And what about CMU?
I left Meta and joined humans& mid-November. People keep asking me how and why I made the decision. To be honest, there wasn't an elaborate process, and I wasn't even looking.
From my vantage point, I'd be starting at CMU as faculty in a year, and I was fine spending that time at Meta. I wasn't super content—the vision and mission didn't match what I wanted—but I was fine. Not actively looking.
Until on a random Thursday, I talked to Eric, just catching up with him about what he's building. And what he said resonated so much that I quit Meta the next Monday.
I work on privacy, LLMs, social interaction, and reasoning. My vision of how we should approach privacy is outcome-based—social, educational, and financial outcomes, rolled out long-term, for different people.
How almost all companies want to do privacy: they want you, as the privacy person, to write down a set of hard-coded rules, then post-train a model to follow them. Same standards. Same rigid rules for everyone. They want an arbiter of privacy who makes decisions for everyone—cookie-cutter, handbook-style.
But that is NOT how people do privacy at all.
Data sharing is incredibly nuanced. You can't really tell if something is okay or not in any given moment. Privacy decisions are usually based on what will happen in the future. If you share the info, will the person get higher insurance premiums? Will they be embarrassed? Or what happens if you don't share? Could it cost lives (like with the COVID vaccine)? Could it save face?
Privacy is contextual and based on long-term social good. You can't sit down with a notebook and be prescriptive about it.
And this isn't just about privacy—I believe privacy is just one application. Modeling, in general, should be like this. It shouldn't be like math and coding where there's always one verified ground truth. There is pluralism. There are multiple co-existing valid truths in the world for almost anything—especially anything human. No two humans are alike, and that's why verification-style modeling of verifiable tasks can only go so far.
People think I do privacy because I like being a moral arbiter or telling people not to do things. But I like it because it's gray. It's not black and white. It's highly ambiguous, non-verifiable, and fuzzy. It's challenging, interesting, and never boring.
Eric's mental model of how we should build models—the long-term good of people, models that consider people and their interactions—was so in line with how I see things. I had given the same pitch during my academic job talks so many times, and it was wild to hear those words come out of someone else's mouth lol.
I talked to him on Thursday, talked to the rest of the co-founders over the weekend, and quit Meta on Monday.
In life, I usually make decisions saying to myself "never a dull moment!" and this was in the same vein. I felt like these are some of the most cracked and KIND people out there, and I wanted to work with them and be surrounded by them and learn from them every day, if nothing else. And tbh, I haven't looked back since.
I had always been anti-startup—mostly because a lot of startups are just trying to ride the wave and the AI hype. People are incentivized by money, fame, power, and some narrow vision of AGI, none of which had ever resonated with me. So I would usually say NO to anything startup-y and nip it in the bud. But this was just so up my alley. It was the vision I've had for a while and a real possibility for it to come through.
I quit so abruptly that even my close friends were shocked—I didn't have a chance to tell anyone. After I posted a goodbye note to Meta, I got texts from people asking what happened and why. I wanted to mess with them, so I said I quit to go backpacking in Patagonia, and folks easily believed it haha. People know how academic-pilled and non-startup, non-AGI-pilled I am. But when I would tell them the truth, they had a harder time believing that I joined an early-stage pre-seed startup. That's just to say how much I care and believe in what we're building.
I get this question a lot. In the past few years, many young incoming faculty have done startup gap years and ended up never going back. So there's now this preconceived notion that the money, the weather, the access will keep you there.
For me: I have ALWAYS wanted to be faculty. I love and get energy from students and mentoring them, and it brings me so much joy to see students thrive. I have no intention of not starting my faculty job.
IN FACT, since starting at the startup, I've been so invigorated and motivated and inspired that I have more energy. I've already started multiple collaborations with students and faculty at CMU, I'm recruiting students, working on grants—and have been awarded some cute little compute grants already, woot woot!
I genuinely feel like I have the best of both worlds now. My two hats (member of technical staff at a frontier lab and faculty at a top university) reinforce and feed into each other. There's so much synergy. I think I'll be a better faculty member because of everything I'm learning at the startup and how I'm growing. And vice versa—everything I learn from my students helps me be more open-minded and versatile as a research engineer.
For me, an ideal world is one where I can have both together.
At the end of the day, here's to all the amazing humans who make every day worth living for me ❤️