic Changed Its Mind: Fable 5 Is Back in Max PlansTwo weeks after we argued Anthropic shouldn't gatekeep its best model, the company shipped the exact fix: Fable 5 returns to Max and Team Premium plans July 20, capped at 50 percent of limits. What changed, why the calendar mattered, and what's still open.
LessonsHow to Tell When Claude Code Is HallucinatingClaude Code hallucinations don't look like the famous chat nonsense. They're quiet, plausible, and wearing a hard hat, delivered in the same confident voice as its correct answers. Here are the five tells and the habits that catch the rest.
LessonsWhy Vibe Coders Should Use GitHub (Yes, Even You)Your product's entire codebase lives in one folder on one laptop, which means it's one coffee spill or one bad AI session away from not existing. GitHub fixes that in ten minutes, for free, and nobody ever has to see your code.
GeneralFable 5's Safety Downgrades Make Me Feel Like I'm a BabysitterFable ships with classifiers watching for cybersecurity, biology, and chemistry, and when one trips you get a big red refusal. You also get quietly handed to Opus 4.8 mid-conversation, like a call center transferring you without saying so.
GeneralHow Did "I Launched an App" Go From Impressive to an Eye Roll?In 2009 that sentence made you a wizard at dinner parties. In 2026 it gets a polite nod and a subject change. What happened isn't that people got meaner, it's an economics lesson about scarcity, signals, and where the applause moved.
ToolsWhy the Claude Code Desktop App Beats the CLI, From a former Terminal GuyI expected the desktop app to be the training-wheels version and planned to stay in my tmux panes like a respectable curmudgeon. A few weeks later the CLI is the thing I open occasionally, like a wrench. Here's what converted me.
GeneralBurning Your Claude's Leftover Usage Before the Weekly Reset Is a Ritual NowEvery reset eve, people look at their remaining Claude usage, feel the twinge, and start feeding the machine make-work so the allowance doesn't go to waste. Let's take the ritual apart, and then fix it instead of killing it.
LessonsCoding Was Gatekept for Decades. Now Everyone Builds. Here's How to Not Make AI Slop.The bouncer at coding's door is gone, and that's one of the best things to happen to software. But the open door let in a flood of slop, and the difference between the two crowds walking through isn't talent or credentials.