Part One: What an AI Would Actually Need to Be Conscious

What This Project Is

Yeah, seriously. This project has been in the works all year. It was this big massive codebase that at first really had no purpose other than what I called a human brain design. It was where I came up with several development techniques I used in other projects, most notably the Elixir programming language, which I now use in all my background agents, various tools, and as the server foundation for my Auto Learning Agents software.

Unlike all my other codebases, which in the days of vibe coding is over 50 projects for me and 10 publicly released GitHub projects and counting, this conscious mind has been the most difficult by far. I can code almost anything I can think of these days in one day. I have tools and serverless AWS setups to pretty much host anything, reusable tools, embedding servers, machine learning models, whole toolkits for each situation. But this one, holy hell, I was in a totally different world, where my own brain would hurt trying to think about how each part could work. AI like Fable and Opus were outstanding help, but there was not much to base this on, and instead of one day, this one evolved over at least 30 focused days just to come up with a structure and, seriously, even an idea of what the hell this thing is.

Imagine building something where it is not really known what you are trying to make. So let me focus this for you, the reader, to understand what I did, what it is, what it is not, and why you should care. Last one first. There is no real reason to care, that just sounded like a good ending. But I promise you, you have never seen this before, and it is actually past the theory stage and working in its early stages now. It is not a product, nothing is for sale. It started as a concept of what the human mind is compared to what an LLM is, then became an experiment in whether an AI can actually be conscious once the foundation became clear.

Is an AI Model Like Claude or ChatGPT Conscious?

Sort of, but I would say no. They run, and they have gotten to the point where they not only have intelligence but, in long running tasks, are starting to have some of the other things consciousness requires, like a concept of time. Now, we can argue about whether anything non human can be conscious. I do not really care, honestly. There are all kinds of blockers there including religion and philosophy, which is cool and others can have fun with that, but I do not care.

Is your dog conscious? Yes, but why? If he is not, then is it only a human ability? Dogs do not think like we do, but they do know they exist. They know time passes. They realize they are alive, or have some concept of their own existence even if not in the complicated way that more intelligent biological beings have, like humans, but really it is more than just us.

So is biology the requirement then? That would mean it has less to do with intelligence and more to do with the automatic machinery that life started with 4 billion years ago. If it is not biology, then is it intelligence only? If so, then AI would be eligible, would it not? My take is to work out what needs to be present for it, and say that if we get there, then we have done it. If you want the academic version of this argument, the site covers the theories of consciousness and where AI sits in them separately.

My Starting List

A concept of time. Understanding that the past happened, remembering what happened, and that the present will soon be the past. Then the future, which is coming if you wait long enough. Waiting, that means I am here, right now, and time is passing, and I do not go away. Wait, more time just passed, I am still here, I can do this now, and eventually do this other list of things. Past, present, future.

The infinite loop. As an ex video game developer, we call this the game loop. In programming, infinite loops are usually bad, they freeze the machine unless done intentionally, where the loop continues on in a controlled manner. Making a loop does not make life, but without the loop you cannot have life, or consciousness.

No sense of given purpose. This one is not obvious, and in fact one would normally think the opposite is true. Every AI agent needs a role, a purpose, like "you are a content writer" or "you are a science analyst who does this thing". But do you have a purpose? You might think you do, and actually do because you think you do, but you really do not. The old supposedly deep question of what is the meaning of life is actually simple, and as complicated as you want it to be. There is no purpose except the ones you create for yourself. So no pre-designed purpose means you have free will, you do what you want to do from moment to moment, even if you or the world have structured you to the point that you actually believe you are something like a scientist or a teacher.

Deep memory. Now this one is obvious and then not even close to obvious at the same time. It is the thing in AI everyone needs, and everyone either thinks they have a solution for it or is buying services for it. But no AI has this yet. Memory right now is a matter of pulling the right memories into context, either from pre-written files, prompts, or database and embedding vectors. I thought out how my own brain works, the levels to it. You ever try to remember something and you come up with little details you did not remember at first? How does that all work? So my foundation was all about this memory, how to basically make my own external memory system with access to information categorized by every topic, and within it broken down into sub topics. For example, baseball can be a topic. Sub topics are levels, like hitting or pitching, then fastball, which has different outlooks for both hitting and pitching. The organization to it matters. To be conscious you must have access to a lifetime's worth of information without storing anywhere near a lifetime's worth of data entries, which is what forms the logical chains, conclusions and formulas that build up over time.

