# David Ha > ~1985– · AI Researcher > > **Recorded contribution:** World models; neuroevolution; Google Brain; co-founded Sakana AI ## How to use this dossier Read for a causal chain, not a hero story: inherited problem → contribution → mechanism → downstream capability → limit. Then close the page and complete the reconstruction exercise from memory. ## 1. Historical orientation AI researcher David Ha is known for visual, interactive work on neuroevolution, generative models, and reinforcement-learning world models. At Google Brain he co-authored World Models, which trained an agent inside a learned compressed environment. In 2023 he co-founded Tokyo-based Sakana AI with Llion Jones and Ren Ito to explore evolutionary and collective-intelligence approaches. Ha's work uses deliberately compact experimental worlds to ask what a model must retain in order to act. World Models separated perception, temporal prediction, and control, while later evolutionary and collective-model projects explored alternatives to one monolithic gradient-trained agent. ## 2. The problem inherited Reinforcement-learning agents often learn expensive policies directly from raw observations and repeated environment interaction without an explicit predictive model they can use for imagination or planning. ## 3. The central contribution Ha co-developed the World Models demonstration and advanced accessible research showing how learned latent dynamics and evolutionary search can support agent behavior. ## 4. Reconstruct the mechanism 1. Compress observations into a lower-dimensional latent representation. 2. Learn a recurrent dynamics model that predicts future latent states and uncertainty. 3. Train a compact controller inside rollouts produced by the learned model. 4. Transfer the policy back to the actual environment and test failures caused by model error. ## 5. What changed downstream - World Models became an influential demonstration of learning behavior inside a model's imagined trajectories. - Ha's interactive presentations broadened access to research mechanisms and inspired model-based and evolutionary AI work. - The work renewed interest in learned simulators, latent-space control, neuroevolution, and small interpretable experiments that expose assumptions hidden by benchmark-scale systems. ## 6. Attribution, limits, and uncertainty - World Models was co-authored with Jürgen Schmidhuber and builds on autoencoders, recurrent dynamics, neuroevolution, and model-based RL. - A controller can exploit inaccuracies in its learned world; toy-environment success does not establish safe planning in open reality. - A controller can exploit inaccuracies in its learned dream, compact latents may omit rare hazards, and success in arcade or toy environments does not establish reliable real-world planning. ## 7. Reconstruction lab Learn a simple transition model for a grid world, optimize a policy inside it, and construct one model error the policy exploits. Compare imagined and real return. Search for an action sequence that earns high predicted reward but fails in the true environment, then add data or uncertainty estimates that expose the model exploit. ## 8. Evidence trail - [World Models](https://worldmodels.github.io/) — David Ha and Jürgen Schmidhuber - [World Models paper](https://arxiv.org/abs/1803.10122) — arXiv - [Sakana AI company information](https://sakana.ai/company-info/) — Sakana AI --- *Research checked 2026-08-09. Dates, roles, and claims about living people are historical snapshots. Linked sources remain the authority; this dossier is original instructional synthesis.*