RMH Deeplink does foundational research toward artificial general intelligence — building the models, methods, and theory that move the frontier, and applying them to humanity’s hardest scientific challenges.
Eight interlocking programs — from the general engine to the science it accelerates and the safety that governs it.
Models that perceive, reason, and act across language, vision, code, and the physical world — the general engine everything else is built on.
Learn more
Reward, search, and self-play remain the sharpest tools for behavior that exceeds its training data. We build agents that plan, explore, and improve.
Learn more
Protein structure, materials, mathematics, fusion, weather and climate — grand challenges where a model is either right or it is not.
Learn more
Interpretability, scalable oversight, and evaluation science so systems stay honest, controllable, and understood as they grow more capable.
Learn moreWe’re a scientific endeavour first. Solve intelligence — understand it well enough to build it deliberately — then use it to solve everything else. We pursue capability and safety as one discipline, never a trade-off.
Selected directions where a learned model becomes a new scientific instrument.
From folding and complexes to de novo design — modelling biology with enough fidelity to design it.
Read more
Reinforcement learning for magnetic confinement; weather models that run in seconds, not hours.
Read more
Formal proof search and autoformalization, where correctness is checkable and the bar is absolute.
Read moreA frontier safety framework, dangerous-capability evaluations, and interpretability research that scale with capability.
Critical capability thresholds, pre-committed tripwires, and red-teaming gate every model before it ships.
Learn more
Scalable oversight, interpretability for assurance, and evaluation science that keep models understood as they grow.
Learn more