## The Paradox One of the most formidable people I know in AI is a founding member of a frontier lab. No PhD, not even a grad degree. Never shipped a line of frontend code. And yet, he came up with the lab's biggest capability breakthroughs for users. He frequently trains half-broken prototypes on weird modalities, that are always just functional enough to learn something real. And because he believes in the bitter lesson, that simple methods with more compute ultimately beat clever algorithms, he spends just as much time collapsing those hacks into a single end-to-end model with the research team as he does inventing them. Anything to avoid the "solutions engineering" purgatory that comes from maintaining a pile of non-general systems that each need their own special handling, their own bug fixes, their own documentation. What he does has a name: forward deployed research. ## What is Forward Deployed Research? It's not academic research in the safety of a clean benchmark, and it's not engineering in the comfort of a fixed spec. It's the discipline of living with the problem, building just enough to learn, throwing it into the wild, and pulling the feedback directly into the next version. You know you're doing it right when you're uncomfortable: too much chaos for a traditional engineer, too much urgency for a traditional researcher. It's knowing which hacks are worth scaling, and which need to be crushed into the foundation before they metastasize into a support nightmare. It's having taste for what users actually want versus what benchmarks measure. It's a craft that rewards curiosity, stubbornness, and a willingness to get your hands dirty on both sides of the research-deployment gap. This is where the real progress in AI is happening now: not in the next percentage point on MMLU, but in that messy space where models meet reality. The people pushing these boundaries didn't set out to create a new discipline. They just refused to accept the traditional divide between research and deployment. They forged something new instead: forward deployed research. And as AI eats software, it might just eat the distinction between scientist and engineer too. --- *If this sounds like you, and you'd like to join a frontier AI lab, shoot me an email at anj at a16z dot com with 'FDR' in the subject.*