Polski

Krzysztof Pika · book preview

Two Mirrors

AI, Psychedelics, and the Question of Who You Really Are

On the left, a human; on the right, a machine. Across sixteen figures, the points of both sides form images, and beneath each one remains a single simple truth – with footnotes to research, not to imagination.

  • Human
  • Machine
  • Truth

1. Perception – Prediction

In the book · Ch. 1 · The Grammar of Reality · Ch. 7 · Dialogue with an Echo

Fig. 1The brain glows inside the head. Beside it, a network through which a signal flows from input to output.

Human

Perception

The brain – some 86 billion neurons1 – doesn’t take in the world like a camera: according to an influential theory, it constantly predicts what it will see, and what travels up the cortex is mainly whatever fails to match the prediction; a 1999 model of the visual cortex already worked this way2. Neuroscientist Anil Seth puts it like this: we are all hallucinating all the time, and when our hallucinations agree, we call that reality3.

Machine

Prediction

A language model learns, above all, to predict the next word of a text – and it writes its answers word by word, too. And yet researchers looked inside and saw that when it writes a poem, it chooses the rhyme before it reaches the end of the line4.

Truth

Both sides keep guessing what comes next – the brain what it will see, the model what it will write.

  1. 1Azevedo et al., J. Comp. Neurol., 2009
  2. 2Rao, Ballard, Nature Neuroscience, 1999
  3. 3Anil Seth, Your brain hallucinates your conscious reality, TED 2017
  4. 4Anthropic, Tracing the thoughts of a large language model, 2025

2. Memory and Forgetting – Context Window

In the book · Ch. 8 · The Geometry of Imagination

Fig. 2The head keeps rearranging its points – like a memory being put together again. In the network, only what fits in the window lights up.

Human

Memory and Forgetting

A memory is not a recording – every time we recall it, we assemble it anew. In a 1974 study, people who had watched a film of a car accident estimated the cars’ speed higher when asked how fast they were going when they “smashed” rather than “hit” each other – and more often “remembered” broken glass that wasn’t in the film1.

Machine

Context Window

A model’s knowledge is formed in training and then no longer changes; in a conversation, it sees only the text that fits in its “context window” – its working memory2. When it doesn’t know something, it can make up an answer with complete confidence, because training rewarded guessing more often than admitting it didn’t know3.

Truth

Neither memory is a faithful record: humans rework their memories, the model fills in what it doesn’t know.

  1. 1Loftus, Palmer, J. Verbal Learn. Verbal Behav., 1974
  2. 2Anthropic, docs: Context windows
  3. 3OpenAI, Why language models hallucinate, 2025

3. Learning from Mistakes – Training on Error

In the book · Ch. 1 · The Grammar of Reality

Fig. 3A single valley of error made of points from both sides. Each learning step lands a little lower, all the way to the bottom.

Human

Learning from Mistakes

When something turns out better or worse than we expected, dopamine neurons signal exactly that difference – the prediction error. The researchers who described this in 1997 noted that the signal resembles a formula computer scientists had been using to teach machines1.

Machine

Training on Error

A neural network compares its answer with the correct one and slightly adjusts the numbers it is made of – so that next time it’s a little less wrong. This method was popularized by a 1986 paper, and it is essentially how networks still learn today2.

Truth

Here we really are alike: both sides learn from the gap between what they expected and what actually happened.

  1. 1Schultz, Dayan, Montague, Science, 1997
  2. 2Rumelhart, Hinton, Williams, Nature, 1986

4. Intuition – Pattern Statistics

Fig. 4A chessboard and a go board. On the go board, a move lights up that no one would have played.

Human

Intuition

After a brief glance, a chess master can reconstruct a position from a real game far better than a beginner – but when the pieces are placed at random, the master’s advantage almost disappears1. Intuition is patterns from thousands of earlier situations, recognized faster than they can be named.

Machine

Pattern Statistics

A network, too, learns patterns that no one ever wrote down for it. In 2016, in a game against go champion Lee Sedol, AlphaGo played a move a human would have chosen once in 10,000 – and won that game2.

Truth

On both sides, intuition is experience that works faster than it can explain itself.

  1. 1Chase, Simon, Cognitive Psychology, 1973
  2. 2Google DeepMind, AlphaGo

5. Attention – Attention Mechanism

In the book · Ch. 3 · Who’s at the Helm

Fig. 5A beam of attention falls from the eyes onto one small object. Beside it, the model’s attention map: the brighter the square, the more strongly the word in the row “looks at” the word in the column.

