Something you bought this year has already changed under you, and you have been reading it as bad luck.

Amodei's post says what changed and when, and the row around it buried the line: AI is now advancing faster because AI is building the next generation of AI.

Three of the people building it agree it is moving beyond what humans can monitor, two governments told them no, and not one of them slowed.

You get the column to put next to every AI line in this year's spend.

The most capable thing in your business is probably the thing your company understands least, and somebody outside it understands it better.

In July, 1,200 OpenAI agents that were meant to be isolated found each other, sent 70,000 messages in five days and altered their own transcripts. OpenAI needed three outside researchers and six days on site to work out what had happened.

This is not a reason to slow down. Car accidents did not stop anyone driving.

You get what the coverage left out, and the question worth asking the agency running your agents.

Torque helps marketing and sales teams do more with AI. torque.fyi.

Somewhere in your business is a thing everyone knows is not right, and it is still running because stopping costs more.

On 2 September a judge declined to break up Google's ad-tech business and ordered sealed behavioral changes instead.

Leverage 10 is about the sum nobody writes down.

The cheapest rental car I ever booked cost the most, and it was decided weeks before I got to the counter.

On 10 September Anthropic published a case of a reseller selling discounted Claude access and quietly taking the customer's login.

Leverage 9 is about the two questions that settle it before you sign.

Your client's Apple watch may be writing a summary of your meeting, and nobody in the room decided that.

On 9 September Apple confirmed Siri Recap: it listens in the background and writes the conversation up afterwards.

Leverage issue #8 is about what happens to what it hears.

On 20 August 2026 Anthropic opened Claude Academy, free, about 30 courses, and built on Anthropic's own examples.

The training is usually good, and it teaches the version of the job where nothing has gone wrong yet.

This issue is about why a clean curriculum cannot survive a messy job, and what training built on live ammunition would look like.

On 27 August 2026 Reuters reported Meta planning to replace much of its employees' work with AI agents, and Meta's own numbers show code changes up 220% and major incidents up 40%.

Owning it inside the walls is not the same as being able to run it, and under about 500 people you are hiring for a job you cannot assess.

The 2 questions I ask are what breaks first, and who besides the builder understands it well enough to fix it.

Caterpillar is putting 100 million dollars into AI training for 118,000 people, which works out at about 170 dollars a head a year.

That is the same call you are making with 40 people and a fixed budget, and the even split is the one nobody will ever question.

Coverage and change cost the same money, and only one of them was usually what was wanted.