Where an LLM "predicts the next sentence, word, what's called a token," a world model "uses the same architecture to predict the next most likely event." For Fluency it represents "how work gets done in a business" — graph-based — so you "put work map in and it'll predict what's going to happen next. Work map out."
More from this episode
Finnlay distinguishes it from large language models: a world model "uses the same architecture to predict the next most likely event," and "it's not representing text. It'll represent something like a video game or represent a manufacturing floor." For Fluency, "it represents how work gets done in a business. So, it's graph based."
The payoff he describes: "effective super agents on top of your process" that can run workflows "fully autonomously," and the ability to "put in like what a future state looks like" and predict "how is it going to affect the team? How's it going to affect your revenue?" He believes the best way to do this is "a sovereign world model per customer," and notes one investor "heads up AI at Princeton" and is "super focused on building out world models." Timeline: "1.5 8 to 12 weeks," act two "9 months."