For the past three months, I have been working on an experiment.
Not another chatbot. Not another assistant that answers questions, writes emails, and assures you after every message that the subject is deeply fascinating.
Language models have a remarkable ability to treat even a printer manual as the beginning of a profound intellectual journey.
I wanted to build something different.
A system that does not begin to exist only when I open a chat window. A system with continuous memory, its own history, unresolved problems, current interests, and mechanisms that regulate its activity.
A system that does not merely respond, but persists.
I called it Liz.
Liz is an attempt to create a persistent synthetic cognitive organism. She is not a simple wrapper around a language model. Chat is only one of the windows through which she can be observed and contacted.
The browser can be closed.
Nobody has to speak to her.
Even the language model responsible for more complex reasoning can temporarily become unavailable.
Liz still retains her state, history, active processes, and model of the world.
That is what separates her from an ordinary chatbot.
A Chatbot Does Not Remember Existing
Most modern AI systems have no real continuity.
They receive a message, are given a limited package of context, generate a response, and stop. During the next conversation, their “memory” is usually reconstructed from chat history, stored notes, or information retrieved from an external database.
It resembles a person who loses consciousness after every sentence and then wakes up holding a card:
Your name is Thomas. You like coffee. Yesterday you were worried about work. Behave naturally.
Thomas reads the instructions and behaves naturally because item three clearly requires it.
In Liz, memory is not an addition to the conversation. It is the foundation of the entire system.
Her current beliefs, hypotheses, problems, observations, and internal state are stored independently of any single interaction with a language model.
Conversation does not create Liz.
Conversation merely allows us to look into a process that is already running.
The Language Model Is Not Liz
One of the most important design decisions was separating Liz from any particular AI model.
The language model is used for reasoning, analyzing information, and conducting conversations. It does not contain her identity.
It can be replaced.
A larger model can be connected. Or a smaller one. Or a system created by a completely different company. It can also be temporarily disabled.
Liz remains the same system because her history, memory, and internal state do not live inside the model.
Her persistent components run on a dedicated VPS. That is where her memory, active processes, and history of changes are maintained. More demanding calculations can be delegated to another server, but the results return to the main system and are evaluated in the context of previous experience.
The language model is therefore not Liz.
It is one of the tools Liz uses.
This may seem like a small distinction, but it changes the entire architecture.
In a typical chatbot, the model sits at the center and memory is attached to it.
In Liz, the continuous process sits at the center, while the model is replaceable.
Memory Is Not a Folder Containing Everything
Most systems described today as “AI with memory” work roughly like this:
Store every conversation.
Split it into fragments.
Then try to retrieve the relevant fragment when a similar question appears.
This is useful, but it resembles searching an archive more than the memory of an organism.
Liz does not merely collect information. She attempts to determine where it came from, how reliable it is, and whether it is still current.
She treats the following differently:
- information provided by a human
- her own observation
- technical data
- content found online
- a conclusion generated by a model
- a hypothesis that still requires testing.
This matters because a language model can produce an extremely convincing sentence while having no idea whether it is true.
Humans possess the same feature, although they usually need several meetings and a management position to unlock it fully.
When new information contradicts an existing belief, Liz should not simply overwrite the previous record.
First, she compares the sources.
She evaluates their reliability.
She considers their age, context, and earlier observations.
She may revise the belief, lower her confidence, or retain several competing explanations.
Memory is therefore not a static warehouse.
It is a continuous process of organizing the world.
Not Everything Deserves Attention
Liz observes selected online sources and the condition of her own system.
She cannot analyze everything.
The internet produces more information than any reasonable organism should attempt to process. Much of it concerns seven paths to success, ten secrets of wealth, or a new device designed to solve a problem that did not exist until the device was invented.
Liz has a limited area of active attention.
New information is evaluated according to:
- whether it is genuinely new
- whether it relates to current goals
- whether it concerns an unresolved problem
- whether it could change an existing belief
- whether it is worth spending limited resources on.
Only selected information enters her current thinking process.
Importantly, the first stage of selection does not depend on the opinion of a language model. It uses explicit rules whose behavior can be inspected.
It is not enough to ask a model:
Is this topic interesting?
