← All speakers

Pim de Witte

Conference affiliation: General Intuition

Pim de Witte is co-founder and CEO of General Intuition, which builds models that perceive, predict and act in interactive environments. He previously co-founded Medal, where millions of players record and share gameplay. General Intuition builds on that platform’s action-labeled video: pairing what a player saw with the controls they used gives a model information about how actions change an environment.

On this page

From game communities to humanitarian software

Pim ran a large private RuneScape server at 14. In 2014, he worked with Médecins Sans Frontières; his humanitarian work included improving health information systems during emergencies and co-creating MapSwipe, a tool through which volunteers contribute to mapping. His personal site also describes work on satellite-data labeling during disasters.

His open-source projects include UnityMainThreadDispatcher, initially written for Firebase Unity. Unity requires operations such as interface updates to run on its main thread. Pim’s utility lets code on other threads queue work for execution there, addressing a practical constraint in building interactive applications.

At Medal, Pim and his team built a recorder used across tens of thousands of games. The resulting footage spans different rules, visual environments and controls. He subsequently created Highlight, an AI assistant built on Medal’s technology. In his announcement of General Intuition, Pim described moving to work fully on the lab while Ken Colton became Medal’s CEO.

Learning from controls and consequences

Pim’s argument for training robots on games starts with a shared interface. A person can use the same controller to fly a virtual aircraft and pilot a physical drone, watching the consequences and making small corrections. In his AI Engineer NYC talk, he proposes training models at that level: visual observations go in, controller actions come out, and the robot’s existing software translates those commands into lower-level motion. The model continues receiving observations as it acts.

Several choices distinguish this approach:

  • Record actions rather than reconstructing them. Images can conceal the input that produced a movement. Pim uses an aircraft’s rudder adjustment as an example of an action that may be difficult to infer from video alone. Synchronized controls provide that missing information. He sees inferred action labels as a promising longer-term approach, while favoring recorded inputs because they avoid depending on a separate model’s ability to recover the actions correctly.
  • Use games’ density of movement. Flying a drone requires many small, sometimes simultaneous corrections. Pim argues that gameplay supplies abundant examples of this behavior, giving action models practice with rapid control sequences. At AI Engineer NYC in 2026, he reported that Medal captured more than 1.5 trillion action tokens a month. He also reported that one model could play games at a high level and control more than ten types of robots using the same control primitives; these were capabilities he described ahead of broader model availability.
  • Separate immediate action from longer-term planning. Pim expects developers to combine these models with an orchestrator that handles longer-term goals. His proposed accessibility example makes the division of control concrete: a player operates the mouse while a model observes the game and supplies keyboard actions, leaving the player’s own inputs to them. He also proposes using action models to explore interfaces during UX testing. These are applications he outlined, rather than established outcomes.

World models as places to learn

In an essay co-written with Packy McCormick, Pim argues that learned world models can approximate unfolding environments without programmers explicitly specifying every interaction. An action model chooses what to do; a world model predicts what follows. Together, they offer a way to explore consequences before attempting actions in a costly physical setting.

MIRA gives this research direction a concrete form. General Intuition and Kyutai developed the multiplayer world model in collaboration with Epic Games. Trained on bot-generated Rocket League matches, it produces an evolving four-player scene in response to player controls. Underlying physics measurements were reserved for evaluation rather than training, allowing the researchers to assess what the model learned from video and actions. MIRA is a collaborative team result, distinct from Pim’s individual arguments about how such models should be used.

Where to invest in AI

Pim’s June 2026 argument for a Dutch AI strategy opposes spending scarce talent and compute on a national frontier language model. He favors evaluation, adapting existing models, and work in fields where the Netherlands has distinctive expertise and data, including biology, chips and the natural sciences.

In a SPRIND interview, Pim also calls for European compute infrastructure and funding that allows researchers to test unfamiliar approaches without continually preparing for another financing round. He places evaluation, interpretability and failure analysis among foundation-model builders’ responsibilities: other developers will extend their models into settings the original teams cannot fully anticipate.

1 conference talk

Loading more talks and highlights…

References