All Horsepower, No Radiator

One signal 🔭 One subtraction ➖ One analogy 🏎️

Created by Sam Rogers, building PAICE.work | Episode 1 with cohost Lee Rodrigues | The Sunday recap of the weekly show at sigsub.show

---

🔭 Signal: Context Over Capability

Between February and June, the frontier moved every 11 days on average. New state of the art, then another, faster than most teams can update a slide about the last one. Fable 5, ChatGPT 5.6, Opus 4.8, all inside a few weeks. The machines have never been smarter.

So watch where they still fail. Not on the hard reasoning. They fail the way a genius fails behind the McDonald's fryer on day one: brilliant, fast, and about to burn the place down, because the situational awareness of how not to burn yourself is not yet loaded to context.

That is the shift in one line. As capability climbs, the bottleneck stops being what the model can do and becomes what it knows about your situation. A model with no context doesn't hedge. It answers, precisely and confidently, the question your missing context actually asked. (Issue 056 watched this exact failure: one overloaded word, one clean falsehood.)

The model has never been smarter. That wasn't ever the part that was going to break.

---

➖ Subtraction: The Seven-Page Answer

Feed a rough idea into Claude, ask for the document, and you get seven polished pages back. They read beautifully. That is the problem: you can't tell which parts are yours, which the model invented, which have nothing to do with your point. Polish hides the seams.

Lee's move, from years of training designers: force it down to a one-page outline before you trust it. Plain text. No formatting, no bars, no just-in-case context.

The one-page test: take the longest thing AI wrote for you this week, demand its one-page outline, and count the lines you can't trace to your own intent. That count is what the polish was hiding. If it can't survive one plain page, you don't understand it yet. And neither does the model.

---

Watch, read, or listen

The full 26 minutes: YouTube. Every format in one place: sigsub.show/episodes/ep-001. Also on Substack and LinkedIn. Podcast on Apple and Spotify from Episode 2.

Jump to a segment: the one-page outline · context over capability · the Ford graybeards

---

🏎️ Analogy of the Week: All Horsepower, No Radiator

Every spec was green. The truck still ran hot.

Ford did what everyone's doing: handed a turbocharger redesign to the AI, let go a stack of engineers, let the specs carry it. The numbers came back beautiful, horsepower, fuel burn, cost, all green. Ship it.

Then they hired the graybeards back. One looked at the AI-approved design and said, more or less: we shipped this turbo five years ago and it cooked the car. It needs a bigger intercooler and radiator to survive towing over a mountain pass in July. The spec was right about everything except the one thing that mattered. It never knew the truck runs hot, because the AI was never on the warranty calls.

The specs had the data. The graybeard had the memory.

The spec was never wrong. It just didn't know what the old engineer knew.

---

🎵 Closing: The Graybeard Premium

Every disruption runs the same play: overinvest in the tech, underinvest in the people holding the context it can't see. Then, a little embarrassed, we hire them back. Ford calls theirs the graybeard army. That memory, still not machine-readable, trades at a premium.

The work is the same on both ends. Subtract your own output to the one page you can defend. Then point the machine at the context it's missing, before it swears the turbocharger runs cool.

So the graybeard question for your own shop: who held the context you just automated away, and are they still in the room?

This context thread isn't finished. More at the next one.

See you then,

Sam Rogers Context Mechanic https://SigSub.show/

---

New here? Subscribe at sigsub.show and the next episode lands in your inbox.

This week's cohost: Lee Rodrigues, who built the first YouTube Certified program at Google with me back in 2013.

Related reading:

• Vocabulary Debt (056) — the confident-wrong-answer this episode calls back to: one overloaded word, one clean falsehood.
• Audit Thyself (057) — why one voice isn't enough, the argument that turned the newsletter into this show.

---

Presented by PAICE.work. PAICE measures whether your organization can collaborate with AI safely, so the context your people carry doesn't vanish the moment a model joins the work.