Story
The long way to the questions I am asking now
Growing up in Morocco, between cultures
I was born in Valencia, Spain, and my family moved to Morocco when I was three: Meknes, then Marrakech, then Tangier. I went to French schools the whole way through, so I grew up with two mother tongues, Spanish at home and French at school, learned German there as a third language, and picked up Arabic partly in class and mostly in the street.
I have always felt a stranger everywhere I have lived, and I grew up surrounded by very different cultures living side by side. It opened my mind early to the idea that no culture is the default reference point, and that what unites people is laughter, care for one another, struggle, and helping each other through it. I have wonderful memories of breaking the Ramadan fast with Moroccan friends and their families, of Catholic mass at the French church on other days, and of coffee and cake with members of the Jewish community other times. That was normal.
A formalism for an open mind
At eighteen, the obvious path was a classe préparatoire in Paris, the two years that prepare you for a grande école. I turned it down, partly to reconnect with Spain and partly to be a contrarian, and went to a brand-new engineering school in Madrid, Carlos III, where the late Aníbal Figueiras-Vidal introduced me to machine learning.
I became restless in Madrid, so I spent my final year at DTU, the Technical University of Denmark. That is where I met Carl Rasmussen. He later became my PhD adviser, and that was the start of my research years: Copenhagen, Toronto, Tübingen and Berlin.
I fell in love with Bayesian learning, where uncertainty is modeled explicitly and predictions come as probabilities. I still think of it as a philosophy for more than machine learning: be humble, hold several plausible hypotheses, let evidence narrow them, and do not let them collapse into a certainty the evidence cannot support.
Learning to ship, then learning what scale means
In 2007 I joined Microsoft Research in Cambridge and learned what it means to ship. We built adPredictor, the click prediction system behind Bing's ads, and an Xbox game developed from research code. Carl had moved to Cambridge too, and in 2012 we redesigned the machine learning course from scratch. Then California.
I spent nine and a half years at Facebook, the most amazing ultra-marathon of my life. Mark wrote in the IPO filing that "Facebook was not originally created to be a company." Working there did not feel like joining one. For me, Facebook and life became one and the same. When I joined in 2012, the culture blew me away. There were posters everywhere: "What would you do if you weren't afraid?" "Move Fast." "Be Open." "Focus on Impact." What struck me was that people really seemed to live by them. Early on I saw an engineer on the ads team deploying a change to the site. I asked who was helping them, and by what authority. Nobody else was involved. They were simply trusted to do their job.
I started in ads ranking, then founded Applied Machine Learning and led it for four years. AML was a labor of love. The team had its own logo and T-shirts, and people were proud to be part of it. We built the platform behind every production use of AI across the company. I did some of the best work of my life there, but I am especially proud of helping people grow and watching them do things they had not thought they could. By the end, though, I had been burning the candle at both ends for too long, although I did not fully understand it yet.
Technology is never neutral
At Facebook's F8 conference in 2017 I was onstage talking about how far computer vision had come. Later that year, at the NeurIPS AI conference, I listened to Solon Barocas and Moritz Hardt talk about fairness, and Kate Crawford talk about the trouble with bias. Joy Buolamwini's talk on the coded gaze had already made an impression on me. Around the same time, I was seeing uneven computer vision performance firsthand: good overall, and still leaving some people out. Whether that was good enough to ship was not a neutral technical question.
I asked to change focus and briefly became an individual contributor so I could go deeper into the ethics of AI. I quickly became a manager again, built the Society and AI Lab, and then led Responsible AI. Later, my boss helped me see how burnt out I was. It became clear to me that carrying on as before was a disservice to the team, so I moved back into an individual contributor role as senior technical lead for Responsible AI. I did wonder at first whether I had jeopardized my career, but the relief was enormous. The move gave me wings and let me work deeply on the ethical challenges of deploying AI at scale.
AI fairness challenges the engineering mindset at its core. We engineers want an explicit objective we can test: the thing is broken or it is not. Fairness is never solved. You can always do better, just like social justice, and there is more than one legitimate definition of fair. Treating everyone the same is not the same as trying to produce equitable outcomes. There is no neutral choice between equality and equity. Declining to decide is also a decision, made on someone's behalf. What worried me in 2017, and what I could not yet put into words, was that the machinery we had built was full of choices like that, about who benefits, who gets to participate, what counts as valuable and who decides. I could no longer tell myself that the values were somebody else's department.
I turned increasingly towards the societal impact of AI, and served on the board of the Partnership on AI, on the Spanish government's AI advisory council and as a Senior Fellow at Harvard's Belfer Center.
Making responsibility real
When I eventually left Facebook, it felt like divorcing the love of my life. LinkedIn gave me the chance to do something I cared about: make AI fairness work in production, not just in papers. I joined as LinkedIn's first Technical Fellow for AI, still as an individual contributor and initially part-time at fifty percent. Before long I was full-time, then working what felt like two hundred percent with amazing people, and loving it. In a cross-disciplinary effort, we formalized an approach that expected equal treatment from models and equitable outcomes from products, with those product decisions made explicitly and owned by someone. We shared this widely in a paper we published at FAccT 2023.
I became restless again, and joined OpenAI with the goal of staying hands-on and working on frontier AI problems. I started in Safety Systems with a very small team, but was later asked to lead Preparedness, including the work behind the Preparedness Framework v2. The work became very real when our evaluations started approaching risk thresholds with consequences for whether and how a model could be deployed.
What I wanted, and what was needed
As I wrapped up Preparedness, I did not want another big leadership role. I wanted to go back to doing things with my own hands. So, pretty crazily, I became an intern on the Health team for a little over two months. The team prepared a ramp-up plan for me. When I landed my first pull request, a documentation change, everybody cheered and clapped. Being helped was humbling; my younger colleagues were much faster than I was. Over time I found bugs and began to feel like a member of the team. I also got to make small contributions to HealthBench. They gave me a "Best Intern Ever" T-shirt, and a "Not An Intern" one when I left.
Then I was asked to lead Recruiting through a period of extremely rapid growth. I hesitated. It felt like a crazy call of duty, one of those times when you choose what is most needed over what you actually want. Stopping the technical work felt like ripping off part of my heart, but I became convinced that Recruiting was the most important problem I could help with at that time. One of my happiest moments every week became my staff meeting. I would look around the table and see trust, ambition and care for one another. It reminded me of the culture I had loved so much at Facebook. We started chanting "One Team." Someone even made a "One Team" Slack emoji, and it became a thing.
For a stretch in 2025 I was also interim head of People. Politics was everywhere, and I learned that politics is not always bad. I could not change every structure to match my values, at least not right away. Sometimes I had to stay in my swim lane, triage the next urgent problem, build trust and wait until I could tackle the structure itself.
What the second half is for
I left OpenAI in May 2026 to create space to decide what the second half is for. The question I keep coming back to is what it means to be a human being in the age of AI. I want to look at it the way I looked at Morocco as a boy, with no assumption that any one way of living is the default. The difference is that now I know how these systems get built, and that technology is never neutral. How does an intelligence superior to our own change how we learn, think, communicate and love? Does the balance of power change? How do we bring everyone along? What is human dignity in a world of artificial superintelligence?
They are personal questions too. I am trying to work out what freedom, no fear and true purpose mean for me. Part of that is making room again for the poet, the philosopher and the teacher in me.
I do not know yet what form this work will take, and I am trying not to answer that too quickly. For now I want to meet people working on these same questions from very different angles, and people at a similar stage of life and career who are also asking what comes next. If that sounds like you, write to me.
