Training models is one job out of several
Ask most people what a career in AI looks like and they picture one job: someone training neural networks in a lab, tuning hyperparameters until a model gets good enough. That job exists, but it is a shrinking share of the actual work happening under the label 'AI.' ML and AI engineers still build and fine-tune models, and applied researchers still push the boundaries of what models can do, publishing or at least internally advancing new architectures and training techniques. But a huge and growing share of AI work now sits downstream of that: taking an existing foundation model and building something useful with it. This is a genuinely different discipline from training models from scratch, and it has opened the field to people who will never touch a training run.
MLOps and infrastructure engineers build and maintain the pipelines that get models into production and keep them running reliably: data versioning, model serving, monitoring for drift, managing GPU costs. Data engineers build the pipelines that feed AI systems clean, well-structured data in the first place, which is unglamorous but is often the actual bottleneck on model quality, more than architecture choices. None of these roles require you to have invented a new model architecture, and all of them are, professionally, AI work that companies pay real money for.