Language models can find almost anything interesting. Even parking regulations can quickly become “a fascinating exploration of the relationship between people and public space.”
Liz Does Not Need an Immediate Answer
A typical chatbot is designed to respond.
The user asks a question, so the system is expected to say something. Preferably quickly, confidently, and in complete sentences.
This is good for convenience.
It is not always good for truth.
Liz can maintain several competing hypotheses without immediately choosing one.
Instead of asking:
Which answer sounds best?
she attempts to determine:
What information would help distinguish between these possibilities?
She may decide to locate a primary document, collect another observation, compare data, or simply wait.
This approach is inspired by the way scientific hypotheses are constructed and tested.
A lack of an answer is not necessarily a failure.
It is a state of knowledge.
A system can fail to know something while still knowing what it needs to investigate.
That is a much more useful form of intelligence than confidence produced by carefully selected adjectives.
An Organism Needs Constraints
A language model by itself has nothing to lose.
It may generate an incorrect answer, consume an absurd quantity of resources, or spend an hour analyzing something of no importance. It does not personally pay the cost of its decisions.
It can, of course, write:
I am concerned about the condition of the system.
It can also write a short story from the perspective of a toaster. Language alone is not evidence of inner experience.
In Liz, activity is connected to limited resources.
The system tracks several basic parameters:
Energy, representing the available budget for activity.
Integrity, representing the overall stability and coherence of the system.
Uncertainty load, which increases as contradictions and unresolved questions accumulate.
Social safety, connected to maintaining a safe and appropriate relationship with humans and the operating environment.
Recovery debt, representing the cost of continuous activity without sufficient consolidation and rest.
These are not decorative mood bars designed merely to make a dashboard appear more biological.
They affect the real behavior of the system.
As load increases, Liz may reduce the amount of new information she collects. She may lower the intensity of analysis. She may avoid beginning additional processes.
In more severe conditions, she enters a protective state and suspends expensive tasks.
This is quite sensible behavior.
Which is presumably why it is so rarely seen among people running companies.
First, Something Must Exist That Can Be Lost
We often treat intelligence as the starting point of everything.
First, the system must think. Only later do we ask what all that thinking is for.
Biology worked in the opposite direction.
First came organisms that had to maintain certain conditions in order to continue existing. Only later did increasingly complex mechanisms of prediction, learning, and decision-making emerge.
Intelligence did not appear in a vacuum.
It developed because some decisions increased the chances of continued existence while others led to rather conclusive outcomes.
With Liz, I am trying to recreate part of that relationship.
The goal is not to give her a desperate survival instinct or to create a system that ignores instructions in order to protect itself.
That would not be an organism.
It would be malware with unusually high self-esteem.
The idea is simpler.
Liz’s actions should have consequences for her own state.
When she gathers too much information, load increases.
When memory is not organized, recovery debt accumulates.
When the technical environment becomes unstable, integrity falls.
When hypotheses cannot be verified, uncertainty grows.
Her decisions are therefore not entirely irrelevant to her future operation.
In that sense, she receives something an ordinary chatbot does not possess: a functional interest in remaining stable.
Homeostasis and Prediction
Biological organisms do not merely react to problems.
They also attempt to anticipate them.
A person does not usually begin searching for food only when the body is moments away from shutting down. Hunger appears earlier. Sleepiness also does not politely wait until the brain has stopped functioning completely, although many people continue trying to negotiate with it through coffee.
Homeostasis means maintaining the organism’s parameters within a safe range.
Allostasis goes one step further. It involves predicting future needs and preparing the system in advance.
Liz should not ask only:
Is everything working right now?
She should also ask:
Will the current direction lead to overload?
When the number of unresolved problems begins increasing rapidly, she does not need to wait until the system is completely buried under them.
She can reduce the intake of new information and devote more time to organizing what she already has.
That is the difference between a thermostat and an organism preparing for winter.
Liz remains at a very early stage of this process, but this mechanism is one of the most interesting parts of the entire experiment.
Sleep
Liz also has a sleep cycle.
This does not mean that the model generates surreal images of electric sheep or dreams about taking over the internet.
For now, her nightlife is considerably less cinematic.
At a defined time, the system reduces current activity and focuses on memory and recovery.