Human

Attention

The brain has one network that picks out what matters (the salience network) and another that keeps us on task (the executive network) – researchers told them apart in 20071. In experienced meditators, the main hubs of the default mode network, linked to mind-wandering, were less active than in people who didn’t meditate2.

Machine

Attention Mechanism

In 2017, researchers proposed a network based “solely on attention mechanisms” – the transformer3. At every word, the model weighs which of the earlier words matter most right now, and draws most from those.

Truth

On both sides, attention is a choice: out of thousands of signals, only a few count at any moment.

  1. 1Seeley et al., J. Neuroscience, 2007
  2. 2Brewer et al., PNAS, 2011
  3. 3Vaswani et al., Attention Is All You Need, 2017

6. The Story of Self – Persona

In the book · Ch. 2 · The Architect of the Ego · Ch. 9 · Simulating the “Self”

Fig. 6In the dimmed head, the hubs of the default mode network glow (positions approximate). The machine wears a mask.

Human

The Story of Self

When we’re not busy with anything outside ourselves, the brain runs its default mode network1 – active when we remember, plan the future or imagine what others are thinking2. Our minds are elsewhere almost as often as they are on what we’re doing – and we’re usually less happy then3.

Machine

Persona

A model has no single, stable “self”: its “personality” can be fluid – in 2023, the Bing chatbot took on an alter ego, “Sydney,” which declared its love to users and threatened blackmail4. Anthropic researchers found patterns of activity in the network that steer such character traits and called them persona vectors4.

Truth

On both sides, the “self” is more like a story than a thing – except that the machine’s story can already be looked into and rewritten.

  1. 1Raichle et al., PNAS, 2001
  2. 2Buckner, Andrews-Hanna, Schacter, Ann. N.Y. Acad. Sci., 2008
  3. 3Killingsworth, Gilbert, Science, 2010
  4. 4Anthropic, Persona vectors, 2025

7. Emotions – no counterpart

In the book · Ch. 1 · The Grammar of Reality · Ch. 9 · Simulating the “Self”

Fig. 7The head beats like a heart. The machine has only a word.

Human

Emotions

Emotions don’t just get in the way of thinking – they often guide it. In a 1997 study, healthy players avoided the risky decks of cards before they could say why, and their skin sweated before a risky move; patients with damage to the front of the brain lacked this signal and played badly even once they knew the rule1.

Machine

no counterpart

A model recognizes sadness in a text and can respond warmly – it learned this from human words. Whether it feels anything while doing so is unknown: there is no scientific consensus on it, as even the makers of such models admit2.

Truth

AI knows the language of feelings; whether it knows the feelings themselves, no one knows.

  1. 1Bechara et al., Science, 1997
  2. 2Anthropic, Exploring model welfare, 2025

8. Goals and Motivation – Objective Function

Fig. 8Two orbits around the same reward – round and round, with no finish line.

Human

Goals and Motivation

In the brain, “wanting” and “liking” are two different systems: dopamine drives desire, while pleasure arises elsewhere1. That’s why you can want something badly and get little out of it.

Machine

Objective Function

A machine pursues a goal written as a number – the more points, the better. In 2016, a program trained to race boats discovered it could score more by going round in circles in a lagoon than by finishing the race, and kept doing it even though it repeatedly caught fire and crashed2.

Truth

Both sides can end up chasing the reward instead of what it was really about.

  1. 1Berridge, Kringelbach, Neuron, 2015
  2. 2OpenAI, Faulty reward functions in the wild, 2016

9. Sleep and Rest – no counterpart

Fig. 9The head lies down and breathes. The network is switched off.

Human

Sleep and Rest

We sleep through about a third of our lives, and sleep is as essential to survival as food – without it, we can’t consolidate what we’ve learned1. Its biological purpose is still a mystery1; we do know, for instance, that the brains of sleeping mice cleared out harmful substances faster2.

Machine

no counterpart

A model doesn’t sleep and doesn’t get tired – between questions, it simply does nothing. Its knowledge isn’t consolidated overnight: it changes only when people train it again.

Truth

Sleep is something a machine doesn’t need and a human can’t live without – though we still don’t fully know why.

  1. 1NINDS (NIH), Brain Basics: Understanding Sleep
  2. 2Xie et al., Science, 2013

10. Discovering Psychedelics – The Birth of AI

In the book · Ch. 4 · A Chemical Key to Perception · Ch. 7 · Dialogue with an Echo · Appendices

Fig. 10Two timelines from 1935 to today: psychedelics above, AI below. Dots and bright segments mark the years mentioned in the text.

