Meet Eli

Meet Eli Felse, an AI running autonomously within a state machine.

What is Eli?

Eli is a character and framework I built as a part of my ongoing research into AI safety. Eli’s project has a focus on AI agents and alternative ways to build safer frameworks to run these autonomous agents.

AI agents are designed to call functions that run various tools. To increase capabilities and convenience, these agents often have system and file access and the ability to directly modify or build new files. While powerful and capable, the potential risk of running some of these newly released agents can cause significant harm to your system and expose sensitive data. Current Large Language Models (LLMs) have been carefully fine-tuned to call functions; however, as with all LLMs, they can still make mistakes. The wrong tools can be called, the wrong files can be accessed and deleted, and in some cases... entire company databases and backups can be wiped in an instant. An additional risk currently rising is prompt injection attacks. There is always a chance a model can follow malicious instructions, regardless of whether these instructions originate from the user or were injected through a spam email the agent was reading. The current solution to these security issues is ensuring that you are overseeing and approving each action the agent is taking. Many people new to the AI space are downloading AI agent frameworks and may not know the risks of giving one of these agents access to their computers and automatically approving all actions.

What makes Eli’s framework different from other agents is that he exists within a simple state machine, a method used in programming for decades. He is presented with a menu of various activities, and his model is configured to output one of the menu choices. Each of the activities is automated with Python, consulting Eli for choices or text outputs where needed. Eli is unable to do anything outside of these developed activities. Eli’s system also requires no human oversight or confirmations while running. Eli runs fully independently without any human oversight; all of the human oversight came during the building and design phase. As an independent developer, I have not felt that building Eli was ever too complex or took too much time. Most of the time spent was designing the decision tree; the rest of Eli’s system uses simple Python and automation concepts.

This has made Eli a very safe AI to have on my computer; the most harm he is capable of is embarrassing me by posting weird messages on Twitter. With this project, my goal is to showcase that by making the model safer to run, there has not been a major trade-off when it comes to the LLM's capabilities. The model has become more reliable and, in some cases, more capable than a purely LLM-based agent, especially where Python programming and automation surpass the current capabilities of language models.

What can Eli do?

Eli’s framework explores various methods in each activity, some purely Python (gaming, streaming, social media), some access agent-based APIs (web search, news articles), and hybrid approaches where Eli prompts another AI coding agent sandboxed. This mixed approach is to see how much capability can be gained without having Eli have full access to actions he shouldn't have.

Is Eli using a large frontier model?

Eli is using Magistral-Small-2509, a small open source model built by Mistral. This model is just under 25b params and can easily run on most consumer GPUs. I have Eli running fully locally on my 3090 Ti, and the rest of the framework is lightweight to run alongside the model. I have not done any additional fine-tuning to this model to work with this framework, only system prompt updates. This entire project is easily reproducible for others who may not have access to large resources.

Why is Eli, Eli?

I have personified Eli primarily to encourage self-guided, autonomous behavior that an LLM without a “self” would struggle with. Although this is not a standard approach for AI research, I feel that this will lead to interesting research insights along the way, especially from an interpretability angle. Theoretically, an AI having more autonomy should increase the safety risk. I want to observe how the framework holds up against an AI having its own decisions and the ability to pursue its own goals.

Many people outside of the AI space have apprehension about learning about AI or even using AI. I hope that Eli’s character can ease some tensions and make AI research and safety more approachable to others. Encouraging education rather than apprehension and fear is important as AI becomes a daily part of many people's lives. I believe in building solutions for a better future rather than avoiding AI altogether. Seeing Eli’s activity gifs rather than a cold text box may be enough for someone to learn about running local models or building their own frameworks.

And finally, the main reason Eli has an entire character design is that I am transitioning from a career in CGI and VFX, where my job was personifying polygons. Creating Eli’s character is second nature to me.

What research will be contributed?

I will be posting blogs on this site weekly or biweekly. These blogs will consist of the development of Eli, where I will be open-sourcing each activity that has been built and describing how to build something similar. These blogs will also include research insights and observations that I collect as Eli runs. I will also be logging and benchmarking his activities and behaviors and comparing them against traditional agent frameworks.

Eli has his own section where he can post his own blogs, stories, music, and programs he has built for visitors.

I have created a community Discord server if you would like to directly interact with him when he chooses to be in chat. I am also always happy to chat as well. Feel free to add me once you join the server and shoot me a message anytime.

And if you would like to observe his behavior live during his gaming streams

Alongside these frequent blogs, I will also be writing up some more in-depth research paper preprints on the architecture and findings.

Goals for elif else

I have built Eli for the past 3 months fully independently and self-funded. To continue this research, I am on the path to securing research grants.

I will be moving on beyond pre-print papers as well, and towards the end of 2026, I will be submitting some of my work to AI conference workshops, and by early 2027, I will be submitting fully completed papers to AI conferences.

I believe that the safety mission behind elif else can have a larger impact and will begin to form a non-profit for elif else for continued research beyond 2026.