The AI gold rush has shifted. Forget training bigger models. The battle now lives in deployment—getting existing AI tech into real-world hands and making it work. Google’s new push to hire hundreds of Forward Deployed Engineers (FDEs) proves it. These embedded coders live inside client offices, turning frontier AI models into production systems that actually move business metrics. This role is neither pure consulting nor traditional product engineering. It’s a hybrid that demands both deep technical skill and relentless focus on outcomes.

FDEs bridge the gap between AI’s promise and its messy reality. They scope customer needs, design tailored architectures, write production-ready code, and troubleshoot relentlessly. The role pays handsomely—average salaries near $238K, with senior engineers pulling close to half a million. Google isn’t alone. Anthropic, OpenAI, Palantir, Microsoft, and AWS are all scrambling for this rare breed. If you want to future-proof your AI career, mastering deployment beats chasing research breakthroughs.

Meanwhile, the era of finetuning is fading. OpenAI recently deprecated its finetuning APIs, signaling a tectonic shift. For years, finetuning was the poster child of AI customization—fine-tune a base model to squeeze out gains without blowing your compute budget. Now, it’s a niche, not the mainstream. The majority of AI engineering is moving toward long prompts, reinforcement learning from human feedback (RLHF), and continuous pretraining. Open models like Cursor and Cognition are doubling down on RLHF, showing that adaptation is less about retraining and more about smart instruction and reward hacking.

Cursor’s recent Composer 2.5 release highlights this transition. It’s tuned for long-running coding tasks and better instruction following. Behind the scenes, Cursor is building a much larger model from scratch with “SpaceXAI” resources—10 times the compute of prior efforts. This investment bets on pushing coding AI into production-grade reliability, not just flashy demos. Early user feedback praises its collaboration and efficiency, underscoring that raw model size alone doesn’t cut it anymore; practical deployment and sustained quality do.

Tech insiders are also sharpening their focus on kernel-level performance tuning. The biggest bottleneck in large language model (LLM) work is making abstract algorithmic improvements run practically at scale. Engineers who can optimize kernels, fuse operations, and squeeze out measurable speedups hold the keys to entry at elite AI labs. Vlad Feinberg’s recent job prep notes emphasize this, challenging candidates to hand-code JAX kernels and outperform existing primitives. This skill trumps fancy prompt engineering or surface-level tweaks.

On the agent front, the shift is from “chatty assistants” to persistent automation loops. AI agents now run complex CI/CD pipelines, monitor production traces, triage bugs automatically, and even generate pull requests. LangSmith Engine, Cognition’s Auto-Triage, and Anthropic’s Claude Code infrastructure all embody this trend. Developers are moving away from interactive chatbots toward embedded, verifiable systems that constrain, verify, and decompose tasks into reliable workflows. The best agents aren’t clever talkers; they’re dependable workers.

Meanwhile, the AI economy itself is a tale of extremes. Anthropic is growing at 10x per year, vaulting past OpenAI to become one of the world’s most valuable private companies. This contrasts starkly with mass layoffs at firms like Block and Coinbase, which blame AI readiness. The AI boom is still mostly hardware and infrastructure—compute and energy—while software teams face brutal market pressure. Strong AI companies scale; weak ones shrink. The divide is brutal and widening.

So where does this leave engineers? The frontier labs want kernel hackers who understand deep model internals. The product teams want embedded engineers who can ship AI-powered solutions that stick. The old finetuning tricks won’t get you there anymore. Instead, master scaling laws, kernel fusion, reward hacking, and persistent automation. Learn to deploy and debug in real-world chaos, not just research sandboxes. The AI gold rush is over—now it’s all about delivery.