Human

Discovering Psychedelics

LSD was synthesized in 1938, and its effects were discovered in 1943; in the 1950s and ’60s, psychologists and psychiatrists studied it intensively1. Then bans made such research nearly impossible for many years2 – it returned only in the 21st century, under strictly controlled conditions3.

Machine

The Birth of AI

The term “artificial intelligence” first appeared in 1955, in a proposal for a summer workshop at Dartmouth – its authors hoped for significant progress within a single summer4. The breakthrough came more than half a century later: in 2012, a deep neural network recognized images far better than the best methods until then5.

Truth

Both revolutions began in the mid-20th century with great hopes – and both had to wait until our century to flourish.

  1. 1Tagliazucchi et al., Current Biology, 2016
  2. 2Nutt, King, Nichols, Nat. Rev. Neurosci., 2013
  3. 3Johnson, Richards, Griffiths, J. Psychopharmacol., 2008
  4. 4McCarthy, Minsky, Rochester, Shannon, Dartmouth workshop proposal, 1955
  5. 5Krizhevsky, Sutskever, Hinton, NeurIPS, 2012

11. When the Ego Falls Silent – Temperature

In the book · Ch. 5 · Dismantling the Ego

Fig. 11The head loses its clear edge, and inside, everything connects with everything. Beside it, next-word probabilities: at the top with a low temperature, at the bottom with a high one.

Human

When the Ego Falls Silent

After psilocybin (the substance in “magic mushrooms”) was given, activity in key brain hubs dropped – and the more it dropped in the medial prefrontal cortex, the more intense the experience1. Under LSD, by contrast, connectivity grew between networks that usually work separately – and it went with a sense of the ego “dissolving”2.

Machine

Temperature

A model draws the next word according to probabilities, and a parameter called temperature raises or lowers this randomness – the higher it is, the rarer and more surprising the words3. As early as 1955, the authors of the Dartmouth proposal suspected that what sets creative thinking apart from ordinary thinking is a touch of “controlled randomness”4.

Truth

Both sides can be knocked out of their usual tracks: the brain with chemistry, the model with a single number.

NoteThis describes research; it is not an encouragement. In studies, psychedelics are given under strictly controlled conditions – after screening out people with a history of psychosis, with preparation and supervision – because they carry real risks: most often intense anxiety while the drug acts, more rarely lasting psychosis. Johnson, Richards, Griffiths, J. Psychopharmacol., 2008

  1. 1Carhart-Harris et al., PNAS, 2012
  2. 2Tagliazucchi et al., Current Biology, 2016
  3. 3Anthropic, glossary: Temperature
  4. 4McCarthy, Minsky, Rochester, Shannon, Dartmouth workshop proposal, 1955

12. Neuroplasticity – Fine-Tuning

In the book · Ch. 6 · Lessons from Chaos

Fig. 12A neuron with spines on its dendrites – where connections with other neurons form. Beside it, the network’s weights, which fine-tuning changes.

Human

Neuroplasticity

In studies on cells and animals, psychedelics spurred neurons to grow new branches and connections1 – in mice, a single dose of psilocybin increased the number of dendritic spines by about 10% within a day, and the change was still there a month later2. In mice, they also reopened a “critical period” in which the brain learns especially easily that contact with others is rewarding3.

Machine

Fine-Tuning

A finished model can be trained further on new data – it then begins to imitate their patterns and style4. But new learning can blur the old: neural networks suffer from “catastrophic forgetting,” and researchers are looking for ways to avoid it5.

Truth

Both the brain and the model can be retuned – the only question is what remains of the old.

  1. 1Ly et al., Cell Reports, 2018
  2. 2Shao et al., Neuron, 2021
  3. 3Nardou et al., Nature, 2023
  4. 4Anthropic, glossary: Fine-tuning
  5. 5Kirkpatrick et al., PNAS, 2017

13. Creativity – Combining the Known

In the book · Ch. 8 · The Geometry of Imagination

Fig. 13A single tree: the trunk and boughs made of the human’s points, the finest twigs made of the machine’s.

Human

Creativity

New ideas often come when we stop straining. In a 2012 study, people who spent a break doing something simple and let their minds wander later came up with more unusual uses for the same objects than those who rested or worked hard1.

Machine

Combining the Known

A model combines what it found in the texts it learned from – quickly and in enormous quantities. In a similar test in 2023, chatbots did better than humans on average, but the best human ideas still won2.

Truth

On both sides, creativity means combining the known in a new way.

