Breaking Into Frontier AI Labs: An Honest Conversation

Chi Wang (Senior Staff Research Scientist @DeepMind) and Edward Cheng (AI/ML @NVIDIA) share their journeys and what really matters when building a career in AI research

The room was full of people with the same question on their minds: How do you actually break into frontier AI labs? Especially if you don’t have the “right” background?

On November 6th, we sat down with Chi and Edward, both with PhDs and now at leading AI organizations, to talk honestly about paths into AI research. What surprised us most wasn’t their credentials, but how much they emphasized that the path in doesn’t require following their exact trajectory.

Two Different Journeys

Chi has built his career across major research labs, from Microsoft Research to DeepMind, accumulating over 15 years of research experience in computer science. He’s the creator of AutoGen (now AG2), a popular and rapidly growing open-source framework for agentic AI that’s won multiple awards including best paper at ICLR’24 LLM Agents Workshop. He’s also created FLAML, a widely used open-source library for AutoML and tuning. Throughout his career, Chi has been a passionate advocate for open source. “These are open to everyone to join. You don’t need a PhD degree to contribute,” he emphasized about his projects. For Chi, open source isn’t just about sharing code. It’s about creating genuine entry points into AI research.

Edward’s path took a different turn. He joined a small startup as the second full-time employee, founded by MIT Professor Song Han. The early days were intense: daily standups at 8 AM California time, rapid iteration on cutting-edge problems in efficient AI. “It was a very intense experience for me, but at the same time I learned a lot,” he recalled. A year later, in February 2023, NVIDIA acquired the startup. His work shifted from optimizing single models for autonomous driving to building generalized software for large language models. He’s been at NVIDIA almost three years now.

What the Work Actually Looks Like

Different research environments operate in different ways. Chi reflected on the cultures he’s experienced. At Microsoft Research, there was a very free, bottom-up approach where individual researchers could set their own agendas and pursue their ideas independently. At DeepMind, while the ambitious, exploratory spirit felt similar, the work often involved larger scope and more collaboration across teams.

Edward described what software-hardware co-design means in practice at NVIDIA. They compress models from 16-bit to 8-bit or even 4-bit while maintaining accuracy. They build tools like Model Optimizer (now open-sourced). They optimize at multiple levels, from calibration algorithms to kernel-level attention mechanisms. “When you do optimization work, you need to think about both software and hardware side,” he said. The team also thinks carefully about what to integrate. “We don’t just integrate one algorithm from one paper showing good results. We test by ourselves if this algorithm can be applicable to multiple large language models and future models, then we select it to integrate into our software.”

The PhD Question Everyone’s Asking

Both speakers were direct: you don’t need a PhD to work in frontier labs.

“I know some very strong people in frontier labs who don’t have a PhD degree,” Chi said. “That’s definitely possible.” The key is demonstrating you can do the work. Contributing to open source projects is one of the most effective ways to show you have the skills. Open source contributions provide a public track record of your abilities. Completing research projects, whether through open source or independently, shows you understand research methodology and can think like a researcher, even without the formal degree.

The path in isn’t limited to research roles either. “You don’t need to go directly to a research lab,” Edward explained. “You can first be very good at an open source project or do very good engineering work, and then you can pivot.” Frontier labs have both researchers and engineers working together, and the boundary between these roles is blurring. You can break in as a strong engineer.

What Actually Gets You Noticed

When we asked what catches their eye in candidates, they got specific.

Chi sees open source as a genuine entry point. What kind of project impresses him? “Find a problem that none of the existing AI can solve. Not only the general purpose products, but also special purpose ones. Find a creative way to combine models, tools, or configurations to eventually solve the problem.” He added a caveat: make sure it’s a problem that won’t be trivially solved by the next generation of models, or the importance diminishes.

Edward’s advice was practical: “If you contribute to popular open source projects (TensorRT, vLLM) and you have a proven record where you contribute PRs and people review your PRs, that’s a very good sign.”

He continued: “Besides that, building a brand is important. Start your own repo on GitHub. Make everything reproducible by others. Have very good documentation. If you have something that can run end-to-end, that’s a very positive sign for us. That’s something that can catch our eye on your resume.”

Startup vs. Big Company

Edward was honest about the trade-offs between his startup days and life at NVIDIA.

In the startup: “Everything can be built from scratch. We read the latest papers and integrate features very fast into our software. It’s a very small team. You don’t need to worry as much about long release processes.”

At a large enterprise: “It’s a really long process when you want to release something. You have product teams, legal review.” But there’s an upside: “There are a lot of different teams with different backgrounds and specialized skills you can learn from.”

Neither is better. They solve for different things. Startups optimize for speed. Large companies optimize for reliability and scale.

Balancing Depth and Breadth

Someone asked a question many of us wrestle with: How do you specialize deeply while also staying broad enough to bring fresh perspectives?

Edward’s answer came from his own path. Before the startup, he worked at semiconductor companies writing kernels for different hardware (GPUs, CPUs, DSPs). That taught him low-level optimization. Then at Baidu Research Lab, he evaluated different AI accelerators, which pushed him up a layer to thinking about models. Later work took him even higher, to model architecture decisions.

“When you make a career change, you definitely can learn something new and then move to a broader knowledge base,” he explained. “First, get most familiar with your technical stack in your current team. Learn from teammates. Then when you move companies, you can specialize in something new while expanding your knowledge.”

It’s not choosing one or the other. It’s going deep where you are, then using transitions to expand.

The Honest Truth

If you’re wondering whether you can build a career in frontier AI labs, here’s what we took away from this conversation:

The path exists without a PhD. Both speakers know people doing excellent work in frontier labs without doctorates. But it requires strong engineering skills, fast learning, and building a body of real work.

Open source is a genuine entry point. Contributing meaningfully to projects like AutoGen, MassGen, FLAML, TensorRT, or vLLM builds both skills and visibility. Start your own projects with good documentation and end-to-end functionality.

Being a builder matters more than credentials. Show you can create things from scratch. Make them reproducible. Solve real problems, ideally ones current AI systems struggle with.

You can pivot. Engineering to research. Low-level to high-level. One domain to another. Each transition is a chance to expand while maintaining depth.

Learn from everyone. The intern, the colleague, the person on GitHub who’s brilliant at one specific thing. Good ideas come from everywhere.

The path isn’t narrow. It’s just not the one that gets outlined in standard career advice. It requires building, contributing, learning, and being patient with the process.

It’s not a guarantee. But it’s a real path, walked by real people.