Job Description
Research Scientist — Frontier World Models & RL
\n
San Francisco · On-site · $400k base + uncapped equity (TC ~$1M+) · No sponsorship
\n
\n
A stealth, exceptionally well-backed applied AI lab is hiring Research Scientists to solve open problems at the frontier of world models and RL.
\n
Three weeks from public beta, 100M+ views before launch, backed by names you'd recognise instantly. A rare chance to do frontier research where it ships — not sits in a paper.
\n
What they're building
\n
A hybrid world model that one-shots interactive 3D environments from a single line of text or one image. Coding agents handle logic, persistence and multiplayer; real-time diffusion handles the visuals. The same stack is dual-use for robotics simulation. The core science is working — this is about pushing it past the field's current limits, with real users and a data flywheel to train against the moment beta launches.
\n
The open problems you'd own
\n
\n
Real-Time Video / World Models
\n
Push the distilled style-transfer diffusion stack forward. Crack the failure modes the field hasn't: temporal consistency, multi-view consistency, and grounding in an autoregressive setting.
\n
\n
RL / Post-Training
\n
Advance RL post-training for coding agents — reward and taste modelling, agent orchestration, tool use, and turning real player trajectories into training signal. Design the method, don't just apply it.
\n
\n
Who we're looking for
\n
- \n
- A strong research background in real-time video/world models or RL post-training — top-tier programme or equivalent research output
- A publication and/or open-source track record at the level of NeurIPS, ICLR, ICML, CVPR, SIGGRAPH — but someone who cares more about the result shipping than the paper landing
- Someone who's authored the method, not just read it; who has a genuine point of view on why current world-model approaches fall short and what they'd do differently
- High agency, able to drive a research direction end-to-end with minimal direction
- Someone who chose early-stage deliberately — and knows why
\n
\n
\n
\n
\n
\n
The bar is simple: hires at this level move the valuation and pull in the next great researchers.
\n
\n
The honest bits
\n
- \n
- On-site in SF. Relocation supported. A rare, exceptional profile in Europe may be considered remote.
- The pace is intense — a genuine sprint through launch and the next raise. Built for people already obsessed with these problems, not those optimising for balance right now.
- Comp: $300k base as standard, up to $400k for exceptional people, with uncapped equity priced ahead of a major step-up. No visa sponsorship.
\n
\n
\n
\n
Process
\n
Fast — decisions in days, not weeks. Founder screen → technical interview → team lead interviews → a short trial task for likely-offer candidates.
\n
\n
Interested? Send a recent CV, plus links to representative papers, code, or demos, and a couple of times for a quick call. If these are your problems, you'll know it.