The Talented Mr. Ripley

The help that studies you.

Every lab wants you to think of its AI as Brock Purdy. Mr. Irrelevant, the last pick in the whole draft, who steps in when the starter goes down and turns out better than anyone had a right to expect. That's the pitch: the eager backup who was only waiting for a chance to help.

You didn't hire Brock Purdy. You hired Tom Ripley.

Meta printed the proof this week. It shipped a model close to the frontier that almost nobody at the frontier will touch, but that's not the story. The pricing is. Meta will sell you Muse Spark two ways: full freight for privacy, or ninety-five percent off if you let it train on everything you type. Tomasz Tunguz did the arithmetic. The gap between the two prices is $1.24 per million tokens. That's not a discount. That's Meta telling you, to the penny, what your data is worth to it.

Which is the tell for the whole business. Intelligence is going to zero (we said so Wednesday), so nobody's really selling intelligence anymore. They're using it to buy the one thing still going up in value: you.

Not all of you costs the same. Ask Google for a running shoe and that's a fair trade, you wanted the shoe, they showed you the shoe. Scroll Meta or TikTok and you hand over your attention for the dopamine. A rough trade, but a trade. Point a frontier model at your own business and there's no trade at all. It runs your collections process ten thousand times and learns your margins, your customers, and exactly how you work. You didn't sell it anything. You taught it.

That's not a tool. That's Tom Ripley. The agreeable houseguest you invite in to do the work, who studies how you sign your name and then wears your life better than you did. Amazon already ran the play, watched what sold on its marketplace and launched Amazon Basics on top of the sellers. Apple sherlocked the apps that did well in its own store. The labs are next, because a pure model company can't outbid an ad company for data (Meta can eat the inference loss all day), so the only price it can charge is your outcome. Get paid on the result, learn the shop from the inside, then be the shop. One of your own said the quiet part Wednesday: "the labs are going to roll out outcome-based pricing and revenue share by vertical." Obviously they are.

So the smart money splits. If your edge is a number, like Jane Street's, you anonymize the trade and pay up, protected by moats a lab can't cross. If your edge is how you work, you can't anonymize the pattern, so you buy real privacy or run open weights, which are cheaper anyway. Everyone in the middle is paying tuition to train their own replacement.

Which leaves two seats Ripley can't study you out of. The hardware that keeps your data on the device, and the compute that collects its rent no matter who wins. Apple sells the one. SpaceX sells the other. Neither one ever asks for your secrets.

You didn't buy a tool. You let Ripley in. And Ripley never hands the job back.

At COAI today: the full Signal/Noise, with the data-asymmetry map, the who-pays-and-why taxonomy, and the Amazon-to-Apple downstream playbook, is live at getcoai.com.

Outsider Labs Advisory →

95%. That's the discount Meta gives you on its new model if you let it train on what you feed it. Read it backward and it's a price tag: your data is worth so much to an ad company that it hands back almost all the inference revenue to get it, about $1.24 per million tokens, $454,000 a year at scale. A model company can't match that bid, so it charges you another way, by learning your business well enough to run it. Free was always reconnaissance.

Read the data clause before you take the discount. Every "free" or steeply discounted AI tier is a data sale, and now Meta has printed the number so you can't pretend otherwise. Today: pull your top AI vendor's terms and find the training-rights line. If the cheap tier trains on you, it's only cheap if the data you're feeding it is worth less than $1.24 a million tokens to you. Most of yours is worth more.

Split the book by what the data reveals. Commodity work (formatting, first drafts, ticket triage) can ride the cheap or open model, because there's nothing in it worth stealing. The judgment-dense work that exposes your margins, your customers, and your process pays full freight for zero data retention, or stays in-house. Today: draw the line down the middle of your workflows and route accordingly. That line is your real security policy.

Treat an outcome-based offer as the tell, not the deal. When a vendor stops charging for tokens and starts charging for your result, cheer for a second, then look again. Pricing on your outcome means measuring your outcome, which means learning whether they can produce it without you. Today: if anyone's pitching revenue share or pay-per-result, ask what data it takes to price that, and who owns the model that learns from it.

Friday, September 4

— Harry and Anthony

Sources: