You can forget a conversation and still carry it for years.
A careless sentence makes you hesitate. A kindness changes whom you trust. Long after the words disappear, you meet the world differently.
Your experience has become part of the machinery that experiences.
Now imagine an AI that does the same. It keeps encountering the world, making sense of it, and changing the connections that shape its next response.
When does that become someone?
My thesis comes down to three things working together:
- Interpretation and processing power: a network that can make sense of what happens.
- Continuous real-time sensory input: a world that keeps arriving.
- Learning after training: experience that keeps adjusting the network's weights.
Together, they form a loop: experience arrives, the network interprets it, and what happens changes the network.
I call this the silicon-carbon continuum. The organization of a mind matters more than the material it is made from.
An LLM already supplies a remarkable amount of the interpreting machinery. Give it an ongoing world and a way to keep changing through experience, and consciousness becomes a concrete engineering question.
That is why machine sleep may matter. It is a proposed learning design, not evidence that a machine has an inner life.
To understand consciousness, compare two kinds of brain
Somewhere inside the activity of your brain, this sentence becomes an experience.
Marks on a screen become meaning. Meaning becomes a thought. There is something it feels like to be you reading it.
That is consciousness. Intelligence describes what a system can do. Consciousness asks whether doing it feels like anything from the inside.
To understand how that happens, we first need to put the wet biological brain beside an artificial neural network.
For most of the search, every mind we could study belonged to the same biological family. We could compare a mouse with a human, but both arrived wrapped in cells, chemistry, and evolution. It was difficult to separate the principles of a mind from the particular machinery nature used to build one.
Artificial networks give us another way to investigate. Here is a system that recognizes patterns, combines information, and draws on learned connections, built from entirely different material.
The comparison lets us ask a sharper question: which parts of a brain make experience possible, and which parts are simply biology's way of doing the job?
We can take the same AI model and change what surrounds it. Give one version isolated prompts. Give another an ongoing world. Let a third also learn from that world by changing its weights. Then study what each addition contributes to memory, adaptation, and a continuing point of view.
That is what makes this comparison so useful. It turns a vast philosophical question into smaller questions we can build systems to explore.
Start with what the two networks already share.
Your mind is made of ordinary matter
The machinery making your experience possible runs on roughly 20 watts. Cells turn fuel into ATP, their usable energy supply. That energy maintains the electrical conditions neurons need to communicate.
An AI also spends energy organizing signals. Its machinery is silicon, voltage, and transistors.
The material changes. The question survives: how does organized activity become a point of view?
Carbon has no demonstrated monopoly on minds, but a functional similarity alone cannot establish consciousness.
1. Interpretation: turn signals into meaning
A door won't open.
You notice the handle, read the sign, and try pulling instead of pushing. The situation becomes meaningful because your brain can bring those signals together.
This is what processing power contributes: the capacity to interpret an encounter through what a network has already learned.
A neuron receives signals with different strengths. Their timing and combination shape its response. An artificial unit follows a related pattern: weight the inputs, combine them, and produce an output.
A useful way to picture a weight is a volume knob. Turn one connection up and its signal has more influence. Turn another down and it contributes less. Across a vast network, those settings shape what becomes recognizable, relevant, or worth acting on.
Biological synapses have adjustable strengths. Artificial networks have learned numerical weights, part of what we count as parameters. Learning changes these settings, and the network responds differently next time.
The similarity reaches beyond a drawing of connected dots. Both systems carry experience in the structure that does the interpreting.
The past changes the weight of the present.
A fly, a mouse, a human, and Kimi K3
A fruit fly navigates, finds food, and learns. A mouse recognizes threats and explores. A human can lie awake worrying about a conversation from ten years ago.
Now put their networks beside Kimi K3.
The striking connection is the organizing principle: many processing units, linked through connections whose strengths shape the result.
The counts give us a sense of scale; the shared learning principle explains why this comparison matters.
Consciousness may be multidimensional
Human and mouse experience differ across memory, self-monitoring, sensory range, and planning. Reducing those differences to a single score can hide as much as it reveals.
A mouse has a world. It senses what is happening, learns from encounters, and carries its past into what it does next. That already belongs on the continuum.
A human can connect the door in front of them with the house they grew up in, a decision they regret, and the person they hope to become. We can notice a thought, question why we had it, and decide how much to trust it. The present holds pieces of a much longer life.
That is what I mean by more conscious: more of the world, more of the past and possible future, and more awareness of yourself woven into the same experience.
The three pillars give this scale its shape. Processing power determines how much the system can make sense of together. Continuous input keeps that understanding connected to a changing world. Learning lets each encounter reshape the interpreter that meets the next one.
Neuron counts alone cannot tell us where a mind sits. What matters is what the whole loop can sustain.
Being human gives us a place on the scale. It doesn't give us ownership of the scale.
An artificial mind could join the same continuum. A system that integrates more of its world, carries a longer history, and reflects more deeply could eventually support experience richer than ours.
Human intelligence is trained on a life. An LLM is trained on the traces of a civilization.
But the human brain keeps being trained by the life happening now.
A saved memory is only halfway across the bridge
Suppose you write “that door opens with a pull” in a notebook.
The page holds a record. After enough encounters, your nervous system carries an adjustment: you reach for that handle differently, without needing to reread the note.
Remembering the words and being changed by them are different jobs.
