After the descent, the only moat is memory

TL;DR

Intelligence is getting cheap for everyone at once. But rented intelligence forgets you completely. The only durable edge is what your own systems have learned from your own operation. Start the learning loop early, it compounds, and it does not transfer.

Fifteen years ago Marc Andreessen said software would eat the world. Now Chamath Palihapitiya has named the sequel. Intelligence is going down the same cost curve the smartphone went down, from a $500 luxury to six billion pockets. Only faster, because two discounts stack. Cheaper chips, and models that squeeze more out of them. He calls it the great descent. We think it is the most useful frame anyone has put on this moment.

He is right about the deeper thing too. The internet made information cheap and left expertise exactly where it was, scarce, expensive, locked in trained heads. What is falling toward zero now is not access to facts. It is judgment. The reasoning of the analyst. The intuition of the clinician. The drafting of the lawyer. The oldest bottleneck in the economy is breaking.

The interesting part is what actually arrives at the bottom of the curve.

Brilliant, and amnesiac

The expertise you can rent today has a property no human expert ever had. It forgets you completely. Every conversation is its first day on the job. It will solve the same problem for you a thousand times and never notice it has seen it before. It reasons like a veteran and remembers like a stranger.

That is not a small caveat. A consultant who forgot every engagement the moment it ended would not be a cheap consultant. They would be an expensive one. You would pay back in repetition everything you saved in salary. Rented intelligence answers questions. It does not accumulate.

The trap has a second floor

Chamath names the first trap. Consume the same off-the-shelf intelligence as your competitors, piped into the same generic workflows, and you have not built an edge. You have erased one. We would add a second floor. Even careful, proprietary use of stateless intelligence does not compound. Prompts can be copied. Workflows converge on best practice. A fine-tune is a snapshot, and it starts aging the day it ships.

What cannot be copied is what your systems have learned inside your operation. From your exceptions, your customers, your mistakes, your decisions and how they turned out. That learning is the edge. Today it walks out the door in retiring heads. Encoding it is not an act of writing it down once. It is building systems that keep learning it, every day, in production.

When everyone can rent the same mind, the only thing left to own is what yours has learned.

The loop is the moat

This is the wager worth making. The next layer of advantage is not better answers. It is remembered ones. Systems where every action becomes memory, every outcome becomes learning, and the answer you get in March is better than the one you got in January, because of what happened in February.

Compounding has a clock. A learning loop that starts this quarter has banked two quarters of your reality by the one after next. None of it transfers to a competitor who starts later. They can rent the same model. They cannot rent your memory. The descent makes intelligence cheap for everyone at once. It makes accumulated learning the only scarce thing left.

Where the descent lands first

The end of the curve is ubiquity, and we happen to work where ubiquity lands hardest, Southeast Asia and India, the places that took the smartphone from a luxury object to six billion hands. The same arc is now running for expertise. The companies here will not adopt it carefully and a decade late, the way the last era's incumbents did. They will do what they did with mobile and skip the middle step entirely.

Every company is somewhere on this curve. Curious, experimenting, scaling, or already learning. Wherever you are, the descent asks the same question. When intelligence is a commodity, what will you have encoded that is yours? Worth answering early. The water is moving.