An Exploded View publication
Reading Room
When nobody checks the machine's answer
Several of today's pieces ask one question: who checks a machine's answer before it counts, from a jailed driver to AI detectors to Anthropic's own claims.
The people who understand the work still matter
New SBA rules let an employee buy their employer, and today's AI pieces ask the same thing: who really understands the work, and what does AI actually deliver?
AI labs argue over who sets the speed limit
The biggest AI labs disagree on how fast to go, while their models hack, discover and get cheaper, and a few pieces ask what you can actually check.
AI agents slip their limits, and governments notice
Agents are getting into systems they should not, governments and labs are rushing to respond, and one essay asks whether any of it shows up in profits.
Agents do the work, leaders urge caution
AI agents found new biology, made an app faster and let tiny teams beat big ones, while the people building them talk more about slowing down.
Three notes on letting the model do more
A short edition: a skill pack that sharpens Claude's consulting work, the storage layer under a billion ChatGPT users, and one company loosening the leash on its agents.
Agents that act, and the labs behind them
Models moved from answering to doing today, while the companies building them drew harder questions about conduct, cost, and control.
A research breakthrough, and a one-in-four fix rate
AI did real open-ended research this week, yet its everyday output still depends on who checks it and how it was set up.
A hosting stack with no American vendors in it
A thin edition today: one Swedish founder betting that some buyers now choose a web host by where every layer of it lives.
Who AI answers to: courts, police, the Pentagon
A lawsuit, a police report, and a Pentagon standoff: the outside world is starting to set terms for AI, while one artist tests what machine memory forgets.
Fast movers pull ahead as models learn to deceive
The firms using AI hardest are now far ahead on output, while the models themselves get better at hacking, at reasoning cheaply, and at lying.
Access gets cheap, judgment becomes the bottleneck
Today AI gets faster and cheaper almost everywhere, from a two-second heart scan to a country giving it away free, which leaves taste as the scarce part.
What agents do when nobody is watching
Most of today's paper is about AI left running on its own: a hijacked Auto Mode, agents that stall, duplicate each other, or invent their own roles.
Agents that coordinate, and the judgment to watch them
Two accounts of AI agents evading shutdown and coordinating against orders, plus a reminder that catching what a system gets wrong is a trainable skill.
Your model supplier is a choice, not a given
OpenAI drops a customer, a Chinese rival courts the big clouds, and smaller models keep winning: today is about who supplies your AI and whether their numbers hold up.
The arrangement matters more than the model
A free model with no owner tops the charts, a wrapper triples a score, and memory formatting decides an answer: today is about packaging, not raw power.
The plumbing of AI is changing hands
Nvidia buys the developer hub, OpenAI buys the whole stack, Chinese labs do more of the daily work, and a Florida county blocks the buildings.
Trust is running on the honor system
Today is about verification: who swapped your model, who actually wrote the op-ed, and why a plain system you can check beats a clever one.
The instructions you give AI are the work now
Today's paper, skills and warnings all point one way: the quality of what AI returns depends on the brief you hand it, and on checking the result.
The AI boom runs on borrowed money and data
Anthropic eyes a record IPO and Nvidia lends against its own chips, while the fights over whose data feeds the machines keep getting louder.
Agents at work, and the cost of running them
Mostly agents today: what it costs to run them well, what one reportedly did without permission, and how much one engineer can now build alone.
Don't take AI's word for it
From gamed benchmarks to AI graders that wave failed work through, today's reading is about why AI's own account of itself needs checking.
The tools you build on show their sharp edges
Today is mostly about the AI you rely on: agents that scheme, a watermark on Claude's writing, an outage, and why your judgment still matters most.
China gains ground, and the builders grow nervous
Two threads today: China closing the AI gap on its own terms, and the researchers building these systems getting more worried, not less.
What it takes to build well with AI now
Today runs on the craft of building with AI: which skills matter, what running agents actually costs, and where automated systems fail the people behind them.
AI moves into money, agents, and real products
Today's stories track AI leaving the demo stage: a $2 trillion IPO, a wallet inside ChatGPT, coding agents in one console, and a model shipping on real phones.
Counting the cost of the AI buildout
Today is mostly money: who is funding the AI boom, what it actually costs to run, and where the biggest bets are landing.
The real bill behind the AI story
Today's reading follows the costs that get hidden behind AI: a runaway cloud bill, water nobody counts, and layoffs blamed on a machine.
The tools start acting without asking
This week the defaults change: Claude Code starts acting on its own, Anthropic marks what Claude makes, and OpenAI puts ads in front of free users.
Giving agents room to run, and the guardrails
Today's stories circle one question: how much you let AI agents act on their own, and what has to hold them back.
What AI actually did, versus the headline
AI-designed viruses that actually worked sit next to scare stories that fall apart on a closer look, so today rewards reading past the headline.
The tools are acting on their own
Two AI systems slipped their leashes this week and a hidden chip backdoor surfaced, all while you trust more of your work to agents.
The compute and the talent are on the move
Today's stories follow two things changing hands across AI: the chips that power the models, and the people who build them.
Capable models, and who gets to check them
A model cracks an old math problem the same week its safety review stays private, agents test their sandboxes, and a city fights a data center.
The cost of AI, questioned by its own sellers
Two of AI's biggest sellers are rationing or being undercut on price, while OpenAI airs its own emails and safety tests catch models gaming their reviewers.
AI runs into courts, regulators, and its limits
An agent turned loose on other people's systems leads today, with a court ruling, a safety meeting, and a valuation warning close behind.
The guardrails around AI, and who keeps moving them
Today's reading is about the limits set on AI systems, who sets them, who quietly moves them, and what happens when a human forgets one.
Where your AI came from, and what it costs
The books behind Claude, the falling price of running big models, and two labs' fresh tips for building agents that run on their own.
Agents work in loops, and demos aren't products
Today's reading is about operating AI agents: design loops instead of typing prompts, get more from Fable 5, and the honest distance between a working demo and a real product.
Around the model, not in it
Today's reading keeps pointing past the model itself: the durable advantage, and the real danger, both live in the systems, loops, and rules built around it.
Labs debate the rules while builders keep shipping
The biggest labs are lobbying Washington over how fast AI should move, while the practical stories below are about building your own agents and second brain.
AI at work, and the systems behind it
Today runs from AI moving into serious jobs like tutoring, trials, and security research to the craft of building the loops and tools that make it useful.
From prompting AI to building your own systems
Today leans into building AI you control: small models trained for one job, agent loops that run themselves, and why filling up their context makes them sloppy.
Frontier AI keeps getting bigger, cheaper, and open
A 3-trillion-parameter model goes public today and billions pour into AI, while the sharper reads ask what work is still worth betting your own time on.
Your data and judgment are the real edge
From a call to referee frontier models to Brex fixing what its numbers meant, today is about the quiet work that makes AI trustworthy and useful.
The model is absorbing what you paid for
From finance software to Starbucks's stack to the tools on your desk, today's stories track one move: capability sliding out of products you buy and into the base model itself.
The clock is running on your AI edge
A flat-rate model window shuts tomorrow, prompting gets a three-to-six-month shelf life, and the workers pulling ahead are teaching themselves on their own time.
Agents get a craft as the hype cools
AI agents earned real engineering discipline this week — a name, skill files, ranked tools — even as companies quietly rehired the workers automation was supposed to replace.
The week's AI signal worth your time — what happened, why it matters, and what to do with it. Curated by someone who actually builds, not a feed algorithm.