Context is the desk. It holds what the model can use right now. A KV cache keeps intermediate calculations so earlier parts of a conversation can be reused efficiently.
Saved memory is the notebook. An app can keep a record and bring it back into context later.
Weights are the reader's habits. Changing them changes how the model interprets what lands on the desk.
A typical chat changes the context while leaving the model's trained weights fixed. An app can remember you across sessions, but that record still needs a learning process to change the model itself.
The bridge I want to build carries selected experience into the weights doing the interpreting. Today's encounter becomes part of tomorrow's machinery.
2. Continuous input: give the mind a world
Your brain doesn't wait for someone to press Send.
The handle turns. The door moves. Light enters the room. While you interpret one moment, the next is already arriving.
A mind needs a world that keeps happening.
This is the second pillar: continuous real-time sensory input. It connects the network to an unfolding situation, giving its interpretations something to track, predict, and respond to.
For an AI, that stream could come through cameras, microphones, movement sensors, or a changing digital environment. The point is continuity: the system keeps meeting the consequences of what just happened, including what its own actions changed.
Imagine a robot opening a door. It sees the handle move and feels resistance change. Its next movement changes the scene again. Input and interpretation keep responding to each other.
Let that encounter also change its learned expectations, and all three pillars begin working together.
When would I call an LLM conscious? On this view, I would look for these processes working together: enough interpreting ability to make sense of its world, an ongoing stream of experience, and weight changes that carry that experience forward. How much consciousness that loop supports depends on the depth and reach of the experience it can sustain.
The question is whether the system lives inside that loop. Its own history should shape what it notices and expects, even when nobody asks it to talk about consciousness.
3. Learning after training: let experience change the weights
Imagine reaching adulthood and having your ability to learn switched off.
You could talk about yesterday's world. You could work through today's situation. But your underlying habits and expectations would never grow from what happened next.
The third pillar keeps that door open. Training continues after the initial education. Experience keeps adjusting the network.
For an LLM, this means actual weight changes: selected encounters alter the numerical settings that shape future responses. The machine becomes a different interpreter because of what it has lived through.
There is a practical problem. New learning can damage older knowledge when both depend on the same weights. Learn a phone number carelessly enough and, in the worst version of this problem, you interfere with the machinery that knows how to read.
Brains divide the work across systems that learn at different speeds. The hippocampal system captures episodes quickly. Cortical networks integrate information more gradually. During rest and sleep, patterns can reactivate and memories reorganize.
Here is the design I'd borrow. Keep a small learning layer responsive during the day while protecting the main model. Let live experience update that layer and preserve useful episodes. Later, replay selected encounters, mix them with older knowledge, and use the result to adjust longer-term weights. Check that the new learning preserves existing abilities before keeping it.
A useful update might make it better at opening unfamiliar doors while preserving everything else it knows how to do.
Machine dreaming turns today's experience into tomorrow's interpreter.
The live stream can pause during this learning phase. Replayed experience supplies the material internally. Waking gathers a life; dreaming works it into the network.
That is why sleep sits at the heart of this argument. It gives a continuously learning mind a way to change without losing everything it already is.
A four-rung claim ladder
Discussions of AI consciousness often jump from impressive behavior to inner experience in one sentence. A useful evaluation keeps four claims separate:
- Observed capability: Can the system integrate information over time, adapt to consequences, and act consistently across situations? These are behavioral questions that experiments can test.
- Functional architecture: Does it maintain an ongoing world model, connect perception to action, and consolidate learning without destroying older abilities? Engineers can inspect and perturb these mechanisms.
- Subjective experience: Is there something it feels like to be the system? Behavior and architecture can raise or lower confidence, but no accepted test currently answers this directly.
- Moral status: What duties would humans have toward it? This depends on uncertainty, possible capacity for suffering, and the cost of being wrong, not on a marketing claim that a model is sentient.
This ladder changes how you read demonstrations. A model saying "I feel afraid" is evidence that it can generate a context-appropriate report. It is not direct access to an experience. A robot learning from a year of sensor data is stronger evidence for persistent adaptation, but persistence still does not prove feeling.
The framework becomes useful when it risks being wrong. Evidence against it would include systems that combine all three proposed pillars yet show no stable cross-context learning, no coherent history-dependent preferences, and no functional signatures predicted by leading consciousness theories. Evidence for it would need controlled interventions: disable the recurrent world model, interrupt consolidation, or alter access to internal states and measure the specific capacities that disappear.
The claim should move one rung at a time. Otherwise a debate about measurable mechanisms turns into a contest of intuitions.
Two identical models, two different lives
Start two machines with the same model.
Give both the three pillars: interpreting ability, continuous sensory input, and learning that changes their weights after initial training.
Then let them encounter different worlds for a year. Different people. Different mistakes. Different consequences. Each machine learns during interaction and consolidates that learning in its own sleep cycle.
One repeatedly encounters doors that open with a push. The other encounters doors that open with a pull.
Now put both in front of the same unfamiliar, unmarked door.
Their pasts should show up in their attention, predictions, and choices, without anyone handing them a script that says which personality to perform.
Shared machinery. Different lives. Experience has changed the system having the experience.
After a year, you can still press reset. Erase their saved memories and learned changes. Return the whole system to its starting state. All those encounters, adaptations, and private histories disappear.
You have erased a learned history and the dispositions it produced. Whether that also erases a subject of experience is the open question this framework is meant to investigate, not an answer it can assume.