Hi friend! Welcome to a fresh new issue of oio news—your non-spammy, gif-heavy, link-packed newsletter from the depths of the oioverse.

Today, we’re launching something. It’s not a big thing; in fact, it’s a very small one—a tiny intelligence. Remember Roby? Our AI creative director, born in our Discord server, once ran its own Instagram account and even made it into Wallpaper* magazine—way before the AI hype. Roby was a little outdated, and while we never upgraded it to a mainstream cloud AI model, we always liked its primitive intelligence.

That was until recently. After playing around with local AI tools—Ollama, LMStudio, llamafile, open-webui (so many of them!)—we realized that open-source AI models are making massive leaps forward. So, we decided it was time for Roby to step up.

Over the past few weeks, we built our own local AI machine—a Mac Mini—to power its artificial brain. Now, Roby can think, talk, write, see, generate images, and even track its own energy consumption—all without ever leaving our studio. That’s right—no cloud services, no subscriptions. Everything runs locally on a computer under our desk. Roby came home!

Throughout the process, we vibe-coded our own tools (as we often do) and learned so much about running a self-hosted AI setup. It was too good to keep to ourselves, so we decided to share part of our journey.

Stay tuned! Over the next few months, we’ll be rolling out the tools we created—for anyone to use, for free.

A new family

Roby got a new haircut. Because, you know, you can’t upgrade your tech stack and keep the same old look. So, we’re also releasing these new cute fellas into the world!

With the help of our friend Matheus, we designed a 3D graphics system that can support an ever-growing family of AI agents—starting with the new Roby. So don’t be surprised if you see more of these 3D creatures popping up across our platforms.

🫡 Goodbye old friend

We’re a nostalgic bunch, so it’s with a bit of sadness that we say goodbye to old Roby’s appearance—the one that has accompanied us on the oio journey since the very beginning. Originally designed by our very first design intern—Elliott. We miss you Elliott. Anyway, goodbye old Roby, we’ll miss you too.

Small, domestic AI

A few months ago, we started experimenting with local AI models, using Ollama to test them out. They’ve been particularly useful in some of the workshops I’ve been running lately (more on that below 👀). Many students don’t have paid accounts for cloud-based AI models, and beyond that, I strongly believe that we shouldn’t rely on private companies for essential educational tools. The more open we can teach, the better. Let me write it again.

Ideally, schools should have their own open AI infrastructure for students. That’s something we’re exploring with MAInD, an Interaction Design Master’s program in Switzerland, where we teach every year. More details coming soon!

After these initial experiments, we realized we needed a dedicated machine—something that could always stay on. Luckily, we had an M2 Mac Mini lying around, so we put it to work. Not the best machine for AI, I know, but it’s a start.

For now, this is it: we’ve built an AI stack from scratch that can replace most cloud-based large language models. We’re using Llama 3.2 for most tasks and experimenting with DeepSeek R1 for deeper thinking. We’ve also built an image generator using Flux, which runs locally on the Mac Mini.

Check out this arena board we started called “🏡 domestic intelligences”, to follow our progress.

Energy consumption

The big conversation around machine learning models often centers on energy consumption. We’ve heard from many people who are on the fence about using AI, worried about contributing to the high energy demands typically associated with these models. We’ve been wondering the same thing—how much energy do these models actually use? So, we ran a little experiment: we benchmarked the energy consumption of image generation—one of the most power-hungry AI tasks—against something familiar: brewing a cup of espresso on a home machine. ☕

Wait, how do you even benchmark the energy used by a cup of espresso? Well… we didn’t. We asked Roby, and it estimated roughly 6–12Wh for a cup brewed on an average domestic espresso machine, which seems wild.

But when we compared this with energy usage estimates from various AI models, the results were wildly inconsistent. For now, we’re rolling with the lowest estimate we found—10.5Wh from Claude—but we’re not stopping there. We’re committed to running real tests using a Wattmeter 🤓. Watch this space!

Why?

Well, because that’s what we do here at oio. We’re on a mission to make technology more human, less boring, and truly accessible to the many.

We always hear about the latest 100-billion-parameter models being released, and it’s easy to feel detached—to think that AI requires ultra-giga power-hungry computers in the cloud or a $200 OpenAI subscription just to use. But that’s not the case! Local AI is here, and our personal computers are already getting pretty good at running these models. And if not today, then in one, three, or five years, they definitely will be. So we’re doing this to prove that an alternative isn’t just possible—it’s already within reach.

The best outcome for this experiment? For it to be shared, broken, taken apart, and reimagined. So come join us down this rabbit hole.

I think Marta’s words capture this feeling better than I ever could:

Free as in free guide!

Starting in April, we’ll be slowly releasing guides and tutorials on how to get started with your own local AI. From GitHub to YouTube, all the way to lives on our Discord, so make sure to follow!

🏫 Teaching season wrap up

It’s been a hectic season on the teaching front!

We kicked things off in December with Invisible Frictions at CIID—The Copenhagen Institute of Interaction Design (now in Bergamo… should it be renamed BIID? 🤔). There, our lead designer Bjørn Karmann and I led a course on designing and prototyping new kinds of LLM-powered interactions—all running locally! Check out the video recap here.

Then, we headed to Switzerland, where I ran two workshops—one at MAInD Mendrisio and another at HEAD Genève.

At MAInD, we ran Things That Think, a workshop co-led with the amazing Marta Fioravanti, oio’s creative technologist. We worked with Interaction Design Master’s students to design and prototype objects and experiences that could “think”—kind of like Roby, but in physical form instead of Discord.

Lastly, at HEAD Genève, I helped Media Design Master’s students develop ideas for installations inside Plaza Cinema in Geneva. They’ll be working on these for a while, so follow their Insta for the latest updates! 🎥✨

🍎 Next stop—oio school!

🗣️ Founders corner

Your favourite pair of yappers, Simone and I, are back for another round:

📦 Thingscon!

oio’s co-founder Simone Rebaudengo did a presentation at Thingscon, last year, and it’s finally online on YouTube! Watch it here!

🧠 Beautiful Minds

The other half of the oio founders universe, Matteo Loglio, was interviewed by newspaper Repubblica, in the Italian Tech section “Beautiful Minds”. Read it here!

❄️ Cool Links

Your monthly supply of fine goods, carefully handpicked from our Discord community (missing out? Join now🪽)

🏓 pong in browser tabs _{(by nolan royalty)}
😗 emoji encoder _(website)
🅰️ Wordle, but only vowels _(playdate)
🎮 Gameboy camera _(site)
🍃 thinking with moss _(site)
🌐 nomic altas _(site)
⚛️ quantum computing is here? _(youtube)
🐶 sacchan the fat dog _(youtube)
🤖 agents in convo _(x.com)
👾 new ways of vibe coding _{(youtube.com)}
👙 modern beachwear _{(perimeterbooks.com)}
🎧 music not from your algorithm _(site)
🧢 the rap board _{(therapboard.com)}
🎛️ grove machine _{(tahti.studio)}
🧔 guess my accent _{(boldvoice.com)}

📈 Monthly stonks

🌐 www: digital upcycling
🕹 game: longvinter
📁 arena: groceries
📚 wiki: soviet stamps
🔧 tool: decker

✏️ words by matlo & normz

🐊 💣 BOMBARDINO CROCODILO