Comment by reissbaker
11 hours ago
I'm confused by both this blog post, and the reception on HN. They... didn't actually train the model. This is an announcement of a plan! They don't actually know if it'll even work. They announced that they "trained over 50 million neural networks," but not that they've trained this neural network: the other networks appear to just have been things they were doing anyway (i.e. the "Virtual Positioning Systems"). They tout huge parameter counts ("over 150 trillion"), but that appears to be the sum of the parameters of the 50 million models they've previously trained, which implies each model had an average of... 3MM parameters. Not exactly groundbreaking scale. You could train one a single consumer GPU.
This is a vision document, presumably intended to position Niantic as an AI company (and thus worthy of being showered with funding), instead of a mobile gaming company, mainly on the merit of the data they've collected rather than their prowess at training large models.
“Concepts of a plan” is often enough to make people think you know what you’re doing. Think most people, here included, got the impression that they had succeeded already.
And I get that; one thing that (I think) especially software developers have is a high level knowledge of many different subjects, to the point where IF they ever have to do something in practice, they'll know enough to figure it out. T-shaped people kinda thing.
Maybe its because the current HN title says "trained" in the past tense?
I'm annoyed that the HN title is still editorialized after multiple hours. It's not like the original title is especially egregious.
Ok, we've replaced 'announces' with 'plans' in the title above. Thanks!
After skimming the article I don't understand what the Large Geospatial Model is supposed to be, what it can be used for.
They have a "VPS" which extracts keypoints from an image and matches them against a 3d pointcloud. Using trigonometry you can work out the 3d position of the camera by matching the keypoints from the image to the keypoints in the point cloud.
What is different is that they are proposing to make a large ML model to do all of the matching, rather than having a database and some matching algorithm.
Will it work? probably, will it scale? I'm not that hopeful, but then I was wrong about LLMs.
According to their website :
Neural mapping. Taking 5 minutes to build a scene in space.
Relocalization estimating camera pose from a single image.
https://nianticlabs.com/news/largegeospatialmodel?hl=en
It looks pretty cool. I imagine it could be a game changer in wearable devices that want to use position like AR.
Intelligence gathering is also another one. Being able to tell where someone is based on a Picture is a huge one. Not just limited to outdoors but presumably indoors as well. Crazy stuff
They have never been a mobile game company and they have said as much themselves on many occasions. They're a data harvesting company. Guess now they're trying to figure out what to do with all of that data.
From what I've heard only the minority of the company works on the games. The rest does 'research'.
> They're a data harvesting company.
I didn't realize this. I used to submit so many "portals" to ingress T_T
Thank you for your service. The robot who will replace DoorDash workers will prioritize your food deliveries higher.