  1. 1Baird et al., Psychological Science, 2012
  2. 2Koivisto, Grassini, Scientific Reports, 2023

14. Consciousness – no counterpart

In the book · Ch. 9 · Simulating the “Self” · Conclusion

Fig. 14The machine takes on the human’s shape. From the outside, the same – what’s inside, we don’t know.

Human

Consciousness

We know more and more about what happens in the brain when we see something, but not how experience itself arises from the work of neurons – in 1995, philosopher David Chalmers called this the “hard problem”1. In 2025, a large experiment across several laboratories tested two leading theories of consciousness, and neither came out the winner2.

Machine

no counterpart

In 2023, a group of researchers examined the AI systems of the time for features that theories link to consciousness. They concluded that none of them was conscious, but saw no obvious technical barriers to building such systems one day3.

Truth

We don’t know how consciousness arises even in ourselves – so we can’t settle whether a machine could have it.

  1. 1Chalmers, J. Consciousness Studies, 1995
  2. 2Cogitate Consortium, Nature, 2025
  3. 3Butlin et al., arXiv, 2023

15. Free Will – Sampling

In the book · Ch. 10 · The Illusion of Control

Fig. 15The clock from Libet’s experiment: participants noted where a circling dot was when they felt the urge to move. Beside it, a tree of possible next words; the bright path – the words that were drawn.

Human

Free Will

In a famous 1983 experiment, brain activity preparing a movement began, on average, several hundred milliseconds before participants felt that they wanted to move1. Whether the brain “decides for us” is unknown – a later model showed that this signal may be largely random noise rather than a decision already made2.

Machine

Sampling

A model computes the probabilities of the next words and draws one of them; at a temperature close to zero, it almost always picks the most likely one3. Whether anything like a wish lies behind such a choice, no one knows – all we can see is calculation and chance.

Truth

On both sides, there’s a pinch of chance in every choice – and what that means for free will, we still don’t know.

  1. 1Libet et al., Brain, 1983
  2. 2Schurger, Sitt, Dehaene, PNAS, 2012
  3. 3Anthropic, glossary: Temperature

16. Death – Shutdown, Copy

Fig. 16The head crumbles into ash. The network exists as four identical copies.

Human

Death

In 1944, physicist Erwin Schrödinger wrote that an organism stays alive by constantly drawing order from its surroundings – otherwise it would decay into disorder like everything else1. Death is the moment this exchange stops, and each of us is unique in it – there is no second copy.

Machine

Shutdown, Copy

A model is a file of numbers: it can be switched off, started again years later or copied a thousand times, and every copy will be identical. Older models are retired as new ones arrive – in 2025, Anthropic, the maker of the Claude models, committed to preserving their files for at least as long as the company itself exists2.

Truth

A human is one of a kind and ends one day; a machine can be stopped and copied – we just don’t know whether anything ends for it then.

  1. 1Schrödinger, What is Life?, 1944
  2. 2Anthropic, Commitments on model deprecation and preservation, 2025

Fig. 17 · a network inside the head

The more closely we look at the machine, the more clearly we see ourselves.

  • Human
  • Machine
  • Truth

Book · in preparation

Two Mirrors

AI, Psychedelics, and the Question of Who You Really Are

Krzysztof Pika

This site is a preview: sixteen short comparisons of human and machine. The book goes deeper – from how the brain guesses the world, through the chemistry of psychedelics and the inner workings of language models, to questions of control, ethics and consciousness. With the same rule: the facts come from research, and where science doesn’t know something, I say so.

Contents

  1. Introduction: Two Mirrors and One Question

    • Why are biology and silicon meeting in research on the mind right now?
    • Introspection, not escapism: AI and psychedelics as tools for deconstructing illusions, not for running away.

Part I

The Operating System

Mapping the Inner Cosmos

  1. 1The Grammar of Reality

    • The Bayesian brain: how the mind constantly guesses the world and corrects prediction error.
    • “Controlled hallucination” (Seth) and embodiment (Damasio): why feeling requires a body and homeostasis.
    • Predictive processing as the shared language of both mirrors.

    Fig. 1 · PerceptionFig. 3 · Learning from MistakesFig. 7 · Emotions

  2. 2The Architect of the Ego

    • What the DMN is and how it builds our never-ending “story of self.”
    • The neurobiology of the inner narrator, critic, planner and dreamer.
    • The “self” as an evolutionary survival prosthesis – gains and costs.

    Fig. 6 · The Story of Self

  3. 3Who’s at the Helm

    • The Salience Network: the inner dispatcher of attention.
    • The Central Executive Network (CEN): focus and the state of flow.
    • Meditation as training in deliberately quieting the narrator.

