Most companies do not need to “own their intelligence” in the way Microsoft CEO Satya Nadella is proposing.
His Reverse Information Paradox is a banger, and the argument deserves to be taken seriously. He breaks the response down into five areas: control, capability, choice, cost, and compound.
But this is not a black-and-white situation for brands. Some of these ideas are relevant. Most of them are not, at least right now.
On control, I agree that companies should create private evals because evals define what good looks like inside the organization. Companies should also retain their memory, traces, feedback, decisions, institutional context, and the outputs generated from their own tasks and queries.
But I agree for a different reason.
I do not want my memory stuck inside OpenAI because I want it to be portable. I want to use the same memory with Anthropic’s models or somebody else’s models. That is less about protecting my knowledge from foundation model companies and more about convenience and freedom.
The same applies to a company’s traces, feedback, decisions, and institutional context. Once that information is retained somewhere inside the company, its brand agents and company agents can retrieve it, use it to answer questions, and build on it as part of the company brain.
On capability, Nadella recommends building proprietary learning environments within the company’s tenant boundary to train or tune models. For the brands we work with, my answer is no. Nobody has the time to build and manage that kind of environment, and it cannot realistically be prioritized right now.
On choice, I agree that the orchestration layer should be decoupled from any single model. If one model disappears, the company should still be able to operate using other models.
Again, this is not primarily about protecting information. It is about having the freedom to choose between models. Some are faster and cheaper. Others are more expensive and more intelligent. They do not all belong to the same company.
The same logic applies to cost. A router system can send simpler tasks to less expensive models and work requiring more intelligence to more capable models. That makes sense, and companies should be doing it.
Bring control, capability, choice, and cost together, and you get the fifth area: a continuous learning loop that compounds over time.
What I do not see as realistic is the broader push for every brand to own its model, load it onto a server, fine-tune it, and train it. Most brands can benefit from portable memory, private evals, model choice, and intelligent routing without actually needing to own their intelligence.
Bora Celik
Founder, Gentic
--
Reply to this email or DM me on LinkedIn to discuss your brand's agentic transformation via gentic.co
