Niantic, the company behind the extremely popular augmented reality mobile games Pokémon Go and Ingress, announced that it is using data collected by its millions of players to create an AI model that can navigate the physical world.
In a blog post published last week, first spotted by Garbage Day, Niantic says it is building a “Large Geospatial Model.” This name, the company explains, is a direct reference to Large Language Models (LLMs) Like OpenAI’s GPT, which are trained on vast quantities of text scraped from the internet in order to process and produce natural language. Niantic explains that a Large Geospatial Model, or LGM, aims to do the same for the physical world, a technology it says “will enable computers not only to perceive and understand physical spaces, but also to interact with them in new ways, forming a critical component of AR glasses and fields beyond, including robotics, content creation and autonomous systems. As we move from phones to wearable technology linked to the real world, spatial intelligence will become the world’s future operating system.”
By training an AI model on millions of geolocated images from around the world, the model will be able to predict its immediate environment in the same way an LLM is able to produce coherent and convincing sentences by statistically determining what word is likely to follow another.
To me this implies that the navigation AI is going to hallucinate parts of its model of the world, because it’s basing that model on what’s statically the most likely to be there as opposed to what’s actually there. What could go wrong?
AI: Dave, turn right and walk across the bridge.
Dave : But AI, there is no bridge
AI: I am 99% sure based on 99 billion images that there should be a bridge
Dave: ok , you’re the smart one
Dave: aaaargh . . . .
SPLAT
Reminds me of Morrowind’s directions, with the frequent east-west mixups, and sporadic north-south mixups.
Fun fact, I worked with several other people on a localization patch for polish version of Morrowind, and we had so many of those east-west mixups fixed. Of course the publisher just translated strings and didn’t QA anything.
I presume the idea is to generate a base idea with ai then correct it with real time data.
Like the way go AI has one part to make a ‘policy’ of moves and a second part to simulate (‘read’) the results of those moves many steps ahead.
It’s only going to hallucinate until it gets new input from reality. Not nearly as precarious as generative models.