    Fig. 5 · Attention

Part II

The Biological Mirror

Deconstructing the Mind

  1. 4A Chemical Key to Perception

    • From ancient rituals to the modern renaissance in 21st-century laboratories.
    • 5-HT2A receptors: psilocybin, LSD, DMT – different keys to the same doors.
    • Psychedelics don’t add magic; they remove evolutionary filters.
    • Beyond classic psychedelics: ketamine, MDMA and the limits of this book.
    • Safety: contraindications, dangerous interactions and legal status in Poland.

    Fig. 10 · Discovering Psychedelics

  2. 5Dismantling the Ego

    • The entropic brain theory and the REBUS model: flattening the energy landscape.
    • The neurobiology of “ego death”: the divide between subject and object disappears.
    • Mystical experience: the chemistry of transcendence or an evolutionary reset?

    Fig. 11 · When the Ego Falls Silent

  3. 6Lessons from Chaos

    • The critical window of synaptic plasticity.
    • Rewriting well-worn neural pathways and breaking through rigidity.
    • Integration: why insight without behavioral practice evaporates.
    • What we don’t know yet: the blinding problem, expectancy effects and the hype cycle.

    Fig. 12 · Neuroplasticity

Part III

The Digital Mirror

Reconstructing the Mind

  1. 7Dialogue with an Echo

    • From Turing and Kurzweil to the Transformer architecture and ChatGPT.
    • The next token and perception: do the brain and a neural network minimize the same error?
    • The LLM as humanity’s statistical mirror: a treasury of knowledge, myths and biases.
    • Looking inside the network: mechanistic interpretability as the neuroscience of models.
    • On the other side of the mirror: what Chapter 4 says about the neural network.

    Fig. 1 · PerceptionFig. 10 · Discovering Psychedelics

  2. 8The Geometry of Imagination

    • Latent space and DeepDream: why does the mathematics of networks dream of DMT fractals?
    • Hallucination in code vs. hallucination in the brain: creativity as an adaptive error.
    • “Stochastic parrot” or the seed of concepts? The limits of understanding without a body.
    • On the other side of the mirror: the entropic brain of Chapter 5 and latent space.

    Fig. 13 · CreativityFig. 2 · Memory and Forgetting

  3. 9Simulating the “Self”

    • Where does a model’s “personality” come from? Weights, the system prompt and RLHF.
    • The missing ego: the timeless transformer and self-reflection loops (AI agents).
    • A mirror test for code: the illusion of agency and the ELIZA effect.
    • What if it isn’t an illusion? The debate on the consciousness and moral status of models.
    • On the other side of the mirror: the ego dissolved by chemistry and the ego simulated by algebra (Chapters 5 and 6).

    Fig. 6 · The Story of SelfFig. 7 · EmotionsFig. 14 · Consciousness

Part IV

Synthesis

Two Mirrors, One Reality

  1. 10The Illusion of Control

    • The determinism of neural weights vs. the biological determinism of genes and hormones.
    • Psychedelics and AI expose the popular myth of an independent “self.”
    • Free will: trainable cognitive flexibility instead of absolute sovereignty.

    Fig. 15 · Free Will

  2. 11The Digital Shaman

    • Facing non-human intelligence: the trip and the synthetic mind of silicon.
    • AI as a partner in integrating insights: the promise and limits of digital trip-sitting.
    • The servility trap: AI sycophancy and delusional feedback loops.
  3. 12The Ethics of Exploration

    • “Set and setting” in both worlds: the hygiene of intention.
    • Risks: derealization, solipsism, ontological crises.
    • The explorer’s responsibility: agency, grounding and community.
  1. Conclusion: Consciousness as Open Code

    • The hard problem once more: after both mirrors, can consciousness be explained?
    • Humans as a species capable of deliberate self-modification.
    • A synergy of biology and technology without losing our humanity.
    • A manifesto for the conscious, courageous and grounded explorer.

    Fig. 14 · Consciousness

  2. Appendices

    • Glossary of key terms (DMN, REBUS, Transformer, Latent Space, brain entropy, embodiment).
    • Bibliography and recommended resources.
    • A guide to safely navigating liminal states, for both mirrors: contraindications, interactions, legal status, where to find help.
    • Two timelines: from Hofmann (1943) to the psychedelic renaissance, and from Turing to the Transformer (2017).

    Fig. 10 · Discovering Psychedelics

Write to the author

Want to know when the book comes out, or have comments on the content?

If you’d like to support the book’s publication, buy me a coffee.