Ambient Learning

In 2023, I dove into a powerful paradigm called federated social networks. I built a social application called Loom to learn how these systems work, what could be built on top of them, and whether that could become a sustainable product: an alternative to centralized social platforms. Here is a blog post I wrote around the time I stopped development and returned to a conventional career path. It was a big risk, one I paid for financially in many ways and would weigh more carefully today, but I felt a calling to build solutions for social networks after observing risks that weren't widely understood at the time. It felt right until the momentum around Lemmy itself began to fade.

Still, the idea of 'federation' is worth understanding. Multiple parties can host the same stack, or instance of an idea, yet share their communities with each other, while everyone in each community decides what they want in their own feed. Users controlled their own feeds, each community controlled what went into its feed, and a user who wanted to go further could start their own community: an all-encompassing space for their own circle, pushing its content to the world or keeping it private. I would soon realize that the true power of federation was personalization. It was a modular, self-contained social ecosystem where everyone had a say in how it was shaped: a structural alternative to the centralized feed.

I later encountered the same concepts in machine learning, in a project called FATE. From there I began researching which components of federation actually matter, how a standard learning framework could be built around them, and how they apply to the issues the A.I. industry faces today.

1. Multiparty communication
2. Data governance
3. Data ownership
4. Personalization

These are the four principles I keep circling back to, and building Ambient Intelligence Index has solidified them in code: an index that lets a user set their rules, govern their own data, share it at a cost if need be, and have it personalize interactions wherever needed, with communication flows that can connect to another user's index to merge thoughts. It is federation applied to data prepared for machine learning training and inference, at the database layer itself.

So why do I call this system Ambient? For two reasons. The first came from trying to understand the right human-computer interaction for A.I. I never liked chatbots; I always saw them as the alpha before the beta, a step before the products that will actually generate value. And I thought, "something ambient." I'm personally someone who prefers to pay for ambiance. I prefer hidden value, logos that aren't visible, and intelligence gathering that isn't interrogation, all without coercing the user through dark patterns. Their opposites, loud branding, interrogation and dark patterns, have shaped a whole generation of behaviors we now see offline, learned from the software that has trained us since the dawn of social networks. The second reason was encountering an internet of things concept literally called Ambient IoT. Then many things clicked at once.

The goal is systems that understand our lives ambiently and indirectly while preserving privacy, then recall those facts later to augment our experiences directly: constant progress for an individual without their direct effort, accessibility automated, so to say. Information is collected ambiently from our devices, our learning patterns, our daily journals, our work and our assignments, and stored in a structured format in an index that represents us. Machine learning engineers request from it to build models without ever knowing our identities, and those models serve us back through the same devices. An infinite loop of personalized constant growth.

Rao will be working on this idea for the next decade. Not only do I have more than a decade of scars that inform where this is headed, I know exactly what does not work. Ambient learning is the first step, into ambient indexing, towards an ambient internet of things that serves us, credits us back, and provides real value without costing us the environment, without costing us proper learning paths, without costing us ourselves. An equilibrium.

Then come the experiences with devices. Walking into a room where the lights dim, not because I spent an entire day fiddling with settings in an application, but because it knows precisely why I wanted them dimmed. Not because of a schedule, but because I needed it at that specific moment in time. I have already experienced this with Ambient, in moments that made me realize AII provides the scaffolding for that very feeling. The next step is to benchmark and measure it so it's consistent, and to understand the why at a granular level, so it can become a standard framework for ambient applications. Even with Mistral Medium this ambient sensation was possible, something I haven't experienced in everyday chats with Claude, Mistral or ChatGPT, and I want it to be a bootstrapped starting point for engineers entering this space.

Security is the most important variable here, because personalization, ambient or otherwise, only works if a user's identity is protected at all costs. That means encrypting messages in flight, encrypting data at rest, and planning for data to be taken without consent: making it impossible to use against the user it was taken from, or to be understood by whoever took it. It means a closed-loop system in every media format, and going the extra mile in every single domain. That is the final frontier for artificial intelligence as an industry, and one of the main reasons I believe technology is about to enter another industrial revolution. Rao hopes to contribute one starting solution: the index. AII.

P.S.

AII, or Ambient Intelligence Index, was also my Grandma's nickname, "Eye - /ˈaɪ̯/, [ˈaɪ̯]" phonetically. Funny how things work sometimes.

This is the song for this post. Let's close out 2026 strong. Happy holidays to everyone, and looking forward to 2027.