I work at Meta's Superintelligence Labs and used to be at OpenAI. Here's what the job is like — and what I've learned.

Navigating the landscape of cutting-edge AI research can be exhilarating yet demanding. Prakhar Agarwal, an applied researcher at Meta's Superintelligence Labs and a former OpenAI team member, shares his experience and insights into the dynamic environment of frontier labs. Agarwal's daily responsibilities fluctuate significantly based on project phases and immediate goals. At both OpenAI and Meta, projects are marked by critical milestones, such as extensive training runs lasting around ten months. As deadlines approach, the pressure intensifies. The work he undertakes is always aligned with the current iteration of the AI model, making it crucial to deliver solutions that address specific issues identified at that stage. Missing deadlines can lead to uncertainty about whether future model versions will retain or resolve previous issues. When deadlines are further out, the focus shifts to evaluations, uncovering failure cases, and identifying problems within the existing models. This environment is characterized by unpredictability; tasks that seem straightforward may take a day, while more complex issues could stretch over a week. One of the primary constraints in these foundational labs is computational power. Unlike larger tech companies that can hire numerous employees to tackle separate project components, the need for compute resources means that the team size must remain manageable to ensure productivity. Effective communication is paramount in these labs, where streamlined interactions among stakeholders can significantly accelerate the pace of innovation. Teams tend to be smaller and more cohesive, enabling a fluid collaboration where individuals frequently work across different projects and teams. The ratio of senior to junior staff is often skewed, allowing for broader project scopes and collaboration beyond immediate teams. Agarwal emphasizes that communication is essential, especially in environments where documentation may lag behind the rapid evolution of code. Researchers must articulate their processes, justify their actions, outline next steps, and convey results clearly to elicit feedback. Familiarity with the code itself is vital, as understanding can often provide insights that documentation may not cover. The strength of these foundational labs lies in their experience with experimentation. Teams develop a keen intuition about what approaches are likely to succeed or fail based on past trials. While the focus on results can overshadow the learning that comes from failures, Agarwal believes that these missteps are invaluable for growth and development. As for managing the inevitable burnout that can come with working on the cutting edge, Agarwal suggests embracing the unpredictability of the field. He advises aspiring researchers to remain open to new ideas and flexible enough to pivot quickly when necessary. In an industry that evolves at breakneck speed, adaptability becomes a crucial skill. For those interested in the world of AI, Agarwal’s experiences highlight the unique challenges and rewards of working in leading research labs. The journey is filled with learning opportunities, not just from successes but also from the myriad attempts that lead to breakthroughs.