The Reverse Information Paradox
- Source
- x.com
- Author
- Satya Nadella
- Date
Context
An organization's corrections to an AI system can reveal how it makes decisions and judges good work. Satya Nadella argues that companies should retain control of this accumulated knowledge alongside their documents. Nadella's essay treats task histories, feedback and private criteria as assets that should remain useful when the underlying model changes.
The proposed design separates company workflows and memory from any single model provider. That would let a business replace a model while preserving what it has learned about its own work. This is an argument for enterprise control, not an announcement that a particular service implements those protections.
- Arrow's original paradox was the seller's problem: to sell information you have to reveal it. Patents fixed that by letting an inventor disclose an idea without forfeiting it. The buyer-side version AI creates has no equivalent instrument yet.
- The leak is not your data, it is the exhaust: prompts, tool calls, and above all corrections. A correction is the one thing a rival cannot buy, and it escapes imperceptibly, trace by trace.
- The asymmetry compounds. The vendor learns more about your operation every time you use the product, while you learn almost nothing about what it learned from you.
- Nadella calls the status quo ironic: providers claim fair use to train on public data, then restrict and reserve the right to learn from customer usage. One-way learning pools the value with whoever owns the learning infrastructure.
- eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
- distillation — Training a small, cheap model to imitate a big one's outputs, keeping much of the capability at a fraction of the cost.
“In the AI age, the buyer risks giving away knowledge, just in order to use what they bought.”
“In consuming intelligence, you are creating intelligence. And what you create should belong to you.”
“What the technical customers want is control over their compute, their models, their data stack, and their alpha. They want to know they own the means of production, and it's not being transferred to someone else.”
Alex Karp
“In the cloud era, enterprises accumulated data. In the AI era, they accumulate learning.”
Checking sign-in…
Loading comments…