This might be the hardest job to crack on the market right now,

SpaceXAI Member of Technical Staff - Model Training Austin, TX - Reposted 1 week ago - Over 100 people clicked apply Promoted by hirer - Responses managed off LinkedIn On-site Full-time ABOUT THE ROLE: You will work on the most critical modeling challenges at any given time. You will be on your first project before an offer. BASIC QUALIFICATIONS: You believe truth-seeking AI is the most important and challenging problem. You are obsessed with building genuinely useful models. You are a power user of AI models. If you previously trained models used by millions of people it's a plus, but modeling experience is not required. Take pride in your work and thrive in meritocratic environments. Median Package xAI Software Engineer Palo Alto, CA Total per year $640K Level 4 Base $240K Stock (yr) $400K Bonus $0 Years at company 0 Years Years exp 5 Years THIS MIGHT BE THE MOST VALUABLE AI PROJECT TO HAVE ON YOUR PORTFOLIO. Advanced 9 steps xAI | Useful Model Loop — Fine-Tune, Ship, Iterate Fine-tune an open-weights model with LoRA for one real task, ship it to users, and let evals = feedback drive every retrain. Python PyTorch LoRA Open-weights LLMs FastAPI Step overview (9) ~12h Jul 2026 Start building ARCHITECTURE Eval Harness Promoted adapter MODEL SERVER streamed replies PROMPTS FRONTEND Chat Surface RESPONSE COLLECTOR response collector USER INPUT USER INPUT IMAGE EVENTS Usage telemetry USER INPUT RESPONSE new data set Task Dataset current user data TRAINING LOOP Fine-tune Pipeline BUILD STEPS LEAN REQUIREMENTS Belief that truth-seeking AI is the most important, hardest problem (yes — it's a listed qualification). Obsession with building genuinely useful models. Power user of AI models — daily, deep fluency. Modeling experience NOT required (models used by millions = a plus, not the bar). Hustle in a fast org — emerge a shippee, no managers-of-managers. Strict communication: common, accurate knowledge sharing. Work ethic = ruthless prioritization. You learn your exact first project before the offer. What It Does: Picks one narrow task real people do weekly (niche summarization, rewriting, domain Q&A). Builds the eval FIRST — a scored set that defines what 'useful' means for the task. Fine-tunes an open-weights model on curated task data with LoRA. Ships it behind a minimal chat surface to actual users. Retrains and re-iterates every iteration, keeping a run log of what moved the eval. Publishes the loop: eval score over time, annotated with each training decision. A documented train -> ship -> measure -> retrain loop with real users on the other end — the miniature of what an MTS does, and exactly what 'obsessed with building incredibly useful models' looks like as evidence instead of a claim. Why It Fits xAI/Model Training: The posting screens for obsession with useful models and power-user depth — explicitly not modeling credentials — and a shipped model with real users, an eval log, and iteration receipts is the strongest possible proof of both.