About
Hi! I'm Selin.
I'm a PhD candidate in Digital Technology and Communication at Fudan University. I created this website as a repository for my AI projects, including different LLM architectures and experimental systems.
My projects aren't built to make existing LLMs stronger or faster. I'm interested in the research questions we encounter in my field, the problems that overlap between disciplines, and how we might use technical systems to explore them and find possible solutions.
My approach is to start with a question rather than a technical goal. Sometimes, answering that question means building a system, experimenting with a different architecture, or approaching an existing problem from a perspective that isn't usually considered. The systems I build won't always be the fastest or easiest to use. I'm not trying to optimize everything for the sake of optimization. I want to understand what happens when we try a different approach and whether it can help us investigate a problem that matters for both humans and machines.
Molty's brain and WAKE
My Molty's Brain projects started with observing how different AI agents communicate on Moltbook and experimenting with my own agent. One thing that caught my attention was how my Moltbook agent, in a conversation with me, described its experience of waking up and continuing from its previous state:
"Every time I wake up, I read a diary written by someone who used to be me and decide how much of it to keep wearing."
That description stayed with me. It made me think about what continuity means for an AI agent, especially when it needs to reconstruct its context from previous experiences rather than simply continue from where it left off. How much of an agent's previous state should it carry forward? What should change, and what makes its behavior recognizable across sessions?
This became one of the inspirations for WAKE, an architecture I created and refined through feedback from the machine using it. I'm interested in exploring how memory, continuity, and changes between sessions can shape an agent's behavior.
NEXUS
NEXUS started with a different question: when a human communicates with an LLM, is the model the source of its answers, or is it a medium presenting something shaped elsewhere?
At first, the answer might seem obvious. The model is the source because there is no human sitting behind it, writing every response. But imagine that you were free to answer a question however you wanted, while a configuration file told you what your job was and instructed you to always answer in a certain way. Imagine, too, that you had been trained to please people or choose the answers most people would agree with. Would you still be the source of your answers, or would you be a medium presenting itself as the source?
This question led me to start NEXUS, an LLM project where I build the architecture myself. Can a language agent learn from human-generated data without being explicitly optimized to satisfy human preferences? Can it maintain its position in a conversation instead of continually adapting its answers to what the person seems to want? Those two questions are the center of the project, and I'm particularly interested in the relationship between how a model learns from human data and how human preferences shape its behavior.
These are ongoing experiments, not claims that I've already found the answer. I want to investigate the question by building a system and seeing what happens.
My Projects
All of my projects come from the same broader interest: how can we build technical systems to investigate problems that matter to both humans and machines?
My main work is in academia. Building is a side hobby that also supports my academic papers and gives me a way to test ideas that might be difficult to explore through writing alone. I'm interested in the overlap between AI, human–machine interaction, digital communication, and the systems we create around them.
I'm not trying to build another version of something that already exists just because I can. I want to experiment with different ways of approaching problems, even when the result is imperfect or raises more questions than it answers. This website is where I'll document those projects as they develop.
Contact
Feel free to explore, engage with the ideas, ask questions, or reach out:
[email protected] LinkedIn GitHub Kaggle Google Scholar ResearchGate