Killing GPS, Jensen's Bets, Chinese Robots, & More

TLDR;

This week, we examine a new use case for AI that could kill GPS, Chinese robots, the main takeaways from Jensen Huang’s GTC keynote, new ChatGPT and Gemini features, Mistral’s latest model, and more.

Enjoy!

NAVIGATION

AI Kills the GPS?

In what could be one of the coolest AI use cases I’ve ever seen, the team at Tern AI has developed a navigation system that operates without GPS satellites. It’s called an Independently Derived Positioning System (IDPS) and uses the vehicle’s own sensors and stored map data to determine where you are.

Instead of pinging satellites, the system “derives a position from nothing” by analyzing movement and map information on the device itself. It doesn’t send out signals; all the location calculation happens on-board, which the company says improves privacy (no external tracking).

In other words, the model just needs a connection to your car’s Bluetooth (it handles vehicles from 2009 and above) and map information. Without any external information (it’s a closed-loop edge compute system; everything the system needs is on board, including sensors and map information), the AI figures out where you are.

In the demo they gave to TechCrunch, the model determined the position of a 2019 Honda Civic using a cached map from Austin and the vehicle’s sensors.

And it worked.

TheWhiteBox’s takeaway:

This could provide reliable navigation in places where GPS fails (like tunnels or urban canyons). It also can’t be jammed or spoofed because it doesn’t rely on satellite signals.

But why does this work, like, at all? All AI models do is map an input variable/s to an output/s.

In AI parlance, AI models simply fit a parametric curve between a series of inputs (the car's sensors and the map information) and an output (a (x,y,z) position in the world).

Tern AI must have figured out a causal relationship that predicts a vehicle's position with high certainty simply by examining the car's sensory and map inputs. So, the model finds hidden patterns in the sensory data that, correlated with map features, can only be possible in position (x,y,z).

In other words, the model has found a function that maps sensory data to position (if we see this particular sensory data and map information, there’s a decent likelihood we are here).

Fascinating.

This is very interesting because it clearly prevents several issues with traditional GPS systems. Most rely on triangulation (the commonly used term; in reality, the word is trilateration), in which the system figures out your position by calculating the distance between your car and three or more satellites, estimating that distance based on signal time, and then computing the most likely position based on solving a set of distance equations for each satellite. This is why your GPS struggles in areas with bad signals.

Additionally, TernAI has a political play, too. As the article outlines, the Trump Administration wants to find alternatives to GPS systems, so national incentives exist to make their business work.

ROBOTICS TAKEAWAY

Chinese Robots Are Really Good

Enjoyed the introduction? Subscribe or sign in to continue with this edition.