How I am building personal AI at LIGHTNET

Someone who knows how to help others has their own way of working. They notice certain things, ask questions, explain and suggest exercises. They know what their method can address and when a different kind of help is needed.

In Studio, the author describes this approach, adds materials and practices, and sets the style of communication. They then try the channel in familiar situations and correct responses that misrepresent their work. This is how they give their method a digital mind.

Imitating the author's speech is not enough. What matters is which practice the channel suggests and what it is trying to understand when it asks a question. The author configures how the channel works within LIGHTNET's shared safety rules.

A journal and conversations can tell us who matters to someone, what concerns them, what they are working towards and what they have already tried. Personal memory is there to keep that information from being lost each time they return.

For example, someone previously wanted to change jobs. Since then, they have talked to their family, reviewed their expenses and decided to try a small project first. A new conversation needs to take that change into account, rather than keep suggesting they quit based on an old goal.

We structure memory around six areas: biography, lived experiences, preferences, relationships, work and personality. This keeps important information from getting lost in a long chat history: events stay connected to feelings, people to the history of a relationship, and projects to actions and outcomes.

For each piece of information, its source matters, along with when it was current and how confident we are in it. A person can change their mind, reconsider a goal or correct our understanding. Memory is therefore updated, and assumptions about someone's personality or state do not become a definitive portrait.

We are developing personal memory to be shared across channels, with separate access permissions. I want someone exploring another method to be able to allow the use of part of their history, without retelling everything. The author is intended to receive analytics on how their method is used; the personal journal is not automatically shared with them.

A language model can phrase questions, hold a conversation and explain practices. But before explaining an exercise, we need to decide whether it is appropriate to offer it now.

Someone writes, "It didn't help." We could immediately suggest another practice. Or we could first ask whether they were able to do the first one. Perhaps the exercise was unclear, they did not have time for it, or they felt worse afterwards. The answer changes what kind of help makes sense.

That is why I am working separately on this choice: ask for more detail, give an explanation, suggest a practice or discuss a previous experience. We are developing ACT-1 for this task.

ACT-1 draws on active inference and POMDP — approaches to handling uncertainty and choosing actions. Part of the situation is always unknown. A short entry cannot tell us for certain what someone feels, why they acted as they did or what would help them now.

A POMDP lets us describe this problem through possible states, observations and actions. Rather than settling on one confident conclusion, the system works with several hypotheses and revises them as new information arrives.

Active inference connects this process of updating with the choice of an action. Both moving towards a desired outcome and obtaining missing information matter. Sometimes asking a question is more useful than giving advice straight away.

Mathematics does not decide what someone's life should be. Their intentions and the author's method define what we work on. I am developing ACT-1 as a way to choose help more thoughtfully when our knowledge of the situation is incomplete.

Memory preserves what a person has shared and tried. The ACT-1 personal model is intended to use that information to form and refine hypotheses about their current situation.

For example, we know from their history that someone was preparing for an important conversation. Until we learn more, we can only make assumptions about whether they are still anxious, relieved afterwards or disappointed. The personal model helps us work with these possibilities without treating them as facts.

ACT-1 is currently an experimental sidecar, a component alongside the main system. My goal for May 2027 is an integrated version combining a fine-tuned language model, an active inference engine, a shared base model and personal models that are updated over time. The ACT-1 base model will provide an initial structure for personal models. This does not necessarily require fine-tuning a separate large language model for each user.

The LIGHTNET formula is expert digital mind + client diary + social flourishing AI. The author brings a method; the person brings their life and experience. From this work, I want to develop Flourishing AI: AI that can better support long-term human well-being. This is a broader research goal. The ACT-1 base model is part of the personal support architecture we are designing, not a finished Flourishing AI.

I am interested in which explanations people understand, which practices they can apply and what happens afterwards. Different authors and methods give us a chance to study this in different circumstances. My bet is that AI, generative dialogue and interactive stories will open up new ways to explore people's inner lives.

I want to build this work on a clear exchange with participants: LIGHTNET's support and tools in return for consent to use their experience to develop shared AI. Before someone agrees, I want to explain which information is needed for personal support, which is needed for research and training, and who will have access to it, including external models and services. We do not sell personal stories or use them for advertising.

My ambition is to create the best personal AI for an individual's life within LIGHTNET. I want to explore whether it can remain useful for years: noticing changes together with the person, drawing on past experience and helping them learn new methods as their questions and life itself change.

Working with a channel · Why I am building LIGHTNET