What it was like working on LLMs and security at Meta (2022-2026)

I loved my time at Meta, and I also counted the days between equity vests and daydreamed about quitting on the morning after almost every one. It was the most exciting engineering culture I’ve ever been part of, and at its worst it could feel like a perverse psychology experiment. I was proud to work on Llama, the foundations of LLM security, and on AI content moderation tech for 3.5 billion people; I was also embarrassed by the internal incentives that motivated us to keep people wasting more and more minutes of their one wild and precious life scrolling Instagram Reels.

Today Meta laid off ten percent of its employees. I left of my own volition in February. Many of the people I looked up to have already gone to Anthropic, OpenAI, and Microsoft. With a couple of months of distance from working there, after layoffs, departures, and the strange feeling of watching a place I loved take one more step down the staircase of its cultural ruin, it feels like the right moment to write something in the hopes it’ll be helpful to myself and others.

A Beehive of Individual Ambition

Imagine someone who freaked out the one time they got an A-minus, who has passed a series of meritocratic gates and received the appropriate accolades and awards, and for whom getting into Meta was like getting into their second-choice elite college, and is now at Meta, and expects a series of good accolades and rewards to continue to be meted out by the new institution they’ve joined.

This felt like the median engineer and product manager at Meta; the company recruited for and filtered for such people, placing them inside a Skinnerian conditioning experiment in which bad performance gets you managed out and good performance is rewarded with money that can change the shape of your life, or at least buy you a house and a financial cushion in San Jose or Fremont; everyone at Meta is running desperately from the former and manically chasing the latter.

The upside of this is that you could assume far more competence from basically everyone you interacted with at Meta than at other places I’ve worked, and sense a kind of background urgency and low-key fight or flight trauma energy animating whatever they were doing, including when they were collaborating with you.

I joined in summer 2022 at the tail end of the pandemic hiring boom, when I suspect they’d lowered the hiring bar, and of the ten or so friends I referred during my time there (once the company had begun the process of layoffs and attrition and raised the bar again), none made it through.

I always felt like an odd duck; I’m self-taught, have a graduate degree in the humanities, live and work remotely from Kansas, and somehow clawed my way up. But I learned that that’s comparing my insides to others’ outsides; I think all of us felt lucky to be there but that our status there was contingent and unstable.

Flatness

Meta’s employee net worth ranges from billions to nothing and is wildly unequal, but the company is also culturally egalitarian. VPs wear shorts and T-shirts and carry themselves like engineers. No one said “I manage a team” at Meta; everyone said “I support a team,” and in most companies engineering managers manage engineers, but at Meta, in the language at least, they served them.

This felt actually meaningful and good, as did the fact that every engineer’s title was simply “software engineer,” not “senior principal architect of X.” There’s probably an IC9 somewhere at Meta who runs all of Meta’s billion-dollar datacenters and undersea network cabling and thus manages billions of dollars in global capital but whose title is the same as the new grad’s, and I really liked this.

There was also an unpretentiousness and a kind of Israeli directness and flatness about the place I really enjoyed; any engineer could propose anything, could give blunt, polite feedback about anything to anyone, although how far the proposal or critique traveled depended somewhat on prestige, social network, and level.

Also, the place was full of nerds who played Magic the Gathering and were part of the Linux club in college, and aside from the competition there was genuine bonding over craft.

Workplace

The internal communication system, Workplace, is a literal fork of the Facebook app. Whereas on Facebook people flex about their vacations, on Workplace people flex about both their vacations and how productive and impactful they are.

Whereas people on Facebook fret about how many likes they got on their wedding photos, people on Workplace fret about that but also whether their director ‘heart’ emojied their confessional piece about the stress they’d been under in shipping that director-priority thing that half.

A typical thing on a team is for the tech lead to start a Google doc with a draft Workplace post that people pile into, which they then post on Workplace the next morning, for maximum engagement, while the tagged team members count the love and like emojis that roll in and think about their performance ratings.

The good part of this is that Meta is Workplace, which is full of nerd sniping about new optimization methods for post-training LLMs, new ideas in low-dimensional embeddings, and also, in the more fun groups, pictures of animals people spotted on campus, the canned food they’re eating while working from home, or dumb, borderline NSFW jokes.

The way I got involved with the Llama effort was by lurking in the Llama team’s Workplace group, reading their posts, and messaging people about how to start a security workstream within it. My team and I did the same to get AI protections into smart glasses, AI agents in Facebook Groups, WhatsApp, and whatever else looked interesting.

The place was alive. People cared about ideas; you could find people who’d debate AI xrisk with you and the fine-grained details of how we should evaluate fuzzers, and who cared about politics and social theory and the world. People argued from first principles. You could message someone working on one of the most important AI systems in the world and, if your idea was good enough, find yourself in the workstream a week later.

Missionlessness

I never sensed a broad cross-company feeling of purpose at Meta. Well-tenured folks I talked to were often lukewarm or privately acerbic about the societal value of our products, even though one could sense they’d been more passionate about mission in the 2010s, and you could see their old Workplace posts from back then.

The company had incinerated tens of billions on AR/VR hardware in a catastrophic attempt to pivot, and then tried to keep up in the AI race partly by firehosing cash onto the problem, failing to make it into the lead there too; there was a muddle of narratives where a cross-company sense of purpose should have been, none of which really landed.