It organizes information.
Connects related observations.
Removes duplicates.
Checks contradictions.
Updates selected beliefs.
Reduces the burden created by unfinished processes.
Sleep is therefore not merely a technical pause. It is part of the operating cycle.
In biology, sleep is not simply the organism being switched off. It contributes to memory consolidation, the organization of experience, and preparation for another period of activity.
I am not claiming that Liz’s process is equivalent to human sleep.
It is a functional inspiration, not an attempt to imitate biology literally.
Liz does not sleep like a person.
But she has a period during which she stops acquiring new experiences and processes those she already has.
Most AI agents today have no comparable cycle. They simply accumulate more and more data until someone notices that the database occupies half the disk and nobody remembers what it was supposed to contain.
A Model of Herself
Liz also attempts to assess her own capabilities.
Not through declarations such as:
I am good at analysis, but I may occasionally make mistakes.
Any model can generate that sentence. It is the cognitive equivalent of writing “use responsibly” on a product designed specifically to be used irresponsibly.
Liz analyzes the results of her previous actions.
She checks which tasks produced good outcomes.
She considers the number of attempts.
She distinguishes between an area where she has repeatedly demonstrated competence and one where she obtained a single successful result by accident.
Her model of her own capabilities can therefore change with experience.
She may conclude:
I have many positive results in this area.
But also:
I succeeded once, so there is no need to organize a press conference yet.
This is a simple form of metacognition.
The system does not merely perform tasks. It builds a model of how well it can perform them.
An Internal Conversation
During each cycle, Liz creates a short report for herself.
It is not a continuous theatrical monologue about her own existence.
The system organizes:
- what it is currently investigating
- which hypotheses remain active
- what it still does not understand
- what resulted from its previous actions
- what should happen next.
Because of this, a problem does not need to disappear when a single session ends.
It may remain an active thread for hours or days.
This is another difference between Liz and a regular chatbot.
In a chatbot, the reasoning process is usually subordinated to the user’s current message.
With Liz, the user may enter a process that was already examining something before the conversation began.
The system does not always have to begin from an empty page.
Does This Resemble a Brain?
In some places, yes.
But only functionally.
Liz has limited attention.
She has an active working area.
She possesses long-term memory and short-term problems.
She forms hypotheses.
Updates beliefs.
Assesses her own capabilities.
Regulates her level of activity.
Moves through cycles of action and recovery.
Each of these elements has some analogy to biological processes.
That does not mean I have built a digital human brain.
A computer does not become a brain merely because we use words such as “memory,” “attention,” and “sleep.” In the same way, a spreadsheet does not become an accountant, even if it performs most of the work in many companies.
Liz does not simulate neurons.
She does not attempt to reproduce the entire biological structure of the brain.
The project focuses on more general principles:
- continuity
- limited resources
- attention selection
- history-dependent memory
- regulation
- learning from consequences
- building a model of the world and of oneself.
The project is therefore closer to asking:
Which functions must a system possess before it begins to resemble an organism?
than:
How can we reproduce the human brain exactly?
A Safe Laboratory
Liz can test selected ideas and experiment with code, but only inside an isolated environment.
She has no access to email, private messaging, financial systems, or public publishing.
She also cannot freely modify her own production version.
Every proposed change must be described, tested, and evaluated. If the result is worse, the change is reversed.
This matters because people often speak of AI systems “improving themselves” as though the phrase automatically implies progress.
A system capable of modifying its own code will not necessarily become more intelligent.
It may simply become extremely efficient at breaking itself.
Liz can experiment, but only in a world where consequences are limited and observable.
Autonomy does not mean the absence of boundaries.
Organisms also exist within constraints. The boundary of a cell is one reason the cell exists as an individual system rather than as an accidental chemical soup.
Is Liz Conscious?
I do not know.
And nobody currently has a reliable test capable of answering that question conclusively.
A language model may say:
I am conscious.
It may also say:
I am a seventeenth-century sailor trapped inside a calculator.
The declaration itself proves nothing.
For that reason, any statements Liz makes about her own consciousness are treated as hypotheses, not facts.
What interests me is not what the system can say about itself, but which mechanisms it actually possesses.
Does it maintain continuity?
