Feel-Good AI Stories

Dr. Ryan Ries here. The news has been pretty heavy recently, so I felt we could all use a palate cleanser this week.

So, for today’s Matrix, let's discuss what’s been happening lately in the interesting, wonderful, and occasionally broken world of AI.

Germany's Underwater Janitor Robot 

German researchers just deployed an AI-powered submarine in Marseille's port that collects underwater trash with a four-finger robotic hand. The submarine can lift a whopping 551 pounds and has enough grip strength to crush plastic buckets, but it has smart enough sensors to handle glass bottles without breaking them.

The Technical University of Munich built this as part of the EU's SEACLEAR project, and the engineering is seriously impressive. 

The system uses a fleet approach: 

• An unmanned service boat with sonar mapping
• A small search robot with turbines
• The main diving robot with its grabber hand.

What is really interesting is that they had to train the AI on underwater garbage identification with almost no existing data. The team manually labeled 7,000 images of "things that don't belong on the seabed, " such as scooters, fishing nets, tires, and broken glass. The AI then converted these 2D images into 3D models to determine optimal gripping points. 

Unbelievable! But it doesn’t stop there.

Let’s not forget about the grip regulation. Four fingers generating 4,000 Newtons of force (enough to lift quarter-ton objects), but with pressure sensors that adjust grip strength based on material detection. Plastic bucket? Gentle squeeze. Metal beam? Full force.

The downside is that the robot's battery only lasts two hours. They keep it tethered not just for power but because the cable connection boosts AI processing performance. The 264-pound submarine needs constant power for its eight mini turbines and onboard compute.

Stefan Sosnowski from TUM made an interesting economic observation: autonomous underwater waste collection only becomes profitable at depths beyond 52 feet. Below that depth, human divers are still more cost-effective.

My knowledge of underwater trash clean-up is pretty slim, but I was honestly surprised to learn how complex it actually is. To pick up underwater trash, you need:

• Sonar mapping from the surface
• Visual recognition in murky water
• 2D to 3D model conversion
• Material classification for grip adjustment
• Coordinated fleet operations
• Real-time processing power that requires a tethered connection

Each piece of garbage requires an entire technology stack. We've built a Rube Goldberg machine for picking up beer cans.

Yet this might be exactly what we need. The oceans are full of trash in places humans can't efficiently reach.

Moflin

Casio just launched Moflin, an AI-powered pet.

This furry blob has no legs, no face (just eyes), and makes sounds somewhere between a cat's purr and a hamster's squeak. It develops one of 40,000 possible personalities based on how you interact with it. Ignore it, and it becomes anxious. Pet it regularly, and it grows affectionate.

I was honestly surprised to see the price point. $400 seems like a lot to me for a furry Tamagotchi.

Casio created a subscription service for your robot pet. For $45 per year, you get "health checkups," fur grooming at their "salon," and even a "revival service" if your Moflin dies (assuming you backed up its personality to the app).

Before you laugh, consider what Casio has actually built: an emotional state machine that adapts to human interaction patterns. The Moflin remembers who interacts with it most, develops trust relationships, and modifies its behavior accordingly. 

To me, the Moflin is a fun idea. My wife and daughter have allergies and can’t have pets, so it could be fun to have a pet that doesn’t cause my household to sneeze incessantly.

Meta's Smart Glasses Faceplant 

Lastly, if you haven’t seen the Meta smart glasses demo videos online from their recent developer conference, you should check them out. Every one of us who works in tech has probably had a demo fail us at one point or another. But this was on another level… 

Multiple demos failed big time. Live. On stage. With Zuckerberg up on stage, too.

Andrew Bosworth, Meta's CTO, went on Instagram live and explained what went wrong.

The first failure happened when a chef said, "Hey Meta, start Live AI" to his Ray Ban Meta glasses. That command activated every single Ray Ban Meta in the building. Hundreds of glasses simultaneously trying to connect to Meta's development server.

The second failure was even more subtle. During a WhatsApp video call demo, the display went to sleep at the exact moment the call came in. When Zuckerberg woke it up, the answer notification had vanished. This was a race condition bug they'd never seen before, appearing at the worst possible moment.

I appreciated Bosworth’s straightforward response to the flop. No corporate speak. No blaming the Wi Fi (even though that's what they initially said). Just straight technical analysis and acknowledgment of what went wrong.

My Take

We're in this messy phase where AI can beat humans at certain tasks but can't reliably answer a video call. Where companies spend billions on training infrastructure while selling robot hamsters.

The path to useful AI isn't a straight line from chatbots to AGI. It's this wandering journey through failed demos, underwater janitors, and robot hamsters. Each experiment teaches us something about what works, what doesn't, and what people actually want from AI.

Keep experimenting. Keep building. And maybe get a Moflin to keep on your desk?

Until next time, 
Ryan

Now, time for our AI-generated image and the prompt I used to create it.