LessWrong's AI safety engineering guide: What to learn and how

Ruben Castaing, an AI safety research engineer based in Christchurch, published a practical career guide on LessWrong titled "How to be an AI safety research engineer." Drawing from his own experience, he warns that this field lacks the structured path of medicine or law. Instead, jobs are found through people, not job boards. His advice: determine which AI safety issues you genuinely care about, then proactively email researchers who publish papers you find interesting. The field is small and mission-driven, so people are usually willing to help newcomers upskill. He also stresses the importance of attending conferences and AI safety coworking spaces, and recommends physically relocating to hubs like San Francisco or London. Though remote work is common, hub-based learning is significantly faster. Castaing expects the entire upskilling-plus-job-hunting process to take roughly one year, so affordability during the transition is a real consideration.

On the technical side, Castaing is unequivocal: Python is king, and you need deep learning expertise in PyTorch to do technical work. He points to Andrej Karpathy's "Replicating GPT-2 from scratch" as a starting point, then the ARENA curriculum as the gold standard for AI safety-specific engineering skills. Foundational linear algebra is essential, with 3Blue1Brown as a recommended learning resource. Practical tools include Git, Docker, Inspect for evaluations, and TransformerLens for interpretability work. Beyond raw engineering, Castaing emphasizes building "research taste"—learning to test ideas quickly, avoid scope creep, and question whether a subfield is genuinely worth your time. He advises newcomers to start by replicating interesting papers and to write jargon-free 1-pagers for feedback. Finally, he notes that research management, operations, and evaluation skills are currently in high demand.