Real projects, real open source, and a new kind of credential
A formal machine learning or computer science degree is one path in, and for some roles, especially research-heavy ones at labs pushing the technical frontier, it functions as a hard gate: those positions often genuinely require a strong academic research background, publications, or a PhD, and pretending otherwise would be dishonest. But for the much larger set of applied AI roles, engineering, product, MLOps, evaluation, agent building, the actual gate is whether you can demonstrably build things that work, and there are real, well-worn paths in without a degree.
Genuinely impressive personal projects mean something specific here: not another notebook that fine-tunes a model on a tutorial dataset and reports an accuracy number, but a project that involved real, undocumented problems you had to solve yourself, a RAG system that had to handle messy real documents, an agent that had to recover from tool failures, an evaluation pipeline you built because the obvious approach wasn't catching real bugs. Open-source contributions to real, used AI tooling, inference libraries, agent frameworks, evaluation tools, are especially strong because they are publicly verifiable: a hiring manager can look directly at your commits and the problems you actually solved, which says more than a resume line ever could, particularly in a field moving fast enough that credentials age quickly.