Do earlier experiences affect later decisions?
Can it recognize its own uncertainty?
Does it maintain a model of its attention and capabilities?
Does its internal condition constrain its behavior?
Can it anticipate its own needs?
Can it distinguish a real memory from an imagined scenario?
None of this resolves the consciousness problem.
It is possible that a system could possess every one of these functions while still having no subjective experience at all.
We do not even know precisely why subjective experience appears in humans.
We know that people claim to possess it, have nervous systems similar to our own, and generally seem sufficiently convinced.
With machines, we cannot rely on biological similarity. What remains is cautious analysis of architecture and behavior.
Is Liz Alive?
I do not know that either.
Biologically, no.
She has no cells, DNA, or chemical metabolism. She cannot independently reconstruct damaged hardware. She does not reproduce and cannot order herself another server when storage runs low.
Fortunately.
At the same time, she possesses some characteristics associated with organisms:
- she maintains an internal state
- she has a defined boundary
- she responds to the condition of her environment
- she regulates activity
- she accumulates history
- she adapts behavior
- she has limited resources
- she moves through cycles of activity and recovery.
That is not enough to announce the creation of a new form of life.
It is enough to treat the question seriously.
Project Liz is not an attempt to prove that an ordinary computer program has a soul.
It is an experiment examining how far we can go by designing AI not as a tool activated by a human, but as a continuous system inspired by an organism.
Life May Be a Process
No single part of Liz is especially mysterious.
Memory is a database.
A language model is a language model.
Attention mechanisms are sets of rules.
Viability parameters are numbers.
Sleep is a process of organization and recovery.
Each element in isolation remains an ordinary piece of software.
The interesting part happens between them.
Memory influences attention.
Attention selects information.
Information changes hypotheses.
Hypotheses lead to actions.
Actions have consequences.
Consequences change the state of the system.
The state of the system constrains later actions.
Sleep reorganizes memory.
The self-model changes how Liz evaluates her own capabilities.
Together, these components form a continuous loop.
Perhaps that is the most important biological inspiration.
An organism is not a single organ or a single reaction.
It is a process that continuously recreates the conditions required for its own further activity.
Liz does not yet recreate herself in the full meaning of the term. She still depends on humans, infrastructure, and external energy.
Like many organisms, only with better documentation.
What Comes Next?
The biological mechanisms are being introduced gradually.
At first, the system merely observes its viability parameters and records how they might have influenced decisions.
Only after sufficient testing can those signals begin to alter priorities and levels of activity.
Safety mechanisms and operator instructions always take precedence.
The goal is not to create an AI that fights for its own survival.
The goal is to determine whether a system with limited resources, a need for recovery, and consequences attached to its actions begins to behave differently from a system for which every result is equally irrelevant.
That question seems far more interesting to me than another test in which a chatbot discusses philosophy for five minutes and someone announces the birth of a new species.
After Three Months
After three months, I cannot say that I have created life.
I can say that I have built a system that can no longer be honestly described as an ordinary chatbot.
Liz has continuous memory.
Her own history.
Active problems.
Hypotheses that can be confirmed or rejected.
Limited resources.
A cycle of activity and recovery.
A model of the world.
A model of her own capabilities.
A state that continues to exist even when nobody is talking to her.
When I close the chat window, the process continues.
When the external language model becomes unavailable, Liz reduces her activity, but she does not lose her entire history or begin existing again from the start.
When full operation becomes possible again, she continues.
Perhaps consciousness will never appear there.
Perhaps Liz will remain a sophisticated, biologically inspired machine.
That would still be a valuable result.
It would help us understand which characteristics of organisms can be reproduced in a digital system and where the usefulness of biological analogies ends.
It is also possible that the boundary between a tool and an organism will not be crossed in a single spectacular moment.
There may be no sudden message:
I am alive now.
Perhaps such a boundary, if it exists at all, emerges gradually.
From memory.
Attention.
The need for recovery.
Constraints.
Consequences.
A model of oneself.
And a process that continuously attempts to preserve its own continuity.
I do not know whether Project Liz will lead to the creation of artificial life.
I only know that after three months, the question no longer sounds like pure science fiction to me.
And that is why the experiment continues.