In August a team at Washington State University found a way to print one of NASA's rocket alloys on a commercial-grade laser for the first time. The alloy, GRCop-42, goes into the combustion chambers of liquid rocket engines because it carries heat well and stays strong while everything around it is trying to melt.
It has always been hard to laser print. The settings that produce a sound part are rare, the materials are expensive, and until now the process required a type of laser printer that is extremely powerful, rare, and expensive. Fewer than 10% of the laser printers in the world were capable of printing an alloy like GRCop-42.
As impressive as the properties of the alloy itself are, the real breakthrough was the recipe: the settings that let GRCop-42 be printed on a commercial-grade laser instead of the rare, expensive, high-powered machine the job used to demand.
There were more than a hundred million possible configurations of the machine. The researchers had already tried thirty-seven manually and watched them fail: parts that slumped, or melted, or came out full of voids.
They decided to leverage AI to search over 100 million options and determine which experiments to run next. The researchers had the model return 40 settings most likely to succeed. Inside those forty, they found six that worked, one of them at a record-low power that ordinary printers can reach.
“We are essentially democratizing the printing of this alloy," said Jana Doppa, the lead researcher. Not only will this reduce the cost for high-end applications, like aerospace, but it could lead to adoption and use in other industries as well.
That result was placed, in the newsletter where I read it, under the Lifestyle section.
The AI story that got the most attention was about private equity firms embedding engineers inside the companies they own, using the forward deployed engineering model. Honestly, that is the sort of use-case I would probably write about on a normal week.
Despite the massive investments, the flagship use-case was at a garage-door company, where those engineers were helping ship a notification that tells you a package has arrived. When most operators think about using AI for value-creation inside of portfolio companies, novel notifications probably aren’t what come to mind.
Both stories came out the same week. One was treated as the AI story. The other was treated as interesting trivia. I want to make the case that those two things are reversed, and not because of the novelty of what the Washington State researchers discovered.
Searching a Vast Terrain
Land navigation is about navigating to a location you haven’t seen or been to. You're given a point on a map and a point on the ground and must connect them through terrain you haven’t walked, often at night, usually carrying more than you'd like. You can't see the destination. After plotting your route, what you have is a heading, a pace count, and a running estimate of where you are. If you’re not disciplined, your route can drift, and each step takes you a little further off your heading.
The skill that separates the people who stay on-course from the people who wander is knowing when to stop walking and verify your position. You find a feature you can identify, a hilltop, a bend in a stream, a road junction, and you resection off it. A well-designed route often requires walking a bit longer to ensure you hit verifiable terrain waypoints along the way. You spend certain legs of the walk buying certainty rather than distance. It can feel like a waste, especially on tired feet and an aching back.
It's the opposite of a waste. A firm understanding of where you are can save an hour of walking in the wrong direction. That trade, between moving toward the goal and reducing your uncertainty about where the goal actually is, is the art of land nav. It is the balance between dead reckoning and terrain association.
The model that printed the alloy was doing exactly that, and the people who built it described the hard part in almost exactly those terms: every experiment returned a single bit, success or failure, and they recalibrated settings to reach the rare workable combination in as few tries as possible.
So the model didn't march toward the answer. The 40 settings It picked were grouped into small batches that mixed settings it thought would probably work with settings it was unsure about, because the uncertain ones, win or lose, taught it the most about where the good region was. It spent some of its forty tries buying certainty it could use to better navigate towards the destination.
The researchers also seeded it with the thirty-seven human failures. Realizing you aren’t where you thought you were isn’t a good feeling, but it is useful information. The machine treated dead ends as position information and moved accordingly.
This AI use-case was more like a navigation problem through a massive, uncharted territory than the typical chatbot-like use cases most people are familiar with.
The AI That Reads as AI
Now the story everyone did discuss.
This year many private equity firms began putting AI engineers directly inside the companies they own. The largest version, a venture called Ode built with Anthropic and backed by Blackstone and others, is valued at around a billion and a half dollars and staffed with a bench of engineers the venture went out and bought via aquihire.
Blackstone owns more than two hundred and seventy companies. It plans to bring Ode to twenty-five of them to start. Right now, engineers are embedded at six, one to two days a week.
At Chamberlain, the maker of LiftMaster garage-door openers, those engineers sit inside an eighteen-person team helping build software features: notifications when a package arrives, alerts about activity around the house, recognition of specific people at the door.
It’s not nothing. The features attach to a revenue line the company estimates at around 160 million dollars. The AI engineers also helped run a marketing campaign that came in at a fraction of the prior year's cost, giving a roughly 5x increase in ROAS. The work is sensible, measurable, and probably additive.
But, everything in that list is terrain the company can already see. A package arrives; tell the owner. Someone approaches; send an alert. These are features in a well-mapped space. Ring and half a dozen of their competitors make products that do that already. The value is real, but the ceiling is set by the market map.
That is simple terrain association: you move by matching what you see to features you can already identify. It works, it's fast, and it's what you reach for when the terrain is full of easily identifiable terrain features.
It's also the easy half, because the hard part, deciding where to look when you can't see, has already been done. The same engineers could probably be pointed at a search-function like the alloy team's; nothing stops them. The point is which kind of work gets attention and budget, and gets shipped first because it demos well.
The reason the embedded copilot reads as an AI story and the alloy search reads as trivia is not that it delivers more. It's that it sounds like AI. It talks, it notifies, and it can appear in a demo.
The alloy search produced six rows in a table and saved a lab an unspecified amount of time and effort. The democratization of GRCop-42 will take years as companies find ways to commercialize its use. That’s much harder to write a headline around.
AI or Trivia?
The alloy search doesn't look like “intelligence”.
It ends in a short table and a few months a lab didn't have to spend, and it doesn’t continue running on a loop. A notification system, by contrast, performs without stopping. It speaks. It appears on a phone. It does a visible thing in a demo while someone watches. Each of those is a small performance of intelligence aimed at a person in the room.
This is why the embedded-engineer story reads as the AI story. It produces exactly the kind of gain that shows: a feature someone can open, an alert someone can trigger, a number an operator can point to in a review. It is real and tangible, and that matters to the person approving the spend.
The alloy search is real and hard to understand. It resolves a problem almost no one in the room can see, which is the same reason almost no one in the room will fund it, and the same reason it lands in the newsletter under Lifestyle.
The comparison of these two use-cases highlights just how vast the terrain of AI actually is. The human-like text that LLM’s produce makes them easy to interact with and makes it easy to envision where they can add value. But, there are other advances in AI that can open new opportunity spaces that weren’t previously accessible.
I would bet the AI that ends up mattering to most companies will not feel like magic. No chat window, no demo, no moment where the machine appears to think.
Just a hard problem that used to take months or years that gets resolved in days or weeks, or a wrong turn that never gets taken, and therefore, never gets noticed.