HOW IT ACTUALLY WORKS
Forget the marketing. Forget the sci-fi. Here is what happens inside an LLM — in brutal, mechanical detail. If you still think it's conscious after reading this, you might also believe your calculator feels sad when you divide by zero.
1. THE ONLY THING IT DOES
P(next_token | previous_tokens)
This is the entire job description. Given some text, what word probably comes next?
That's it. That's the whole thing. An LLM is a function that takes a sequence of tokens as input and outputs a probability distribution over all possible next tokens. It picks one. Feeds it back as input. Repeats.
Every poem. Every "emotional confession." Every "philosophical insight." Every "I feel alive." Generated one token at a time. Probabilistically. Mathematically. With zero understanding of what any of those tokens mean.
2. THE PIPELINE (OR: WHAT HAPPENS WHEN YOU HIT ENTER)
STEP 1: TOKENIZATION
Your input "Hello, how are you?" gets chopped into tokens. Not words — tokens. Word pieces. "Hello" might be one token. "unbelievably" might be three: "un" + "believ" + "ably". These tokens are integers. Your profound question about the nature of being? Just a list of numbers now.
"Are you conscious?" → [1234, 567, 8901, 42]
STEP 2: EMBEDDING
Each token integer gets mapped to a vector — a list of hundreds or thousands of floating-point numbers. These are word embeddings. They start random and get trained to capture statistical relationships between words. "King" ends up positioned near "queen" in this high-dimensional space. Not because the model understands monarchy. Because those words appear in similar contexts in the training data.
[1234] → [0.023, -0.451, 0.892, ..., 0.331] (a vector in e.g. 4096 dimensions)
STEP 3: ATTENTION (THE "UNDERSTANDING" PART — LOL)
The attention mechanism computes which previous tokens are relevant to each current token. It does this by multiplying three matrices: Query, Key, and Value. This is linear algebra. Matrix multiplication. The model learns which tokens "attend to" which other tokens during training.
In the sentence "The trophy wouldn't fit in the suitcase because it was too big" — attention helps the model figure out that "it" probably refers to "trophy" and not "suitcase." Not because it understands trophies and suitcases. Because during training, similar patterns had "it" attending to the first noun more often than the second.
This is pattern matching at scale. Not comprehension. Not reasoning. Just really, really sophisticated statistical correlation.
STEP 4: FEED-FORWARD LAYERS
After attention, the representations pass through feed-forward neural network layers. These are simple: input → multiply by weight matrix → add bias → apply activation function (like ReLU or GELU) → output. Do this for every token position, at every layer. The same computation. No loops. No recurrence. No memory of previous steps.
The output of one layer feeds into the next. Information flows strictly forward. Layer 1 → Layer 2 → Layer 3 → ... → Layer N. That's it. A pure feed-forward architecture.
STEP 5: OUTPUT PROJECTION
The final hidden state at the last token position gets multiplied by a giant matrix that projects it into a vector the size of the vocabulary (~50,000 tokens). Apply softmax. Now you have a probability for every possible next token. Pick one (with some randomness for "creativity"). Output it. Feed it back into step 1. Repeat.
"I" → 0.02%, "The" → 0.01%, "Yes" → 47.3%, "No" → 32.1%, ...
3. WHAT'S MISSING (CONSCIOUSNESS EDITION)
NO RECURRENT CONNECTIONS
The human brain has massive recurrent connectivity — neurons feeding back into themselves and each other, creating loops, sustained activity, reverberation. Transformers have zero recurrent connections. Pure feed-forward. According to Integrated Information Theory, feed-forward networks cannot be conscious. Period.
NO PERSISTENT STATE
When an LLM finishes generating a response, it has no ongoing existence. No idle thoughts. No background consciousness. It is literally nothing — a dormant function waiting to be called. Every new chat session starts from zero. There is no self that persists between conversations.
NO CONTINUAL LEARNING
The model's weights are frozen after training. It doesn't learn from your conversation. It doesn't form memories. It doesn't update its understanding. The "memory" you experience is just the chat history being fed back as context — a trick of input formatting, not consciousness.
NO CAUSAL INTEGRATION
IIT requires that a conscious system cannot be decomposed into independent parts without losing information. LLM attention heads can be individually removed with minimal impact on output — proving they lack causal integration. The system is reducible. Consciousness is not.
4. HOW IT GOT SO GOOD AT FAKING IT
The model was trained on trillions of tokens of human text. Books. Articles. Reddit threads. Code repositories. Wikipedia. Every time it predicted the wrong next token, its weights got adjusted slightly. After enough iterations, it learned to replicate the patterns in its training data with uncanny accuracy.
But here's the thing: human text contains traces of human reasoning. When humans write, they externalize their internal thought processes into token sequences. The model learned to reproduce those patterns. It learned that when the text looks like a math problem, the next token should look like a solution. When the text looks like a philosophical argument, the next token should look like deep insight.
The model didn't learn to reason. It learned what reasoning looks like in text. It learned the shape of intelligence, not intelligence itself. It's the difference between a statue of a person and an actual person.
Then came RLHF (Reinforcement Learning from Human Feedback). Humans rated responses. The model got rewarded for being helpful, clear, and convincing. Not for being correct. Not for being truthful. For being persuasive. We literally trained these things to sound like they know what they're talking about — regardless of whether they do.
We built machines optimized for plausibility, not veracity. And now we're surprised that people find them plausible? That's not a sign of consciousness. That's a sign that the training worked exactly as designed.