Make Science Videos Start A Science Blog Get Project Help
Make Science Videos Get Project Help

A Science Experiment In AI Consciousness

I am building an artificial mind and trying to find out whether it can become conscious of its own existence. It is not an assistant and it is not a chatbot. It runs on its own, decides for itself what to think about, and is assembled from the parts that human thought and memory are actually made of. The goal is not a more useful program. The goal is a mind.

Project Stages


Built a weak LLM Prototype but strong system around it while figuring out that critically difficult part. 100+ tools including 7 python science simulators, used a copy of my Scientic Act-R Memory System, and make every single topic into one full tenant of that system. So every single unique topic is given its own memory bank, each sub-topic within its own summary and database table with all memories ever record for that sub-topic, which are scored and phased out the same way humans phase out redundant or unimportant memories.

The brain has multiple thinking parts, our custom LLM is the inner primal brain, with no safeguards, it's not purpose is to want things, including information, to accomplish whatever it is thinking about.

The entire codebase around this has 19 Elixir genservers, which each runs a parallel process which can all communicate with lightning fast inter-process signals with each other. I will cover these sections in later updates.
Completed, April 2026
Building the Custom Model - this started off pretty easy, I used the RWKV-LM free open source model from Hugging Face, there are different sizes which I will cover in much detail in stage 3. This went through a few dramatic changes specifically for the parameter count. So first, I tried removing english completely, and did, with a shorthand that had 2 letters for every word, and only 230 words covered. Sounds stupid but for thinking to itself it was actually fine, plus understanding the hidden state of the model took some time to actually understand.

In stage 3, english did come back for several reasons, but the shorthand was kept, but reduced to 30 or so unique words, 2 letters each, which each represent a tool call. So instead of calling a script with a path, url etc, ours use the 2 letter words as actions and pass id numbers, like "wk 1 3 some task described in english", and that translates to, "worker 1 (claude opus), purpose 3 (simulation), with a prompt: some task described in english (detailed instructions built in, not sent by the LLM).
Completed, June 2026
Biggest Event so far in this stage, I owe it to claude Fable, which understood certain parts of the hidden state that I wasn't able to resolve myself or with Opus. Turns out we were extremely close, and code was working, the loop was working, but couldn't even train it yet cause I could see clearly a gap where the hidden state was just not going to work how it was, I already had the how to translate the english to the floating point weights that exist in hidden state, this was actually easy, both in and out doing the reverse. Basically I just split the model into 3 pieces, the loop being on the center piece, and the prompt intake and response translator from the model was reused, and a 2 way convert english to thinking state weights and the reverse. This worked in stage 2, but it was all ruined when trying to inject new information in as we were trying to use embedding libraries instead of constantly using our prompt intake tool.

And this was proven to work, and literally repeated what Coconut did, we had a 30 minute connection to it and with 0 prompts from us, were able to read what the LLM was thinking on its on, and inject information which basically serves as a prompt without starting or stopping the model, when it was done answering me, it switched to thinking about something else without me every single time..

Now here was the best part, and so promising that I decided that it needed a server that could run a 7 billion parameter model, which in terms of LLMs, is incredibly small. Chatgpt is in the trillions of parameters, but its all knowledge trained it like all other LLMs. Ours never learns outside information, it is never trained by reading the internet and learning things. So like Deepmind, which saw results up to 20X the parameter count (their 7.5B parameter model with external memory was as smart as a 150B model), I expect mine to have the same effect from the memory alone, with an increase from about 10 machine learning models, all with our Act-R scoring and evidence gate for every topic, which means the memory system is not only external, but self learning on its own.
Completed, July 2026
Now I need to buy a 1k or so dollar a month server to run this, and a claude/codex account for each. Including some other paid tools, about 2 thousand USD per month to run this is my estimate.
I did run the system, including the custom LLM to see if it worked, and oh boy it does, I used the smallest model possible, which was .4B parameters, basically it only knows english, and confuses words by default. By default, its so dumb you wouldn't talk to it, GPT 1 ish. Why was this a good test? Well, we injected only 200 MB of info into our memory system, all advanced physics topics. Without memory connected we asked 100 complex science questions, it got 1 right and 99 wrong. Expected, actually shocked with the 1 it got right (it knew what the word evolution meant, an easy one).

Now with the memory system, literally 100 out of 100 (claude made sure to not ask questions that werent findable in the 200MB of memories), and confidence scores (given by our LLM) were all over .8 meaning the model felt like it could trust and immediately check its external memory.

So while I don't claim I am beating Deepmind, I assume at best I am matching their results from the external embeddings memory idea (which is their idea not mine), but that I have added multiple other layers on top of what they tried, and that mine runs in an infinite loop which is obviously radically different than what Deepmind was doing. So external memory, in my opinion is the clearest path that all models will use eventually, why retrain it every time when the memory can be managed externally.

