Every month the same thing happens. A frontier lab posts a research engineer role, and ten thousand people apply to it by Friday. Almost all of them are qualified on paper. Almost none of them hear back.
The front door at OpenAI, Anthropic, and Google DeepMind is the most crowded door in tech. You can stand in that line. Or you can notice that the building has other doors, and most of them are barely guarded.
This is the honest version of how people actually get in. What the labs screen for, why competing head on is the slowest path there is, and the specific work you can do in the next month to walk in through an edge instead of the crowd.
What Frontier Labs Actually Screen For
Strip the brand names off a resume and three things predict who thrives inside a lab. Intent, mathematical maturity, and grit.
Intent is taste about what to work on. The strongest researchers do not chase interesting problems, they chase important ones, and they can tell the difference before the result is in. Mathematical maturity is not a transcript. It is the ability to pick up an unfamiliar problem, strip it to its structure, and reason about it cleanly. Grit is the willingness to stay with something hard for months after the novelty wears off.
None of those three fit in a cover letter. All of them show up in work you have actually shipped. That is the entire game, and most people never play it. They polish the claim instead of building the proof.
Stop Competing at the Center
The center of the field is the part everyone can see. Pretraining a flagship model. Sitting on the core research team. Touching the biggest cluster. Those roles are real, and they are also where every strong candidate on earth is pointed at the same time.
The center has a second problem. You cannot rehearse it from the outside. You do not have a hundred thousand GPUs in your bedroom, so you cannot show a lab you can do the thing they are hiring for. You can only say that you could. That is the same position as the person who sends four thousand applications for two offers, just with a fancier resume on top.
The edges are different. They are the parts of building a model that are essential, that decide whether the lab ships, and that you can actually do on a single rented GPU or no GPU at all. There are two of them. Below the stack, and above it.
Below the Stack: Make the Model Faster
Below the stack is systems work. Kernels, memory, the physics of getting tensors through an accelerator without wasting half of it. This is where Flash Attention came from. It is where quantization methods like LLM.int8 and QuIP came from. It is where the languages people write fast kernels in came from. Most of it started outside the labs and got adopted because it was undeniable.
It is unglamorous and it matters more than almost anything. A lab notices the person who made a layer two times faster on the same hardware, because that person just handed the whole team more compute for free. No tool writes this for you. You have to understand what the accelerator is actually doing, and almost nobody bothers to learn it. That is precisely why it stays open.
Above the Stack: Make the Model Useful
Above the stack is everything that turns a raw model into something that does real work. Agents that plan and use tools. Evaluations that measure whether a system is good instead of flattering it. Careful experiments on how models behave when you push on them.
This is the newest edge, which means it is the least crowded. The wave of open agent and program search work exists because people ran rigorous experiments on top of models they never had to train. You can do the same this weekend. The cost of entry is a willingness to run real experiments and report honest numbers instead of vibes.
The Crowded Center vs the Open Edges
Put the two paths side by side on the things that actually decide whether you get a reply, and the choice gets hard to argue with.
| The crowded center | The open edges | |
|---|---|---|
| Who you compete with | Everyone, at once | A small, self-selected few |
| Compute you need | A frontier cluster | One rented GPU, sometimes none |
| How you prove it | You can only claim it | You can show working code |
| Time to a real signal | Years | A month |
| What the lab sees | Another strong resume | A contribution they can run |
| Who reaches out first | You do, into a filter | Sometimes they do |
The center and the edges are both real work inside a lab. Only one of them is something you can prove from the outside this month.
The Curriculum: What to Build This Month
None of this is theory. Here is a concrete month that ends with something a lab can actually look at, in order.
- Work through the JAX tutorials until you can write and debug your own training loop without copying one.
- Do the exercises in the scaling book until you can reason about what a model costs to train and run, not just recite the numbers.
- Implement a transformer from scratch, end to end, with no framework doing the hard parts for you.
- Then pick an edge. Write a custom kernel that speeds up a mixture-of-experts layer, or build a small agent and a real evaluation that proves whether it works.
- Put all of it in public, with the numbers, so anyone can run it and see that the result is true.
Finish that and you have done more than ninety percent of the people who applied through the front door. You have a thing that demonstrates intent, maturity, and grit at the same time, instead of a paragraph that claims you have them.
Where the Credential Still Matters
This is not a cheat code, and I am not going to pretend it is. A demonstrable contribution that people adopt gets you the conversation, the referral, the second look. It does not erase the parts of hiring that are still about pedigree and timing, and for a few pure research roles a PhD is still the price of entry.
What the edge work changes is which line you are standing in. Instead of being the ten thousandth identical applicant, you are the person who shipped the faster kernel that someone on the team already starred. That is a different conversation, and you can start it without anyone's permission. It is the same reason the job market keeps absorbing change instead of breaking under it, a thread we pull on in The Next Billion Jobs.
The One Line That Decides It
The way into the hardest companies in the world was never to be a slightly better version of ten thousand other applicants. It is to build the one thing they cannot get anywhere else and put it where they can see it.
Build the rare thing at the edge. Let the search take care of itself. The front door was never the way in, and the edges were open the whole time.
Build the proof. Let Yara run the search around it.
Join the Yara waitlist at yara.soFrequently Asked Questions
How do you get a job at a frontier lab without a PhD?
By showing work instead of claiming it. Frontier labs hire for intent, mathematical maturity, and grit, and a demonstrable contribution proves all three faster than a degree does. The most reliable path is to build something real at the edges of model development, like a faster kernel or a rigorous agent evaluation, publish it with the numbers, and get it in front of someone on the team. A PhD still helps for some pure research roles, but adopted work is what gets you the conversation.
What skills do frontier AI labs actually look for?
Three traits predict who does well inside a lab. Intent, which is taste about which problems are worth solving. Mathematical maturity, which is the ability to take an unfamiliar problem apart and reason about it cleanly. And grit, the willingness to stay with something hard long after it stops being fun. None of these show up in a cover letter. All of them show up in work you have already shipped.
What should I build to get noticed by OpenAI or Anthropic?
Build at an edge you can reach without a frontier cluster. Below the stack, that means systems and kernel work that makes a model faster on the same hardware. Above the stack, it means agents and honest evaluations that prove whether a system actually works. A concrete month: work through the JAX tutorials, implement a transformer from scratch, then write a custom kernel for a mixture-of-experts layer or build an agent with a real eval. Publish all of it so anyone can run it.
Do you need frontier compute to break into AI research?
No, and assuming you do is why most people never start. The center of the field, training flagship models on giant clusters, is closed to outsiders by definition. The edges are not. Kernel optimization, quantization, agent design, and evaluation are essential to every lab and can be done on a single rented GPU, sometimes none at all. That is exactly why those edges are the least crowded way in.
How long does it take to land a frontier lab job?
Longer than a week, shorter than the years people assume. The skills take months of deliberate work, not a weekend. But the signal can come fast. A single adopted contribution, a kernel someone stars or an eval a team starts citing, can turn into a referral in days. The slow part is building the proof. Getting it in front of the right people is mechanical, and that is the part you can hand off.