Introducing Vivax Academia: Türkiye's First Academic Course on World Models
— by Vivax
Vivax is launching Vivax Academia, and its opening program is Türkiye's first academic course on world models, World Models 101 — hosted in mid-August 2026.
Today we are announcing Vivax Academia — the education arm of Vivax — and its opening program is, to the best of our knowledge, a first for the country: Türkiye's first academic course dedicated to world models: World Models 101. The course will be hosted in mid-August 2026, and it exists because of a gap we kept running into: world models have moved from a research curiosity to the most discussed frontier in AI, yet nearly everything worth learning about them lives scattered across papers, lab blogs, and conference talks. There has been no structured, academic-format course on the subject in Türkiye — until now. Vivax has spent the past years building its entire medical AI stack around clinical world models, and Vivax Academia is our decision to teach what we have learned, in a rigorous and structured way.
For newcomers, the subject itself: a world model is an AI system that learns an internal simulation of how a world changes over time — how a scene evolves, how an action alters a state, what consequences follow from an intervention — rather than merely mapping inputs to outputs. The lineage runs from the early recurrent world models of Ha and Schmidhuber, through latent-predictive architectures like JEPA, to today's large video-pretrained systems and world-action models that first imagine an outcome and then act to bring it about. The course treats this as a living field, not settled history: participants will follow the same debates the research community is having right now, from what a world model should predict to how anyone can tell whether it truly understands the dynamics it claims to capture.
The first half of the curriculum is about training world models. We walk through the full pipeline: assembling and curating the temporal data a world model learns from; the main training objectives — next-frame and video prediction, latent-state prediction, and action-conditioned dynamics; how text, actions, and other signals are injected as conditioning; what changes when you scale from a toy environment to web-scale video; and how trained models are evaluated, from reconstruction and rollout fidelity to whether the model's imagined futures respect physical plausibility. The emphasis is practical throughout — the aim is that participants leave able to read a modern world-model paper critically and to reason about what it would take to train one in their own domain.
The second half is about using world models — and about being honest on where they shine and where they do not. On the advantages side: a good world model gives you simulation and rehearsal before acting in the real world, planning over imagined futures, synthetic data generation, and a grounding in dynamics that pure pattern-matching systems lack — properties that matter enormously in high-stakes domains like medicine and robotics. On the disadvantages side: training them is computationally expensive, evaluation is still an open research problem, generated futures can look plausible while quietly violating physics or causality, the tooling is far less mature than the LLM ecosystem, and for many everyday tasks a simpler model is the better engineering choice. The course teaches both columns of the ledger, because knowing when not to reach for a world model is part of knowing the field.
Vivax Academia is the natural extension of the mission Vivax was founded on: medical AI grounded in how the world actually behaves — and now, a course that trains the people who will build it. The first edition of World Models 101 will be hosted in mid-August 2026. If you would like to learn more about the program — as a prospective participant, an institution, or a partner — write to us at info@getvivax.com. Applications will open on all Vivax social platforms; follow us there and stay tuned.
Curriculum. Six modules across two arcs: the first three teach how world models are trained; the last three teach how they are used in practice — and when they should not be. Module 1 — Foundations & Lineage: What a world model is and is not; from the recurrent world models of Ha & Schmidhuber, through latent-predictive architectures like JEPA, to today's video-pretrained systems and world-action models. Module 2 — Data for World Models: Assembling and curating the temporal data a world model learns from — sequences, actions, and consequences — and why data design decides what the model can ever understand. Module 3 — Training Objectives & Conditioning: Next-frame and video prediction, latent-state prediction, action-conditioned dynamics, and how text, actions, and other signals are injected as conditioning. Module 4 — Scale & Evaluation: What changes from toy environments to web-scale video; evaluating rollout fidelity and whether imagined futures respect physical plausibility — still an open research problem. Module 5 — World Models in Practice: Simulation and rehearsal before acting in the real world, planning over imagined futures, and synthetic data generation — with case discussions from medicine and robotics. Module 6 — The Honest Ledger: Compute costs, immature tooling, plausible-but-wrong futures, and the engineering judgment of when a simpler model is the better choice. Knowing when not to reach for a world model is part of knowing the field.
Who this course is for: Graduate students and researchers entering the world-model literature; ML engineers who want to reason about training and evaluating one in their own domain; Clinicians, roboticists, and domain experts working with high-stakes AI; Institutions and partners exploring collaboration with Vivax Academia. Who it is not for: Those looking for a no-prerequisites general introduction to AI or chatbots; Anyone expecting a settled textbook — this is a living field taught with its open debates; Teams that need product training for LLM apps; this course is about dynamics, not prompting.
Program details: Course — World Models 101. Format — Academic-format lectures with paper readings and critical discussion. First edition — Mid-August 2026. Applications — Opening on all Vivax social platforms — stay tuned.