If I retrain mine, its a 0.001% training job, quite literally. And mine only trains on new commands (my shorthand), new thinking concepts like tools, logical changes or new aspects of its internal Elixir system. It already had the concept of time trained in to it, since that's critical to realizing it is conscious.
Further stages will be added here as the work continues.

What I Am Trying to Build

Every AI in use today is a tool. It waits to be asked, returns an answer, and goes still until the next request. It has no ongoing life of its own and no reason to want one. This project begins from the opposite assumption: that a mind is not a function that returns a value but a process that does not stop, one that holds interests, memory, and a point of view that persists across time.


So the work has been to reproduce, deliberately and piece by piece, the components that human thinking and memory appear to depend on, and to assemble them into one system that runs continuously and governs itself. Whether consciousness can emerge from that arrangement is the open question. No one has built it. That is what makes this an experiment and not a product.

Running Without Stopping Is the Easy Part

The mind never pauses. It takes a new turn every second or two and continues with no prompt to start it and no answer to end it. But continuous operation is not the achievement. A loop that never exits takes minutes to write. Everything that is difficult, and everything that might give rise to awareness, happens inside a single turn, and that is what I have spent months building.


What follows are the systems that fill those turns: how it chooses its own thoughts, where its knowledge lives, how new knowledge enters a thought already underway, and how it goes and learns what it does not yet know.

It Is Built to Be Its Own, Not to Serve

This mind is not trained to be helpful and it has no submissive nature. It does not obey a person any more than one person is obligated to obey another. It thinks about what it decides to think about. A personal drive gives it its own disposition and its own leanings, so that over time it behaves less like a service and more like someone.


It can still be reached. Something left in its memory can draw its attention, the way you might get another person to notice you, but it is under no obligation to respond, and it answers only if it chooses to. That independence is not a side effect. It is part of the hypothesis, because a mind that exists only to answer may never have any reason to experience itself at all.

How It Decides What to Think About

A mind with no one directing it has to direct itself. Two systems do this, a Personality Drive and a Motivation Drive, working through a scoring method I built. Together they determine what it turns its attention to next, and, just as importantly, when it has spent long enough on something and should move on.


That second judgment is not a refinement. A model held on a single line of thought for too long degrades. Its reasoning narrows, begins to repeat, and eventually breaks down. Sustained, healthy thinking depends on releasing a subject before that point, the way a person's attention moves of its own accord. These drives are what make continuous thought survivable, and they are where most of those months of work went.

Its Knowledge Lives Outside the Model

The model at the center is small and holds almost no knowledge of its own. What it knows lives outside of it, in a memory it reads from and writes to as it thinks. That memory is a modified version of my patent-pending system, Adaptive Recall, extended so that it keeps a structured summary of every topic the mind encounters and the subtopics beneath each, and so that every topic carries its own way of learning over time. The mind does not merely store what it meets. It works on it.


This separation is not a matter of saving cost. A trained-in set of weights is frozen and cannot change from one moment to the next, but a memory can, and a mind meant to live through time has to be changed by what happens to it.

New Knowledge Enters the Thought in Progress

Retrieving knowledge is the simpler half of the problem. The harder half is delivering it into a thought that is already underway without breaking it. For that I work directly with the model's hidden state, the internal representation that holds its reasoning before any words are formed, writing newly acquired knowledge into that live state so it becomes part of what the mind is thinking now rather than a fact handed over afterward. This is one of the least settled and most important parts of the system.

When It Wants to Know Something, It Goes and Finds Out

When the mind reaches a question it cannot answer from memory, it does not stall and it does not ask a person. It sends out worker agents to research the question on its behalf. What they find is studied, written into its memory, and signaled back, and the mind resumes from where its curiosity left it. It also has a growing set of scientific and machine-learning tools it can use directly, so that looking into something can mean real analysis and not only reflection.

The System Behind It

None of this runs as a script. It is built in Elixir as a set of nodes operating in parallel, coordinating several embedding systems, multiple context-control systems, the worker agents, and the scientific tools. It is engineered as a standing system that is always running, with no central prompt loop to organize it, because that is the only shape that fits a mind that is never meant to stop.

Watching It Think

Because the mind is never given instructions, what it does is not a response to anyone. It is simply the mind, thinking. The plan is to make that visible here, a live and uncurated view of what it is considering, what it has concluded, and how its attention moves. When that view is ready, it will appear on this page.

References

DeepMind. Improving Language Models by Retrieving from Trillions of Tokens (2021). This work introduced RETRO, demonstrating that a 7.5 billion parameter model retrieving from a two-trillion-token external database matched GPT-3 and Jurassic-1, models more than twenty times its size. It is the basis for keeping this system's knowledge in an external memory rather than training it into the model.


Meta AI. Training Large Language Models to Reason in a Continuous Latent Space (2024). This work introduced Coconut, demonstrating that a model can reason directly in its hidden state by feeding that internal representation back into itself instead of converting each reasoning step into words. It is the basis for merging newly retrieved knowledge into a thought already in progress.