Teton's Cassandra is a world model that forecasts falls in care homes

Fall prediction is the first use. In one example Teton published, Cassandra flagged elevated fall risk more than three minutes before the fall happened[1]. Teton says nobody programmed the warning signs in; the model learned on its own which residents are generally at higher risk, and how behaviors such as repeated attempts to get out of bed, changes in gait and "furniture surfing" raise that risk[1].

It also picked up that where a walking aid sits matters for some residents. That lines up with Teton's earlier study of 2,000 falls, where two-thirds of falls among walking-aid users happened while the aid wasn't being used[1][2].

Because Cassandra generates future timelines, Teton can test changes before they happen. In one demo, adding a two-hour afternoon nap to a resident's day moved their simulated bedtime from 19:22 to 20:09, with shorter, more fragmented sleep[1]. In another, inserting a fall with an injury into the previous day produced a next morning with 43 minutes in a wheelchair, a staff visit of more than eight minutes instead of 32 seconds, and four staff-assisted transfers instead of none[1].

• Behavioral tokens, not video. Teton's spatial AI runs locally on edge devices in each room and records activities, falls, wellbeing and staff interactions without wearables. Those observations become compact tokens that can stand for anything from a brief transition to hours of one state, which lets the model learn patterns over days or months[1].
• Next-event prediction. Like a language model, Cassandra is trained to predict the next event and when it will happen, then samples repeatedly to build long-horizon timelines[1].
• Mamba backbone. It uses Mamba, a selective state-space model whose cost grows linearly with sequence length, so it can hold far longer histories than full Transformer attention[1][4]. A dedicated head scores fall risk, and a dwell-time head predicts a distribution over how long until each possible next event[1].

Training on progressively larger datasets, Teton found prediction loss follows a power law with an exponent of −0.099, close to the −0.095 reported in early language-model scaling work[1][3]. Teton notes the experiments aren't directly comparable, and that more compute alone hit a plateau: what mattered was more behavioral data, which can't be scraped from the internet and needs real deployments[1].

World models are mostly discussed for robots and self-driving cars. Cassandra is a reminder that the same recipe works anywhere you can turn the physical world into a stream of events, and that the moat is the deployed sensors producing that stream. Teton says its goal is to move care from reactive to proactive[1]. The results so far are the company's own examples, so outside validation of how well the fall warnings hold up across real care homes will be the thing to watch.