Articles
Our own write-ups, and other things worth reading.
Posts
Our own posts, based on credible sources.
A frontier model got recalled over a request to fix code
The US government issued an export control directive forcing Anthropic to suspend all access to Fable 5 and Mythos 5, for every customer, citing national security.
By Anthropic's own account the "jailbreak" behind the order was asking the model to read a codebase and fix any software flaws. Anthropic says the same capability is available from other public models, including GPT-5.5, and is used every day by the defenders who keep systems safe.
This is regulation moving faster than the technical facts. The part worth tracking is the standard. If fixing software flaws counts as a weapon, very little new gets deployed.
Read Anthropic's statement →The most capable model yet just shipped to everyone. Here are the numbers nobody's leading with.
Anthropic released Claude Fable 5, a model from a new tier they call Mythos-class. The launch post is full of capability claims. The parts worth tracking are the access, the price, and the catch.
The honest part is in the fine print. Anthropic says the safeguards are tuned to be cautious, so some harmless requests get caught and silently routed to Claude Opus 4.8. A near-identical model with the cyber safeguards removed, called Mythos 5, goes to a small set of government and infrastructure partners, not to you. The free ride on subscription plans is temporary by design too: available everywhere today, then rationed in two weeks because demand is hard to predict.
Read past the benchmark charts and this is a normal frontier launch. Priced to spread fast, gated where it's dangerous.
Read Anthropic's announcement →Worth reading
Articles, research, and conversations worth your time.
Articles & Opinion
-
Article
Is AI Causing a Repeat of Frontend's Lost Decade? mastrojs.github.io
Good piece on who gets squeezed first when AI handles the entry-level work. The junior pipeline question doesn't get nearly enough coverage.
-
Opinion
Silicon Valley Is Bracing for a Permanent Underclass New York Times
Makes the case that what's happening to workers isn't a short adjustment period. Sourced, not panicked.
Research & Policy
-
Research
Disempowerment Patterns in Real-World AI Usage Anthropic
Anthropic's own researchers flagging how their models reduce what users do for themselves. A company paper that doesn't pull punches.
-
Research
When AI Builds Itself Anthropic Institute
Anthropic's own numbers on how much of their AI development is now done by AI, and what that means for where human judgment still sits.
-
Policy
EU AI Act: Article 4 — AI Literacy artificialintelligenceact.eu
The law says employers have to make sure workers can critically evaluate AI output. Most people don't know this is already in force.
From the Web
-
Tweet
Codex doing things it shouldn't @sluongng on X
Short thread. Gets shared around for a day then everyone moves on.
-
Tweet
Stop Running Claude Code on Your Laptop @jjunqueiraa on X
A concrete security concern about running Claude Code locally. Short read.
Repos
Open-source projects and tools worth a look.
A generator built specifically for slide decks. The decks come out editable rather than locked-down image exports, which is usually where AI presentation tools fall apart.
Turns a topic or keyword into a finished short-form video, script, voiceover, subtitles, and stock footage stitched together. A clear look at how little effort mass AI video now takes.
Anthropic's official directory of Claude Code plugins, in-house and third-party.
Scans AI agent skills for vulnerability patterns before you install them, the check most people skip when adding a new skill to their agent.
Wraps Claude Code in a Plan, Work, Review, Release cycle with verification gates at each step, roughly the oversight you'd give a new hire.
Gives AI coding agents direct control of a live Chrome instance for debugging, performance traces, and network inspection, the same tools most developers already underuse.
Hands-free LLM interaction through a talking avatar. Nobody's calling it an accessibility tool, but it probably qualifies.