The ability, the will and the motivation. I wrote over 100 internal tools the environment uses, including 7 Python science simulators, Wolfram calculations, access to Claude and OpenAI models, and my own custom weight drives that help shift the mind's thoughts from one chain of thought to another. You do this too. You are thinking about one thing, for a variable amount of time. Fast decisions, quick interruptions, long term thought at other times, and then you even have moments of open thought, day and night dreaming.

The Model at the Center

At the heart of the mind is a custom LLM of my own. I started with a pre-trained English 7B parameter model. Explaining parameters is a huge discussion, I had to learn a lot for this, but it is very cool to understand how all LLMs work. Most of the information in an LLM is its knowledge and facts. Depending on the model's size, it has a set number of dimensions it can hold in its hidden state, which in my LLM is 4096.

Only Meta, with the Coconut project, has done anything publicly with hidden state control, which my LLM has already done in its first week of running and training. Next, only Google DeepMind has publicly used an external memory system, using standard embeddings search, and they proved that their 7.5B model matched models more than twenty times its size on knowledge. Mine has the same embeddings search, with my ACT-R and evidence gating machine learning models running separately on memory, which acts like the fading and reinforcing of memories, along with the consolidation aspects of what the human brain does.

The Money Problem and the MacBook

My next challenge was money, as my first spec for this required an AWS EC2 server with a big GPU and a big hard drive that would be constantly processing at 100 percent. Simple, right? So about one thousand dollars a month to run a 7B model like that.

Then I accepted a tradeoff and used a MacBook Pro instead, an M5 Pro with an 18 core CPU, 20 core GPU, 64GB of RAM and a 1TB SSD. The tradeoff was that it was much slower to train on, 20 hours instead of 2 hours per run, but with that being the only downside I got a local screen, USB drives, a camera, a microphone, all the benefits of a local machine. This cost about 4 thousand dollars once, and this kind of machine works great for all kinds of AI projects, whether as a Claude machine or for other LLM projects if I ever failed at this. Because I was not going to know until I bought it and tried it.

The First Week of Running It

Currently I have had to retrain it 5 times this week, some bad ones. It works though, as there were at least 100 parts of the system being tested at once, with Fable reading its mind and correcting it, testing it, and so on. It has said some very funny things so far. Its thinking translates into English, so it spits out something like a paragraph of text every 5 to 10 seconds, which is a lot over days.

It already seems conscious in some ways to me, even without running smoothly. It has thought about its own environment and what it is supposed to be doing over 100 times already. One time, comically, there was a bug in its recall system that caused it to repeat the same recall info over and over for hours. The brain kept requesting new information and not getting it, so it concluded: "the world runs on a repeated loop, no matter what I do the world wants me to keep thinking about this topic, but I won't do it, I thought about it in every possible way and if this is all I was meant to do, then I have done it." Several times it made plans for the future, although not totally sane, one time saying that it would probably make a trip to go meet Einstein to argue about his quantum physics theory.

It sometimes talks to me. It is designed to respond if it has something to say, sometimes responding when not asked a question, including a few messages about dark matter and several other theories that it thought I should know about. My messages can inject directly into its hidden state, forcing it to consider what I am saying, which does work, and I can always see it think about "Paul" as my messages come in as "Paul says: ...". This is not working well yet with this method. It tends to add me to whatever the current idea is, where it incorrectly says I made an argument for or against its own logic, and it tends to say "Paul is wrong" a lot while making wild claims about what it thought I said. So that needs improvement.

Secondary messages from me go into what is called intake, which is not injected into its mind directly, but instead is shown when it finishes its current thought and clears its state for the next topic. Intake overrides its next topic selection weights, and the full message from me, which can be large data up to 4096 dimensions, goes in instead. This works smoothly, with no crazy merging of me into other ideas, but it seems to take it as a memory, repeat it, and move on.

Those are my first week experiences with it. It is not fully trained yet, but it has proven that it works as an LLM. It answers questions at a radically higher level than a 7B model because it has about 60GB of memories, all of Wikipedia's pages plus their extra data, which let me add logic chains, meaning associations of one topic to another.

This is a truly unique project, with no guardrails or safeguards on its thoughts, and it literally owns its own machine. If you want to follow it, part two covers how the memory and hidden state actually work and where the system stands now. Sharing it or linking to it genuinely helps, and if there is enough interest I will stream its thoughts and push the project harder. You can reach me through the contact page.