Prakhar Agarwal
- Prakhar Agarwal has worked at OpenAI and is now a researcher at Meta Superintelligence Labs.There's a way for people to secure roles at top AI labs, Agarwal said.His advice: Use models a lot, spot their gaps, and learn beyond the classroom.
This as-told-to essay is based on a conversation with Prakhar Agarwal, an applied researcher at Meta Superintelligence Labs. The following has been edited for length and clarity. Business Insider has verified his employment and academic history.
I started my career at Apple in 2020. I spent five years there, then moved to OpenAI in the OpenAI API team. I moved to Meta Superintelligence Labs this summer when a lot of folks were making the shift.
I was in graduate school at the University of Washington, specializing in machine learning, when I applied to Apple. Later, OpenAI, Meta, and a bunch of other companies began reaching out, so I didn't have to explicitly apply for any of those.
I don't deny that experience plays a huge role. In most of these companies, the number of positions is pretty small, so naturally, they're converging more toward experienced folks.
These roles are very high autonomy. You don't have a traditional setup and hierarchy. Your role involves identifying a gap, then going to solve that problem. It's up to you to prioritize what is the right thing to address in the limited time and resources that you have access to.
Once you're in, you're pretty much thrown in the deep end. You define your own problems and try to come up with solutions. At OpenAI and Meta, they spend a lot of time hiring smart people. You need to tell them what needs to be done, rather than the other way round.
Interviewing at a top AI lab
The interviews test for a couple of things. First, do you understand the required nomenclature, and do you understand what LLMs are?
You still have to write code, but it's much more involved and related to the actual work you're doing at the job. You are fitted for scenarios.
The second thing they're trying to understand is whether you can operate in an ambiguous domain. Given an abstract problem, how are you concretizing and making it a workable metric-driven solution?
Having a Ph.D. helps. It conveys that you're able to work in an abstract domain. But if you can convey that in a different form, be it at a startup or in your role in building an integral piece of software, that is a good enough scenario to get a résumé accepted.
I recommend that people get their hands dirty and actually work on problems and solutions. Practical experience gives you the required skillset and a base to build on. It'll also teach you what not to do and what won't work. Building that intuition will differentiate you from the crowd at interviews.
Top tips for getting hired
At a minimum, make sure your theoretical understanding is good and work to understand the nomenclature required to do your job.
You also have to use these models a lot. Once you're using them, you'll understand what they are good at and what they are not good at, which is something people may overlook.
The ability to find gaps in AI models is actually one of the most important things that all of these companies are looking for. What is a gap that needs addressing in the next version of Llama? And once you've identified it, can you quantify that in a metric?
You'll also want to demonstrate that you know where things are trending. These are the capabilities that I think the model could be good at three or six months down the line.
High-bandwidth communication is really valuable
These top-tier AI companies are focusing on high-bandwidth communication.
The handling of the problem statements is happening at a much higher pace compared to Big Tech, where you spend a week trying to create a presentation. Here, you'll just go to a meeting room and discuss the problem over a whiteboard session before going to your own spaces and working on these problems.
These work conversations are usually one-to-ones, one-to-twos, or three-person conversations, so you should be able to articulate the gaps and problems well to people above you and people in the same peer group.
How to actually learn AI
What I've noticed about the AI communities is that they're very open about ideas or feedback.
If you get stuck with something, reach out to people on Twitter or LinkedIn. They are very likely to respond and help.
It might feel like a lot of information is beyond the classroom because the structured class coursework is pretty outdated. When you want to learn about these domains, don't just rely on your coursework or your professor or the books that were written probably five, 10 years ago to bring you to that level.
Consume knowledge from wherever it's coming from: a blog post, a YouTube video, or a conversation on Twitter.
Start following people who are sharing a lot on these domains. You might not be able to understand everything on day one, but you'll start picking it up.
Do you have a story to share about working at a top AI lab? Contact this reporter at cmlee@businessinsider.com.
