How The Current Approach to AI Falls Short
What is the future of AI? How will it evolve over the next years and decades?
Yan LeCun, formerly the Chief AI Scientist at Meta, believes that AI will progress beyond the Large Language Models (LLMs) like ChatGPT, DeepSeek, or Claude, which have dominated the discourse so far. Instead, LeCun argues, the future of AI lies in systems that understand the physical world, which he calls “world models.”
The core of LeCun’s argument is that today’s LLMs are fundamentally limited because they are based on language alone; as a result, they cannot fully understand reality.”
“Most of what we learn and most of our knowledge is through our observation and interaction with the real world — not through language,” LeCun explains in one interview. He adds that, “The vast majority of human knowledge is not expressed in text; it’s in the subconscious part of your mind…most knowledge really has to do with our experience of the world and how it works.”
For AIs to reach their full potential, they must move beyond LLMs; rather than predicting predicting the next word or next token, AIs must evolve towards Joint-Embedding Predictive Architectures (JEPA), which:
- Learn abstract, high-level representations of the world, enabling systems to simulate scenarios, understand causal signals and consequences alike, and plan actions;
- Ingest multimodal inputs, such as video, audio, spatial signals, and sensor data. This permits the AI to move beyond text, and grasp physical constraints like time, motion, causality, and object permanence;
- Reason, plan, and maintain persistent memory, a stark contrast to the transient, temporary context windows of LLMs (which LeCun believes will become obsolete within five years).
World Architectures in Action: Healthcare
In fact, these architectures already exist, and are in use today in healthcare. Anyone who’s ever worked in healthcare, whether they are vendors, clinicians, operations team members, or leaders, will be familiar with the following saying: “If you’ve been to one hospital, you’ve really only seen one hospital.”
When it comes to challenges and circumstances, every healthcare facility is unique in its own way, and can even change from one day to the next, shaped by patient safety, staff constraints, and unpredictable events. Because the system is constantly in flux, it cannot be fully captured by documentation alone.
Kontakt.io builds a functional world model of healthcare systems, so that hospital teams can directly understand, address, and improve these realities. By providing spatial, temporal, and causal awareness of hospital operations, the Kontakt.io platform moves beyond written documentation or electronic health records, neither of which can provide the full picture in isolation.
Instead, LeCun’s framing and Kontakt.io’s mission stem from the same idea: shifting from reactive systems to proactive ones. Today, we hear statements like:
- “This unit is always congested.”
- “These pumps are underutilized, why do teams keep asking for more?”
- “Transport took longer than expected, so care will be delayed.”
In contrast, world models will generate more accurate, relevant predictions, such as:
- “If six patients are admitted from the ED in the next 90 minutes, ICU throughput collapses at 2:30.”
- “If pumps are rebalanced now, nurse search times will decrease by 18% on the night shift.”
- “If staffing drops by one full-time employee, this ward will become a bottleneck.”
Patient Journey Analytics: The World Architecture that Models Care Delivery
Within the Kontakt.io platform, the best example of this world architecture is Patient Journey Analytics. By combining real-time signals from RTLS devices with clinical and operational data from EHRs, Patient Journey Analytics can move beyond reactive fixes and historical reporting, and instead model the hospital as it exists in the real world: continuously, spatially, and over time.
For each patient, this world model captures:
- The temporal state: where each patient is in their care progression, and how long each phase will likely take;
- Spatial state: where the patient is physically located, and where they are expected to go next;
- Resource coupling: which staff, rooms, and equipment are required to move care forward;
- Constraints: capacity limits, staffing availability, and downstream bottlenecks.
By grounding itself in physical realities, Patient Journey Analytics not only provides accurate, real-world intelligence, but also a clear chain of reasoning. For example, estimated discharge predictions are easily explainable: users can understand why this discharge date is likely, which milestones remain, and which factors are most likely to change the timeline. This same understanding informs discharge disposition, flagging patients when they are likely to require prior authorization or face non-clinical delays.
Importantly, this world model also allows ancillary teams to prioritize work dynamically. Medically urgent tasks are always addressed first, but beyond urgency, procedures are sequenced based on which actions will most effectively resolve care barriers and unlock the next phase of a patient’s journey.
At the system level, Patient Journey Analytics also supports ICU stepdown planning, identifying which downstream units should prepare for incoming patients, and unit census forecasting, highlighting when patient-to-staff ratios are likely to become unsafe. This enables teams to intervene before risks and dangers arise.
Kontakt.io has long understood LeCun’s analysis, that the future of AI will belong to reality-based (rather than language-based) systems. Like LeCun’s world models, Patient Journey Analytics understands hospital operations deeply, and provides a real-time, real-world representation for technologies and teams to work off of. Importantly, it also helps teams shift their operations, actions, and mindsets from reactive firefighting to proactive, intelligent orchestration.
