You have to align it in the world, not in some simulation.
In a conversation that moved between Persian poetry and planetary ambition, Microsoft CEO Satya Nadella offered a rare glimpse into the human intuition behind one of the largest corporate bets in technology history. His conviction was not born in a boardroom but in the quiet recognition that a machine had preserved the soul of a Rumi poem across two languages — a moment that pointed toward something larger than efficiency or profit. Microsoft's partnership with OpenAI, its safety-first deployment philosophy, and Nadella's insistence on reaching all eight billion people rather than racing toward artificial general intelligence together form a vision of AI as civilizational infrastructure, not merely competitive advantage.
- Nadella's eureka moment — GPT-4 translating Rumi's Persian verse into Urdu and English while preserving its poetic essence — became the emotional anchor for Microsoft's entire generative AI commitment.
- Microsoft's rushed launch of the AI-powered Bing produced public hallucinations and triggered a competitive panic across the industry, forcing rivals to release half-ready products and critics to demand a full development pause.
- Rather than acquiring OpenAI outright, Microsoft engineered a novel partnership structure — sharing foundation model development while retaining control of tooling, safety, and responsible AI — betting that alignment beats duplication.
- Nadella dismissed calls for a six-month AI moratorium, arguing instead for a 'safety harness' approach: rigorous pre-deployment work combined with real-world feedback loops, accepting that true alignment can only happen in contact with actual users.
- His deepest concern is not the speed of AGI's arrival but the historical pattern of transformative technologies bypassing the developing world — he wants AI's abundance to reach everyone, not to repeat the slow, uneven spread of the industrial revolution.
Satya Nadella opened with something unexpected for a CEO — a personal story. Using GPT-4, he had translated Persian poetry by Rumi into Urdu and then into English, and the model had done something remarkable: it had preserved the poem's essential quality across two language boundaries. For Nadella, who grew up in Hyderabad longing to read Persian verse without the means to do so, it was a eureka moment — and a window into why Microsoft had committed so deeply to generative AI.
That commitment had been building for years. When Microsoft's models moved from GPT 2.5 to GPT-3, Nadella noticed the technology doing things no one had explicitly trained it to do — teaching itself to code, among other things. He became a true believer. Microsoft had already been developing its own model, Turing, across Bing and Azure, but rather than doubling down alone, the company chose partnership over acquisition. OpenAI's nonprofit status made a direct purchase impossible, so the two organizations constructed a hybrid structure: a for-profit entity that preserved OpenAI's nonprofit identity, with shared profits capped at a ceiling beyond which OpenAI would revert to its original form. The logic was simple — why run parallel foundation models when you could build one together and let Microsoft handle the tooling, safety infrastructure, and responsible AI work on top?
When asked about the risks of reaching artificial general intelligence too quickly, Nadella's answer shifted the frame entirely. He was not afraid of AGI arriving fast. He was troubled by a different historical memory — that the industrial revolution had taken generations to reach the world where he grew up. With AI potentially as transformative, he wanted it to reach all eight billion people, not just the privileged few.
The messier reality was harder to sidestep. Microsoft's AI-powered Bing had hallucinated publicly after a rushed launch, embarrassing the company and triggering a wave of premature releases across the industry. Critics demanded a six-month pause on AI development. Nadella refused that framing — stopping innovation was not the answer. Instead, Microsoft was building what he called a safety harness: careful pre-deployment work, followed by real-world alignment with actual users and real consequences. Hallucinations would improve. The technology was getting better. And when pressed on whether AI posed an existential threat, Nadella drew a steady parallel to other powerful technologies humanity had learned to govern — the power grid, nuclear energy — and expressed confidence that the same disciplined vigilance could work here too.
Satya Nadella sat down with Steven Levy at Wired and found himself describing a moment that felt almost personal—the kind of thing a CEO doesn't usually lead with, but which seemed to crystallize everything Microsoft was chasing in artificial intelligence. Using GPT-4, he had translated Persian poetry by Rumi into Urdu and then into English. It was not a mechanical word-for-word conversion. The model had preserved something essential across two language boundaries—the sovereignty of the poem itself. For Nadella, who grew up in Hyderabad wanting to read Persian verse but lacking the tools to do it, this was a eureka moment. It was also, he suggested, a window into why Microsoft had bet so heavily on generative AI.