In the vacuum where company mission should have induced some selflessness and company-level alignment was really everybody’s individual ambition. Sometimes, when good people clustered together on a team, there was a team mission shared by a dozen people who banded together as in a zombie apocalypse movie to try to succeed and make sure no one was part of the 20% or so that’d be pushed out that year.

I was fortunate to work with some of these teams. But in aggregate, what this dynamic wound up meaning was that for big pushes like virtual reality or scaled AI, the company often struggled to ship coherent products that would have required hundreds of people marching passionately behind one banner.

Conway’s Law in a Company of Individualists

I don’t know if you, dear reader, have ever tried using Horizon Worlds, Meta’s VR app, but it feels like a dumping ground of individual engineers’, individual PMs’, and maybe individual small teams’ ambitions with no coherence; it feels unfinished, basically, as do Meta’s smart glasses and VR headsets.

How this happens at the pattern level is as follows: first, Zuck has a new idea about how to solve the problem that Meta is basically a 2010s-era social media company few really deeply believe in. He then broadcasts to the company (and the world) that this is the next big thing. People have a range of degrees of belief in the new thing, but there is a uniform desire to grow one’s career at the company, and hopping on board the new thing thus seems like a smart move.

Then people flood into the new orgs Zuck creates for the new thing, all at very different positions on the spectrum of actually believing it’s a good idea. Regardless, everyone struggles to position themselves with the most glamorous scope within this new thing so they can get promoted to the next level, in a career-optimizing algorithm that serves their own goals but not a larger goal. The result is a massive muddle of a product or product family.

The same thing happened with AI. You had product leaders from mature, 2010s-era social products flooding in; some had backgrounds in older AI models, many had little experience with current AI, and the pivot to AI was obviously a career opportunity and an opportunity for individual advancement or at least survival.

This stood in contrast to more focused cultures like OpenAI, Anthropic, and DeepMind, which don’t necessarily have smarter people, but do have a stronger shared center of gravity.

Performance Measurement Panopticon

At Meta, data scientists were the performance monitors for the entire company; the way reward and punishment are meted out at Meta runs through what the data science org says about you, and downstream of those judgments are layoff decisions, manage-out decisions, promotions, bonuses, equity refreshers.

Literally anything you did at the company was measured. It didn’t matter if you were fixing people’s laptops at the help desk, growing ad revenue by improving recommendation algorithms, managing the free bicycles that were everywhere on campus, or reducing the likelihood our production servers would crash; there were measurements, and they’d be used to help determine your financial and career status at the company. I think even the kitchen staff were measured based on how many people were badging in at each cafe.

This meant data scientists wielded enormous power, even though they were compensated far less than engineers and product managers, and it meant people metric-hacked constantly, expending energy on bad metrics that didn’t model the problem they were actually solving, just to look good at promo and performance review time.

Performance Purgatory

I got lucky with the teams I worked on and the projects I picked and always got the coveted ‘greatly exceeds’ performance rating, but I did have one stretch, about two months long, that gave me a taste of what performance purgatory felt like. I’d advocated for and bootstrapped and done 60-hour weeks to start the workstream applying large language models to content moderation, and it had gone well; the workstream grew into a whole org.

Once we had several other IC7s on it, some of whom had been at the company for years and had big social networks (I was an IC7 in my first year), I gave my scope away to them, figuring delegation was a virtue; what I’d actually done was render myself scopeless inside the org I’d been the first cause of forming, and I floundered for a month and a half, not knowing what to work on, while my manager became increasingly alarmed and started sending the kinds of signals you learn quickly to read there: that if I didn’t figure out a role for myself, I’d be managed out.

It got to the point where I cried in a meeting with him while he pressed into my need to show more progress; this meeting was happening on a Saturday, over Zoom, with me standing at my standing desk, my family downstairs, wishing I’d come down and hang out with them and stop working, not really able to understand the pressure I was under, not understanding the nature of the performance panopticon I had entered in taking this job that was unlike any job I’d had until then.

Fortunately an opening came up in Meta’s security org around then, and I’m a security person through and through; I rescued my performance rating for the year, and did great after that.

The feeling of being on track to fail at Meta is awful in a particular, hollowing way; you can sense when it starts, people begin to shun you, don’t want to be in your meetings, your invites thin out, and you descend into a slow spiral that usually ends with people quitting of their own volition before the company quite gets around to firing them.

If they stick it out they become the walking dead of the company, going through the motions of a job that’s already over, knowing the writing is on the wall, commiserating with others anonymously on Blind, almost never getting back on track.

What I Loved

I lasted three years and seven months there, which transformed my career. With people far smarter than me, I got to help define the modern practice of AI for security, and security for AI. I got to represent Meta in a NATO working group, at big cross-industry meetings, at private strategy meetings with my peers at OpenAI, Anthropic, and Google, and to publish and speak publicly about my and my teams’ work.

I joined at the perfect moment; I’d been obsessed with neural scaling laws and LLMs when I got there and rode that wave; it felt like Type II fun, like running a marathon. The singular focus of it and the trees and air and road blurring by exhilarate you, but you’re always looking at your watch wanting it to be over and you’re always painfully fatigued and increasingly unsure of whether you have the stamina to continue.

The reward and punishment machine at Meta led to some great AI security work done by the sweat and tears of the folks I worked with. I left a few months ago, right after my promotion to IC8, and right as I felt I had a tenured position helping to lead the company’s AI security work. I counterfactually incinerated the financial and career windfall that would have come with that by leaving to start a company. I miss Meta, and I’m also so glad I’m gone.