In consuming intelligence, you are creating intelligence
Platform risks, Hardware secrets, & why every company must own its harness
I want to connect something Jason said, something Satya wrote and something which Sam Altman maybe did.
Jason Calacanis recently issued a blunt warning to YC founders…
If you accept OpenAI tokens in exchange for equity, there is a non-zero chance that OpenAI will study exactly what your startup is doing, copy the idea, and fold a free version of your app into their offering. Classic platform playbook.
Be careful.
A few weeks later the stakes rose dramatically. Apple filed a lawsuit accusing OpenAI of encouraging former employees to bring over confidential hardware presentations, secret prototypes, supplier details, unreleased parts, designs and project documents.
Apple claims a pattern of trade-secret theft that allegedly turbo-charged OpenAI’s nascent hardware efforts.
OpenAI has denied any interest in other companies’ trade secrets. The partnership that once put ChatGPT inside Apple’s ecosystem now sits beside a courtroom fight.
Against this backdrop Satya Nadella published a post that quietly reframes the entire conversation for companies of every size.
He calls it the Reverse Information Paradox.
It is the most important due-diligence framework for the AI era, and it maps directly onto what many of us have been calling harness engineering.
Arrow’s paradox flipped
Maybe Nobel economist Kenneth Arrow described the classic information paradox decades ago?
The seller of knowledge must reveal the knowledge to prove its value, yet once revealed the buyer has it for free.
Patents were the institutional fix.
AI creates the reverse problem.
When you buy intelligence (API tokens, hosted models, agent platforms) you must feed it your proprietary knowledge to make it useful.
Every prompt, every tool call, every correction, every private evaluation, every institutional “this is how we measure success” becomes training signal for the seller.
You pay twice: once in cash, again with the unique knowledge that makes your firm different. The better you want the model to perform, the more of that knowledge you must hand over.
Over time the asymmetry grows.
The model provider learns more about your workflows, preferences and failure modes.
You learn almost nothing about what they are learning in return.
From data exhaust to institutional know-how
Models do not only learn from public internet data.
They learn from the “exhaust” of real usage: the prompts people write, the tools agents invoke, and especially the corrections humans make when the model is wrong.
Every correction is distilled institutional know-how. It is knowledge a competitor could never buy on the open market. Yet under most current terms of service it leaks almost imperceptibly, trace by trace, eval by eval.
In consuming intelligence, you are creating intelligence.
And what you create should belong to you.
This is Hayek’s knowledge of time, place and circumstance.
It is the particular intelligence of your firm. It should not quietly transfer to the owners of the learning infrastructure.
Trust boundary…own the harness
Satya Nadella breaks the practical response into five interlocking requirements.
Together they form what the market is increasingly calling the harness (the continuous learning loop, the hill-climbing machine that sits above any single model):
Control
Create private evals that define what “good” looks like inside your organisation. Retain full ownership of your organisation’s memory, traces, feedbacks and context.
Capability
Build proprietary learning environments inside your tenant boundary so models can learn against real workflows without the knowledge ever leaving.
Choice
Decouple the orchestration layer from any single model. If one frontier model disappears tomorrow, can you still operate and optimise against your own evals with other models? Does the “company veteran” expertise remain with you even when the “generalist” model is swapped?
Cost
Decoupling also lets you compose context, models and tasks in the most efficient combination without sacrificing quality.
Compound
Bring the four together and you create a continuous learning loop. Your AI investments stop being pure consumption and start compounding the unique value of the firm.
So, a company must be able to use a model without giving up the knowledge that makes it unique.
Due Diligence for Founders and Enterprises Alike
Jason’s warning to YC startups and Apple’s lawsuit against OpenAI are not isolated dramas.
I think they are surface symptoms of the same structural risk Nadella describes.
When the platform that supplies the intelligence also has structural incentives (and technical means) to study, absorb and re-productise the unique ways customers use that intelligence, classic platform risk returns in a more powerful form.
The practical due-diligence checklist…
Who owns the traces, the private evals, the corrections and the memory?
Can I fine-tune or distill on my own data using the model outputs without restrictive terms?
Is my orchestration layer model-agnostic so I retain choice and cost control?
Do I have a real trust boundary that stops intelligence exhaust from leaving?
Am I building a harness that compounds, or merely renting intelligence?
But, I personally think…
The ultimate expression of that trust boundary is ownership of the model itself.
For many workloads the cleanest answer is no longer “use the frontier API with carefully negotiated terms.”
It is to run open-weight models (or your own fine-tunes and distillations) on infrastructure you control.
When the model lives inside your tenant, every prompt, every tool call, every human correction, and every private eval stays inside the same hard perimeter. The intelligence you create by consuming intelligence never leaves.
This is why the companies that will compound hardest are the ones treating the harness + the model weights as owned assets rather than pure consumption.
The model can still be rented when it makes sense for frontier capability. But the learning loop and the institutional memory must not be.
Chief Evangelist @ Kore.ai | I’m passionate about exploring the intersection of AI and language. From Language Models, AI Agents to Agentic Applications, Development Frameworks & Data-Centric Productivity Tools, I share insights and ideas on how these technologies are shaping the future.
COBUS GREYLING - At the intersection of AI & Language
Cobus Greyling is an AI Evangelist & thought leader dedicated to exploring the intersection of artificial intelligence…www.cobusgreyling.com