The company's commitment to the technology had crystallized gradually. When Microsoft was moving from GPT 2.5 to GPT-3, Nadella noticed something shift. The model began doing things that impressed him—not because engineers had explicitly trained it to code, but because it had learned to do so anyway. That was when he became a true believer. The technology felt genuinely transformative, not incremental. Microsoft had been working on AI for years, building its own model called Turing that ran across Bing and Azure. But rather than doubling down on that path alone, the company made a strategic choice: partner with OpenAI instead of trying to acquire it. OpenAI was a nonprofit at the time, which made a straightforward purchase impossible. The two companies found a workaround—a complicated structure that allowed them to create a for-profit entity while maintaining OpenAI's nonprofit status. Profits from their collaboration would flow to both companies, but with a cap. Once that ceiling was hit, the arrangement would dissolve, and OpenAI would revert to pure nonprofit status. Nadella explained the logic plainly: why train multiple models in parallel when you could align around a single foundation and build everything else on top of it? They would do the foundation models together. Microsoft would handle the tooling, the safety work, the responsible AI infrastructure. Two independent companies, deeply partnered, moving toward one goal with discipline.
When Levy asked whether Nadella worried about hitting artificial general intelligence too quickly, the CEO's answer revealed a different preoccupation. He was not anxious about AGI arriving fast. He was haunted by something else—the memory that the industrial revolution had taken decades to reach the parts of the world where he grew up. Now, with AI potentially as transformative as that revolution, he wanted it to reach all eight billion people, not just the wealthy corners of the globe. Abundance for everyone. That was the vision.
But there was a messier reality to address. Microsoft had rushed Bing's new AI-powered version to the public, and it had hallucinated—given wrong answers, confabulated details, embarrassed itself and the company. The release had triggered a cascade: other tech companies, not wanting to be left behind, had hurried their own AI products out the door, ready or not. Critics called for a six-month pause on AI development. Nadella rejected that framing. Stopping innovation entirely was not the answer. Instead, Microsoft was building what he called a safety harness—elaborate measures to establish control before and during deployment. The company had not launched its Bing chatbot with GPT-4 the moment the model was ready. There had been work to do first. But Nadella also acknowledged something crucial: you cannot align an AI system entirely in the lab. You have to align it in the world, with real users, real feedback, real consequences. The hallucinations would improve. The technology was getting better. There were innovations in the pipeline to help correct when the model went wrong.
When Levy pressed on whether AI posed an existential threat to human life, Nadella acknowledged that uncontrolled power was indeed dangerous. But humanity had learned to manage powerful technologies before. The power grid. Nuclear energy. These required elaborate safeguards, constant vigilance, deep expertise. The same approach could work for AI. We can deal with powerful technology, he said. As for his own legacy—whether he would be remembered as the CEO who transformed Microsoft through AI—Nadella deferred. That was not his to decide. What he knew was that Microsoft had been making strides in recent years that kept it relevant, and AI transformation was at the center of that work. The rest would be for others to judge.
Citas Notables
GPT-4 did it in one shot. It was not just a machine translation, but something that preserved the sovereignty of poetry across two language boundaries.— Satya Nadella, Microsoft CEO
I am haunted by the fact that the industrial revolution didn't touch the parts of the world where I grew up until much later. I'm looking for something that may be even bigger than the industrial revolution and really doing what it did for the West, for everyone in the world.— Satya Nadella, Microsoft CEO
La Conversación del Hearth Otra perspectiva de la historia
You mentioned that moment with the Rumi translation as a eureka moment. Was it the personal satisfaction that mattered, or was it something about what it revealed about the technology itself?
Both, I think. Yes, it was personal—fulfilling a childhood dream. But what struck me was that GPT-4 didn't just translate word-for-word. It understood the poetry. It moved meaning across two language boundaries without flattening it. That's not trivial. That's the difference between a tool and something that might actually understand what it's doing.
You chose partnership with OpenAI over acquisition. That's unusual for Microsoft. What made you confident that structure would work better than owning the whole thing?
We could have tried to build everything ourselves. But OpenAI was already doing something remarkable with foundation models. Why duplicate that effort? We each had different strengths. They focus on the models. We focus on everything around them—the safety work, the tooling, the integration into products people actually use. One foundation, one platform effect. That's cleaner than two companies competing.
But there's a cap on profits. Eventually, the partnership ends. Doesn't that create an incentive to rush, to extract value before the clock runs out?
It's structured that way intentionally. Once we hit the cap, OpenAI reverts to nonprofit. That's the deal. Our goal isn't to maximize extraction. It's to get the technology right and get it to people. If we're fighting over profit margins, we're not focused on what matters.
You said you're not worried about AGI arriving fast. Most people in your position would be cautious about that statement. Why aren't you?
Because I'm haunted by something else. The industrial revolution took centuries to reach the parts of the world where I grew up. If AI is as transformative as that, I want it to reach eight billion people, not just the wealthy ones. Fast AGI means fast abundance. That's not a threat to me. That's the point.
The Bing chatbot hallucinated. Critics called for a pause. You rejected that. Why not just slow down?
Because you can't align an AI system entirely in the lab. You have to align it in the world, with real feedback, real users. Yes, we should have done more safety work before launch. But stopping entirely isn't the answer. The answer is to keep moving, keep learning, keep building the harness as we go.