Hello, and welcome back to Inc.'s 1 Smart Business Story. OpenAI is launching a global online challenge. The prize? A full‑time job at the company. The competition, which will roll out in multiple rounds, is designed to mirror the kinds of problems OpenAI researchers work on internally and to surface candidates who think creatively under pressure.

In an interview with Inc., OpenAI’s chief research officer says the company expects participants to come from a wide range of backgrounds, including self‑taught builders and founders who have been quietly experimenting with AI on their own.

In this article you’ll learn:

BY BEN SHERRY, STAFF REPORTER

OpenAI is launching an expansive online challenge aimed at finding the next generation of AI researchers. The reward? A full-time job. 

In an exclusive interview with Inc., OpenAI chief research officer Mark Chen says that the OpenAI Model Craft Challenge will be a contest that will task applicants with solving some of the common problems that the company’s researchers face on a daily basis. The contest will involve multiple rounds, testing candidates’ abilities across several fields of machine-learning research. 

The core question that OpenAI is asking applicants, says Chen, is “Can you come up with creative ideas in a sandbox setting?” Chen has been at OpenAI since 2018, when the company was a small research laboratory with no commercial products. He says that through years of work with a rapidly growing team of engineers, he’s learned that it’s easier for an inherently creative person to learn the technical aspects of machine-learning research than it is to teach a technical expert how to think outside the box. 

Chen expects that most applicants will be recent college graduates, young founders, and people who have been casually experimenting with AI systems.

The company points to Will DePue, an OpenAI researcher, as a model of the nontraditional backgrounds it’s looking to recruit from. DePue dropped out of college in 2022 after selling a company he co-founded in high school. Much of his machine-learning education came from OpenAI co-founder Andrej Karpathy’s popular YouTube channel, in which Karpathy walks through the process of building language models. DePue followed along with these lessons, training his first models while working from his bedroom. 

Today, DePue runs his own research team within OpenAI, and several of the people on his team also don’t have a formal machine-learning education; they’re former mathematicians, neuroscientists, and physicists—people who have tons of experience solving hard problems with novel solutions. 

The first challenge, named Parameter Golf, is all about recreating the problems that OpenAI researchers deal with in pretraining, the initial process of building a model by having it ingest tons of training data. Chen compares this process to building “an efficient rocket ship.” Applicants will need to build a small model under tight limits on how big the model can be and how much computing power they can use, and will be judged on how well it does on a test that includes data it wasn’t trained on.

“We’re often solving problems under compute constraints, or certain performance or time constraints,” explains Chen. Often, he says, these constraints result in the most creative solutions.

Applicants will need to adjust the model’s weights and training code to optimize its performance, but will only have 16 megabytes of space. Once the weights and code have been finalized, the model will have 10 minutes to train, powered by just eight Nvidia H100 GPUs. AI inference startup RunPod will provide up to $1 million of compute credits to the contestant pool. 

“We have no idea what the optimal solution might be,” says DePue. Instead of testing contestants’ ability to find a single correct solution, he says, the challenge is intended to identify people who are “trying weird and interesting and exciting things that prove that they can probably do ML research or have great ideas.” 

To excel in this challenge, says Chen, applicants will need to exhibit high levels of creativity, deep problem-solving skills, and a willingness to make tradeoffs. Contestants who do particularly well can even opt in to recruiting conversations with OpenAI’s hiring team. 

DePue says that people who are constantly thinking about the “if-then” nature of systems and algorithms often make for great researchers. “Many of the principles of machine learning are surprisingly simple,” DePue adds, involving some “basic linear algebra” and calculus. 

For those who have been closely following the AI industry but don’t have any experience training models, DePue recommends starting slowly. First, he says, learn how to build a basic transformer model, like the one originally introduced in Google’s 2017 paper “Attention Is All You Need.” Then, read a paper about some improvement to the original model, and attempt to replicate it. This should provide a solid base to build on, along with an understanding of the methods researchers use to eke out additional efficiency from models.

Applicants can try their hand at Parameter Golf here. This is just the first task that applicants will have to complete, Chen says. More information about future challenges will be revealed later.

OpenAI isn’t the first tech company to offer a competition like this: Google held its own international coding competition, the Google Code Jam, for years before shuttering the program in 2023.