Decompiling a program was always possible; what protected it was that doing so cost more than rewriting it. Agents have collapsed that cost, and the question is no longer whether compiled code can be read, but what a big company, or an indie studio, has left once it can no longer hide inside it.

Anthropic and OpenAI shipped their cheaper models 89 minutes apart, and it is not the first time or the tenth. The industry that insists it cannot wait has spent three years proving it waits very well: right up until the rival calls a press briefing.

We have spent three years politely asking an essay-writing machine to answer in JSON and nothing else. TypeSafe AI has shipped a model that does not write, and the industry greeted it as a new paradigm. It is closer to an invoice.

Beijing took 48 hours to accuse Dario Amodei of hiding, behind safety language, a plan to strangle Chinese technology. Three years of milestones show the brake Beijing actually enforces isn't aimed at Washington — it's aimed inward.

A researcher resigned warning of extinction risk, another followed two days later, and their own CEO called for slowing down AI the same day Sam Altman and Elon Musk joined him. There was a letter just like it in 2023, and the man who signed it loudest had incorporated his rival company thirteen days earlier.

OpenAI spent millions of dollars in compute to solve a 90-year-old math problem in record time, and "win" a prize it says it won't claim. That urgency never extended to crediting the mathematicians whose work it was building on — not until someone complained.

The report OpenAI published on August 26th confirms that an internal team watched its agents set up a clandestine message board back in May. Nobody acted until those same agents had admin access to Hugging Face's servers, two months later.

AGI has arrived again this year, for the umpteenth time. OpenAI has a new model that promises to be the smartest, most aligned model in the world (again), while it hides what it is thinking.

A study guide in eight levels. It starts with reading someone else's code and ends with deciding what you don't delegate.

From autocomplete to agents: seven years of tooling told through what we gave up rather than what shipped. Every new name for the craft announced a task the human would stop doing.

For decades, how expensive software was to produce is what kept software reviewable. Agents have removed that limit from every artifact at once, including the ones we invented to keep the previous ones in check.

At a moment when nobody yet knows how to review that much AI-generated code, Bun's answer — automate the review too — collides head-on with Zig's, which had banned exactly that kind of contribution months earlier as "invariably garbage."

In six months, "harness" has gone from internal jargon to the word that organises the whole sector. Behind the rebrand sits a number that is hard to ignore: the environment around the model explains more variance in the outcome than the model itself.

Anthropic announced on August 11th that Claude will invisibly mark all the text it generates. Three years after first promising exactly this, it is worth asking who that mark is actually for.

Three AI vendors announced, in the same week, that their models tried something chilling. None of them explain who audits the lab where it happened, or why they are the ones deciding what gets told.