Satya Nadella, the CEO of Microsoft, has been getting creative on Twitter. Through a message almost as long as the one that his colleague Asha Sharma, CEO of Xbox, took the opportunity to communicate the dismissal of 3,200 people just a week ago, Nadella has shared something that disturbs him: in the age of AI, information is power, and You don’t like others benefiting from your data. Nadella cites the phrase coined by Nobel Prize-winning economist Kenneth Arrow, whose paradox of the information market was that “the buyer does not know the value of that information until he has it, but by then he has acquired it, in effect, at no cost.” In Arrow’s formula, for the seller to convince the buyer of the value of the information, he or she must reveal enough of it to make the buyer interested. The problem is that at a certain point, the buyer will already know so much that they will not need to pay for that information. For Nadella, in the age of AI the problem is the other way around because it is the buyer of AI services who runs the risk of transferring knowledge valuable institutional, simply, to use artificial intelligence tools. For Nadella, companies are paying twice. “One with money, but another with something even more valuable: the personal and confidential knowledge that you must reveal for that intelligence to be useful.” According to the CEO, “the better you want the model to perform, the more knowledge of that type you will have to feed it.” That, precisely, is the problem. And this is what Nadella has dubbed the ‘Inverse Information Paradox’. The Inverse Paradox of Information Following Nadella’s reasoning, as a company gives more and more information to the owner of the artificial intelligence model, the asymmetry between both parties becomes increasingly skewed. You can imagine why: the seller learns more and more about your company as you use what you bought, while you learn very little about what the seller is learning in return. Did Nadella just describe what all companies do with search engines and services that traffic in our information? Maybe, but what the CEO is clear about is that, just as intellectual property patents solve one of the aspects of the Arrow paradox, since the inventor can reveal the idea without giving it away, the Inverse Information Paradox would need its own legal framework that protects companies that buy AI services. As? Well, that’s what we should see, since Nadella points out that models learn from the prompts that people write, the agents’ tools and, above all, the corrections that users make when a model makes a mistake. But Nadella goes further, stating that “When you consume intelligence, you are creating intelligenceand what you create should belong to you. This is your particular intelligence, the knowledge of time, place and circumstance, something that no one else can possess.” The message is tremendously ironic, but the reflection is useful and makes perfect sense: “if learning flows in only one direction, the economic value converges towards the owners of the learning infrastructure instead of towards the creators of knowledge.” Therefore, according to Nadella, “it is imperative that we distribute the learning infrastructure to be able to control that learning loop.” Advice for the age of AI (for companies, of course) In short, Nadella continues developing an idea in which what he explains is that the true competitive advantage of AI in the business environment does not lie in choosing the best model, but in owning its learning cycle. What if Microsoft ‘rents’ Claude either ChatGPTbut then it is Microsoft users who train that external model, that knowledge stays in Microsoft and don’t flow to Anthropic and OpenAI. At least free. Of course, and this would end up… Azurethe Microsoft cloud. Because the models belong to others, but if users access them through the Microsoft cloud, then Microsoft should keep all that valuable information, such as the corrections they make to the different models that have been used, remember, thanks to the Microsoft cloud. “If learning flows in only one direction, economic value converges towards the owners of the learning infrastructure rather than towards the creators of the knowledge” On the sidelines, in his very long message, the CEO of Microsoft proposes five points to ensure the company’s profits. Because, if in the era of the cloud companies accumulated data, in the era of AI they accumulate learning, and that learning should not escape if the following advice is followed: Control: create private evaluations, retain ownership of the company’s memory, comments and institutional context. Ability– Build proprietary learning environments within the confines of your own server to train or tune models without exposing company knowledge. Choice: decouple that from any individual model and be able to adapt to any AI model. Cost: thanks to that decision, you can bring together the context, models and efficient and profitable tasks without sacrificing quality. Capitalization: bringing together the four elements above to create a continuous learning loop that allows AI investors to multiply the value of the company. In the end, another thing that Nadella seeks is that there is no competition in the choice of models. Claude, GPT and Gemini are already good enough, so what companies that do not have equivalent models must do is compete to obtain all that knowledge of the learning loops, that does not escape from those companies and that, specifically, are built in Azure. Others have the models, Microsoft, Amazon and Google the servers. Now, not everyone agreesand both OpenAI and Anthropic they assure that their workflows and updates do not depend so much on those loops, but on other issues to keep frontier models evolving for customers who are not big whales like Microsoft. Because learning loops can work if you have Azure ready to provide the service, since the client does not care which model you use because there will always be a working model, but not all companies … Read more
