How to Get Hired at a Frontier AI Lab (When Your CV Won't Help)

If you’re a new grad or a software engineer trying to move into AI, the standard advice — polish your resume, apply through the portal — is quietly broken. The traits that actually get you hired at a frontier lab don’t fit on a CV, so the portal is rarely how anyone gets in.

Simon Du knows the hiring side. He’s Reasoning Chief Scientist at Apodex, an ML theorist with a stack of honors (Sloan, NSF CAREER, MIT TR35), and someone actively building a team. Apodex builds deep-research agents that beat the frontier labs on hard search tasks, and Du was direct about what he looks for in a hire — and why the conventional path so often fails.

The job is less about IQ than you’d think

The first myth Du dismantles is that frontier AI work requires extraordinary raw intelligence. “I don’t think large model job is super hard. It requires a lot of IQ intelligence, I would say,” he says — then immediately reframes what actually matters. “It’s really about you should be very responsible, and you need to be hardworking. Those are more important characteristics of a good candidate we are looking for.”

Responsible. Hardworking. Detail-obsessed. These are the signals he’s hunting for. And here’s the problem: “It’s very hard to identify those signals.” None of them show up cleanly on a resume.

Why referral is how labs actually hire

That difficulty explains a hiring pattern most people notice but don’t understand. “If you look at all this frontier lab, how they hire — referral is probably the most common way to get a job in those frontier labs or any place that’s building the large models.”

The logic is straightforward once you see it. “Whether you are hardworking, whether you are responsible, whether you’re keen to detail — those are characteristics that are hard to be identified on your CV. So the most reliable way is just asking some referrals.” A referral is a trusted person vouching for exactly the traits a resume can’t prove. That’s why it dominates.

This is uncomfortable advice if you don’t have a network. But Du doesn’t stop there — he offers the move that builds the signal without needing a connection first.

The on-ramp: build an open-source agent framework

The most actionable thing in the conversation is a specific project, chosen for a specific reason. “If you have some open source project, especially on agent,” Du says, that’s the near-term direction for anyone trying to break in.

Why agents specifically, and not a model? Compute. “It’s very hard to get resource to train a large model. But for writing some agent framework, you only need API. So you can still build a pretty actually very comprehensive, strong agent framework just using API.”

That’s the whole insight. You can’t train a frontier model in your spare time — you don’t have the GPUs. But an agent framework only needs API access, which means a single motivated person can build something genuinely strong, in public, that demonstrates being responsible, hardworking, and detail-obsessed. The open-source project becomes the referral you couldn’t otherwise get.

Learning by doing — the deeper reason

There’s a second payoff Du points to, beyond getting hired. Later in the conversation, talking about how the next generation will develop judgment when AI does so much of the reasoning, he lands on the same advice. “I generally believe in learning by doing something. Even with AI’s help, you still learn the whole process. I think that’s still very helpful and really can improve your understanding about large model itself, also the reasoning or general system-building capabilities.”

So building an open-source agent framework does two things at once. It produces the visible proof of work that substitutes for a referral, and it forces you through the system-building process that actually teaches you how these systems work. Du’s broader theme — that “system-level thinking” is the skill to develop — is exactly what that project builds.

The practical takeaway

If you want into a frontier lab, stop optimizing the resume and start optimizing the proof. Build an open-source agent framework — it only needs an API, it’s in reach this weekend, and it demonstrates the traits a CV can’t. Ship it in public. That’s the most reliable on-ramp Du describes, and it doubles as the fastest way to actually learn the work.

FAQ

How do frontier AI labs actually hire?

Mostly through referral. According to Simon Du of Apodex, the traits labs look for — being responsible, hardworking, and detail-obsessed — are hard to identify on a resume. A referral is a trusted person vouching for exactly those qualities, which is why it’s the most common path into labs building large models.

What do AI labs look for in a candidate?

Less raw IQ than people assume. Du says the work doesn’t require extraordinary intelligence so much as responsibility, work ethic, and attention to detail. The challenge is that none of those traits show up clearly on a CV, which is why labs lean on referrals to surface them.

How do I break into AI work without a network?

Build an open-source agent framework. It only requires API access — not the compute to train a model — so one person can build something strong in public. The project demonstrates the responsibility and detail-orientation labs hire for, effectively creating the signal a referral would otherwise provide.

Why build an agent framework instead of training a model?

Compute. Du notes it’s very hard to get the resources to train a large model, but an agent framework “only needs API.” That makes it the realistic project for an individual: you can build something comprehensive and strong using existing model APIs, without GPUs you can’t access.

Does a computer science degree get you a job at an AI lab?

A degree alone is rarely enough, because labs hire for traits a CV can’t prove. Du emphasizes visible proof of work — an open-source agent project — and referrals over credentials. The degree may help you learn fundamentals, but demonstrated, public work is what surfaces the signals labs actually want.

What skills matter most for working on large language models?

System-level thinking, plus the soft signals of responsibility and detail-obsession. Du frames the whole agent system as an “operating system” and says understanding how such systems are structured and organized is increasingly important — more so than raw intelligence or any single technical credential.

Why is “learning by doing” the best way to build AI skills?

Because building a project forces you through the entire system-building process, even with AI’s help, which deepens your understanding of how large-model systems actually work. Du says this hands-on path improves both your grasp of the models and your general system-building ability — exactly what the work demands.

What’s the fastest way to demonstrate AI skills to employers?

Ship an open-source agent framework in public. It’s achievable without large compute, it shows responsibility and attention to detail, and it proves system-building ability — the traits labs hire for and can’t read off a resume. Du points to this as the most practical near-term on-ramp into LLM work.

Is software engineering experience still useful for moving into AI?

Yes — Du frames the move as a transition many software engineers are making. The bridge is building agent systems, which leans on system-level thinking software engineers already have. An open-source agent framework lets you apply existing skills while demonstrating the work ethic and detail-focus AI labs prioritize.