This told essay is based on a conversation with Prakhar Agarwal, an applied researcher at Meta Superintelligence Labs who previously worked at OpenAI. The following has been edited for length and clarity. Business Insider has verified his work history and educational background.
How I spend my day varies greatly depending on where I am in the project and what my immediate deliverables are.
OpenAI and Meta will achieve these milestones (for example, training at scale and running reinforcement learning) within 10 months. I get nervous when a deadline approaches.
The work I identify is always based on the current iteration of the model. When I say that the model is bad at X and my solution will help fix X, it’s based on that version of the model. If the deadline has passed, we don’t know if the next version will have the same problem.
If you’re further away from that deadline, you’re primarily working on evaluation and trying to find failures and pain points in existing models.
The work is very dynamic. Sometimes we think something is so easy that we’ll be done with it in a day. And vice versa. It could take a week because there are so many unknowns.
I feel that working at Frontier Labs is very different from Big Tech.
The limitation in these basic labs is computing. Unlike Big Tech and other places, we can continue to employ large numbers of people and give them small tasks.
Everyone needs computing to actually do anything, but as soon as there are too many people, computing becomes divided and no one can do anything.
It also requires high-bandwidth communication between stakeholders. No need for 10 different communication layers. The speed of iteration is much faster. These core groups tend to be much smaller and tight-knit.
The concept of “team” is also very fluid. Each has their own project, but they work with others to work on joint projects. Meta and OpenAI have a lot of senior people and not many juniors, so everyone has projects of a certain scope.
At times, I collaborate with people externally more than within my immediate team. The scope is not limited to 4 or 5 people. Your scope is the problem you are trying to solve.
Deepening communication and coding is key
Communication is paramount in these labs. Many things are undocumented, so you need to be clear about what you’re doing, why you’re doing it, what the next steps are, communicate your results, and be able to get feedback on your work.
Being comfortable reading through code and identifying details is one of the most important skills I’ve ever seen. Code evolves much faster than documentation. If you get stuck on something, read the code and try to figure it out yourself.
Having some understanding of what’s going on in different industries also gives you an overview of the ideas and approaches people are trying. It’s all closely related, so you might be able to learn something from it or find a way to contribute.
The biggest advantage of these labs is that they know what works and what doesn’t.
The research paper says, “We did X, Y, and Z in this particular order and it worked.” But what you don’t see is that before you did X, Y, and Z, you tried 50 different things that didn’t work. And people aren’t talking about it.
To me, that is the real strength of these basic institutes. Thanks to all the experiments and all the work done so far, the team has built a very strong intuition. They know what doesn’t work, what doesn’t scale, and what works well.
Outsiders often look for profits, but they miss the point that even failure can be extremely valuable.
Advice for those who want to work in a top-notch laboratory
There is no good answer to managing burnout. I’m mostly just going with the flow. You work on the cutting edge, but simply put, if you want to be here, you can’t be thinking about strictly day-to-day things.
What I would tell my younger self is to be comfortable exploring new paths and new ideas. What I’ve seen is that we either try to play to our strengths or we try to stay in a decisive situation where we know it will work. But things move so fast in these fields that you need to be able to switch to new topics.
Build muscles that can withstand being thrown into something completely new. In some cases, it may be more a psychological issue than a skill issue.
Do you have a story to share about working at a top AI lab? Contact this reporter at: cmlee@businessinsider.com.
