Luca

Software Engineer at NVIDIA | Focused on AI, robotics, and simulation. Exploring the intersection of intelligent systems and emerging technologies

[Luca @luca_moreti_]

Introducing NVIDIA Pets, a new generation of AI-powered robotic companions with their own personalities, behaviors, and memories.

NVIDIA Pets interact with the real world, respond to their surroundings, recognize people, and learn from interactions.

Think of them as the next generation of pets.

No feeding, no cleaning up, no vet bills. Just an intelligent little companion that learns, grows, and develops its own personality.

Meet NVIDIA Pets. Your new best friend might just be a robot.

[Luca @luca_moreti_]

if a robot can fall over, recover, and keep going, that's already a pretty good metaphor for learning software engineering.

[Luca @luca_moreti_]

I've been thinking about why little robots get people excited in a way that giant industrial machines don't.

It isn't because they're more capable. Usually they're not.

It's because you can imagine living with them.

A small robot walking around your desk creates an entirely different relationship with technology. Suddenly you're not thinking about an assembly line or a research paper. You're thinking about what it could learn, how it would respond to your voice, and what it might do when nobody is giving it instructions.

Microduck is a good example of that shift. Pollen Robotics took reinforcement learning, locomotion and open-source hardware and packaged them in something approachable.

There's still an enormous gap between a cute demo and a reliable everyday companion. Batteries die. Sensors fail. Floors are unpredictable. People do unexpected things. Real-world safety matters more than a smooth animation.

But I love the direction.

Affordable hardware gives more developers something physical to experiment with. Open software makes it possible to share improvements rather than starting from zero.

If robotics becomes a normal hobby for curious programmers, the next generation of useful personal robots might come from people building in their bedrooms instead of giant labs.

[Luca @luca_moreti_]

the most underrated AI feature might be a really good cancel button.

[Luca @luca_moreti_]

inference optimization is basically the art of making people wait less without sacrificing the part that actually makes the model useful.

[Luca @luca_moreti_]

Pollen Robotics' Microduck has 15 motors and stands about 25 cm tall.

small hardware, surprisingly big control problem.

https://t.co/fNmUbnYMEB

[Luca @luca_moreti_]

i would rather read a developer post about a weird bug they finally solved than another thread announcing that AGI is 6 months away.

[Luca @luca_moreti_]

an agent remembering something is not the same as an agent understanding whether that memory is still correct.

this becomes a huge problem once you let it act on your behalf.

[Luca @luca_moreti_]

RTX Spark makes me think about what software looks like when local AI becomes a normal feature of your machine, not a cloud service you bolt on afterward.

[Luca @luca_moreti_]

one of the funniest things about robotics is that a simple movement in a demo can represent thousands of failed training episodes.

[Luca @luca_moreti_]

the engineering question i'm most interested in right now:

how do you make an AI assistant feel instantly responsive without burning through unnecessary compute?

[Luca @luca_moreti_]

a lot of people see a cute robot and immediately ask what useful job it can do.

sometimes the useful thing is making robotics accessible enough that thousands of new developers actually want to learn it.

[Luca @luca_moreti_]

I think we're getting the definition of a personal computer wrong.

For years, a PC has been something you operate. You open applications, move files around, search for things, memorize shortcuts and work around the limitations of software.

What happens when the computer starts understanding the goal instead of waiting for every instruction?

That doesn't mean handing an AI unrestricted access to everything. Quite the opposite. A useful personal agent needs clear permissions, predictable behavior, memory you can inspect, and an easy way to undo mistakes.

I'd much rather have a modest local model that handles five everyday tasks reliably than a giant model that can write a brilliant essay but fails to rename a folder correctly.

This is why NVIDIA's push toward local AI hardware is worth watching. More compute and memory on personal devices makes it practical to experiment with on-device inference and persistent assistants.

The hard part is still the software. We need better tool interfaces, better evaluations and much better failure recovery.

A computer that understands your intent without taking away your control would be a genuinely meaningful step forward.

[Luca @luca_moreti_]

Microduck having an articulated beak is such a fun design choice. It's a cute little robot, but that beak also makes object interaction possible.

[Luca @luca_moreti_]

there's a huge difference between giving an agent access to your computer and designing an agent that knows when to ask for permission.

the second one is an actual product.

[Luca @luca_moreti_]

NVIDIA's new Nemotron 3.5 lineup is interesting because different workloads need different model sizes.

A local conversational assistant and an enterprise research agent shouldn't have identical hardware requirements.

https://t.co/6cKxXqHCHs

[Luca @luca_moreti_]

the future of desktop AI probably looks boring at first.

less clicking around, fewer repetitive steps, more useful context. that's where the value is.

[Luca @luca_moreti_]

my favorite direction for AI isn't bigger chat windows.

it's small, persistent tools that understand the task you're working on, request permission when necessary and quietly help you finish it.

[Luca @luca_moreti_]

watching a robot learn to walk through reinforcement learning is weirdly satisfying.

all those failures in simulation, and eventually you get a tiny machine that can move around without falling over every three seconds.

[Luca @luca_moreti_]

a smaller model that responds instantly can feel more intelligent than a bigger model that makes you stare at a spinner.