Klemen Kotar announces PSI-0.5, a promptable physical world model
Kotar framed PSI-0.5 as contrasting with world models focused on moving around scenes; the reposted announcement we saw did not include links to code, a demo, or a paper.
By Ryan Merket ยท Published
Why it matters
If PSI-0.5 truly lets builders steer a world model with prompts, it could compress simulation setup and iteration from scripting to instruction, making embodied AI workflows faster for small teams.
Klemen Kotar (@KlemenKotar) says he is releasing PSI-0.5, a promptable physical world model, writing: "We're releasing PSI-0.5: a promptable physical world model. Many world models let you explore by moving around a scene, ..." The announcement surfaced via a retweet by Zoey Chen (@ZoeyC17), and the reposted snippet did not include links to a paper, demo, or repository.
The phrase "promptable physical world model" is the tell. Kotar is positioning PSI-0.5 against a class of world models that primarily let you navigate a simulated environment step by step. Promptable suggests steering the environment or its dynamics directly with instructions, rather than only issuing low-level moves. If that is the direction, it points to a more accessible interface for interacting with physical simulators and learned dynamics models.
What we know
- Name and version: PSI-0.5.
- Kotar's framing: a "promptable physical world model."
- Contrast: he notes many world models emphasize moving around a scene, implying PSI-0.5 focuses on prompt-driven control.
- Public assets: none linked in the post. No code, weights, paper, or demo URL was provided.
- Team and org: not specified. The post does not identify an affiliation or collaborators.
Why promptable might matter
Founders building embodied AI, robotics, and simulation-heavy products often struggle with the interface gap: domain experts can work with actions and state variables, but most product teams and end users want higher-level control. If a world model can be guided by prompts, it could lower the barrier to prototyping behaviors, generating scenario variations, and shaping data for training loops. It might also simplify how non-specialists ask a simulator for counterfactuals and edge cases.
There are also practical implications for iteration speed. Prompt-driven interaction could help teams describe scene setups and dynamics adjustments in natural language or structured instructions, then get immediate rollouts without hand-authoring every trajectory. That kind of loop can make simulation more like prompt engineering than scripting, which matters when a small team needs to explore many what-ifs quickly.
What we do not know yet
- Availability: it is unclear whether PSI-0.5 is downloadable, gated, or demo-only. No links were included in the post.
- Modality and fidelity: the post does not describe inputs, outputs, physics accuracy, or benchmarks.
- Mechanism: "promptable" could mean text prompts, image-conditioned controls, action primitives, or something else. The post does not specify.
- Backing: there is no stated company, lab, funding, or collaborators attached to this release.
For now, PSI-0.5 reads as a directional statement from Kotar more than a full reveal. The version tag hints there has been prior iteration, but without a repository, paper, or examples, the community is still guessing at the specifics. If and when links appear, the interesting test will be whether promptable interaction actually compresses the distance between an idea for a scenario and a runnable rollout.