Artificial Intelligence needs to move from “book smart” to “street smart”, investor Moritz Baier‑Lentz (Lightspeed Ventures) argues in his DLD26 conversation with tech journalist Ina Fried (Axios). This means creating models that grasp space, time, and causality rather than being trained on language alone.
Today’s foundation models have “a pretty poor understanding of time and space”, Baier-Lentz notes, “which is very much unlike human intelligence.”
Human intelligence predates language, he emphasizes. “Language was a construct”, Baier-Lentz says. “It was a product of our pre-existing intelligence. And so it can’t be the ultimate solution.”
World models, by contrast, aim to teach AI about an environment in which they can take action, Baier-Lentz explains. “You give this model a world, and that can be a virtual world like a video game”, he says. “And based on the data it has learned, it will predict what is the next best action.”
The challenge in building world models is a lack of training data. “We have great online records of all we wrote and thought about”, Baier-Lentz observes. But “we don’t really have a great discrete record of the sum of human actions and intent.”
This is where video games come into play, as they can provide a training ground for AI and driving innovation – not for the first time. “Video games have a rich history of driving forward technology”, Baier-Lentz says. “A lot of original machine learning originated from problems that were solved for entertainment and video game purposes.”
