We are an applied research lab based in San Francisco, building the verification infrastructure to scale omnimodal intelligence. Our founders come from frontier AI labs where they built multimodal systems, video world models, and large-scale evaluation infrastructure.
We are hiring interns across two tracks. The first is a general track spanning video understanding, generation, and gaming. The second is a robotics track. Internships run 2 to 6+ months, and we prefer working in person in San Francisco over the summer.
Track 1General TrackMultimodal RLSan Francisco2-6+ months
The general track: agentic work spanning video understanding, generation, and gaming (see our AgenticVBench and Nitrogen).
What you’ll do
- Design tasks and reward signals for agentic world modeling, iterating them out of successive training runs.
- Verification-focused: discover programmatic checks together with rubrics.
- Push the frontier of rubric-based RL, with rubrics-generator and grader-model co-evolution.
- Push the generator, verifier, and agent stack forward together.
What we’re looking for
- Enrolled in a CS, ML, or related program (BS, MS, or PhD), with a solid deep learning foundation.
- Familiar with large-model post-training, with hands-on experience in VLM or multimodal agentic RL.
- Strong engineering ability, and fluent with coding agents for fast day-to-day iteration.
- Run-driven: you read loss curves and eval breakdowns, and let the results drive the next decision.
Bonus
- Publications at ICLR, NeurIPS, ICML, CVPR, COLM, or similar venues.
- Experience with video or image generation, agents, or RLHF / DPO.
- A film or game background, with taste for what good creative output looks like.
- Core contributor to a well-known open-source project.
- Available for six months or more.
Track 2Robotics TrackEmbodied / SpatialSan Francisco2-6+ months
The same agentic world modeling work, pointed at robotics: connecting the capabilities of video generation and world models to robotics deployment.
What you’ll do
- Design tasks and reward signals for agentic world modeling, iterating them out of successive training runs.
- Build the reward model for physical and spatial consistency: its architecture, training data, and training recipe.
- Co-evolve failure modes and rewards: mine failure modes from training runs and fold them into the next reward iteration.
What we’re looking for
- Enrolled in a CS, ML, robotics, or related program, with a solid deep learning foundation.
- Familiar with VLM and/or diffusion model post-training, with exposure to embodied or spatial reasoning or world models (see spatial-intelligence work like ESI-Bench).
- Strong engineering ability, and fluent with coding agents.
- Run-driven, and genuinely excited about recursive self-improvement (RSI).
Bonus
- Publications on robotics, embodied AI, or video generation at top venues.
- Experience with sim, physics engines, world models, or robot deployment.
- Awards at IOI, ACM-ICPC, NOI, or similar competitions.
- Available for six months or more.
What we provide
- Frontier collaboration. Work closely with our frontier-lab research partners on the most advanced omnimodal intelligence.
- An open technical culture. We encourage independent exploration of the frontier, not experiments in a sandbox.
- Real support. Competitive internship compensation and a path to full-time, onsite in San Francisco.