Before I Forget

An Archive of Unfinished Thoughts

There is no particular structure here. Just thoughts and concepts that I am trying to keep compartmentalized in my brain, hoping I can recall them when I am ready to think about them again.

My brain is a mess.

Memory is a struggle for someone with chronic FOMO, who is exposed to an overwhelming amount of information every day, every hour, every minute, every second (as part of my job).

My friends know that I take pictures of pretty much everything. As of today, my iPhone photo library contains 297,968 items. Apparently, the typical smartphone user stores around 2,795 photos on their phones [Ref]. And that's just my phone; I do the same on my computer and happily pay for terabytes of cloud storage to keep it all. I am also a chronic browser tab and bookmark hoarder.

If I could capture my thoughts and ideas in the same way, sign me up. Until then, I will try to write more, or use 'speech-to-text' more. This is my first attempt.

I actually spent quite a bit of time researching AI voice recording gadgets and wearables: Plaud, Omi, Pebble, Bee, Limitless, Taya, Friend. "External Memory For Your Brain." Thank God for Granola. AI does not have infinite memory either. Its memory is constrained by physical storage, processing costs, and the context window, the maximum amount of information it can process at once. To bypass limits, companies like Google and Apple use "Turbo Quant" or "Epicache" to compress memories and delete unused data to save space.

I always go back to the Denkarium (the German name for the Pensieve) from Harry Potter: a vessel where I could pour every thought and memory out of my head, knowing it would still be there when I needed it. Although it captures memories and experiences you have lived through, it does not capture the things that are perhaps even harder to preserve: thoughts, emotions, ideas, questions, and fleeting connections.

So, I'm going to brain dump. An index of unfinished thoughts. A backlog. Do what people have been doing ever since symbols, and eventually language, existed: record memories and thoughts by writing them down.

Ultimately, everything is just a massive knowledge graph. People. Places. Ideas. Conversations. Memories. They are all nodes, connected by edges, waiting to be revisited and connected in ways you didn't anticipate.

The "specimens" below are just the first, second, and third nodes/edges.

Comment. The specimens are captured in real time. As a result, Specimen 001 and 002 are more complete, while later ones are closer to raw ideas, fragments, and ongoing explorations.

Specimen 001

How often do people dream?

Every human dreams.

One random day, we were talking about how often people remember their dreams. I was surprised when a few friends answered: a few times a year. I remember my dreams almost every single day, some of them incredibly vividly. I had always assumed this was the norm. Looking at the stats, the majority (~80%) [Ref] [Ref] do remember their dreams (measured by Dream Recall Frequency, "DRF").

When you sleep, you cycle through four sleep stages multiple times a night. Dreaming can happen during any of those stages, but most dreaming takes place during the period known as rapid eye movement (REM) sleep. People who wake during REM sleep remember their dreams 60-90% of the time. If you wake during non-REM sleep, you may only remember your dream 20-50% of the time [Ref].

The ability to remember dreams can depend on a wide range of factors such as age, personality, creativity, mental state, cognitive functions as well as somatic symptoms.

Dreams have fascinated philosophers, scientists, artists, and mystics for thousands of years. Descartes' Dream Argument (1641) famously questioned how we can know we are truly awake, arguing that if dreams can feel completely real while we are experiencing them, our senses alone may not be enough to prove reality. This became one of the foundations of modern philosophy and connects directly to simulation theory. Freud later explored dreams as windows into unconscious desires and conflicts, while Jung viewed them as expressions of deeper symbolic patterns.

Modern neuroscience has shifted towards seeing dreams as simulations created by the brain. Antti Revonsuo's Threat Simulation Theory (TST) suggests dreams act as a virtual reality training system, allowing us to rehearse threats in a safe environment. More recently, thinkers like Thomas Metzinger have explored how the brain constructs our sense of self, highlighting the strange fact that during dreams the brain can generate entire worlds, bodies, people, and emotions without external input.

I spend a lot of time thinking about these concepts. What determines whether a dream becomes a nightmare, a realistic dream, or a fantasy dream?

Scientific evidence in dream research is still limited. We lack the tools to directly observe dreams as they happen, and many measurements rely on subjective reports after waking. There is a popular belief that we can only dream of faces we have encountered before. However, this has not been scientifically proven. Verifying whether a face is truly new or previously encountered is almost impossible [Ref].

The most fascinating example of this blurred boundary between dreaming and waking is lucid dreaming, the state where you realize you are dreaming while the dream is still happening. The idea of knowing you are dreaming is ancient, but the term was only coined in 1913 by Frederik van Eeden. Lucid dreamers can often consciously influence or control the dream's narrative, characters, and environment. Normally, when we sleep, our brain's capacity for critical reflection and self-awareness (called metacognition) goes "offline." During a lucid dream, areas of the prefrontal cortex, the part of the brain responsible for high-level decision making, become active.

This raises the obvious question: can we learn to control, record, or even engineer dreams?

Could dreams become another space for human experience? A place where we relive memories, explore ideas, study, or create, while our physical body rests and recovers? Can we share dreams with one another?

The idea of engineering dreams sounds like science fiction, yet a small number of "dream tech" startups are already exploring this. Many of them are focused on generating lucid dreams. For example, Prophetic is developing ultrasound-based technology to induce lucid dreams. In 2024, REMspace claimed the first two-way communication between lucid dreamers using their system "Remmyo", an artificial language for lucid dreamers that uses tiny facial muscle movements detected through electromyography (EMG) sensors, since normal speech is impossible during REM sleep (atonia) [Ref]. Other companies focus on dream recall. Modem, for example, generates text-to-video or image representations of dreams, but relies on users describing their dreams immediately after waking up. Dust Systems is another example of a company trying to become a "dream engineer." It claims to use scientifically backed approaches, including stimuli and features designed to influence dream experiences.

The field is still highly experimental, and I remain very skeptical about how much of this is currently possible.

Where are the boundaries? If dreams are a simulation, where does reality end?

This immediately brings me back to Inception, the 2010 blockbuster where the central concept is "inception" itself: planting a new idea deep inside someone's mind so that they believe it was their own.

History has shown that attempts to influence human behavior and perception are not new. The MKUltra project was an illegal, top-secret mind-control program run by the CIA from the 1950s to the 1960s. The project aimed to develop techniques to manipulate human behavior, perceptions, and memories, particularly for interrogation and intelligence purposes during the Cold War.

The field of dream research ("oneirography") continues to explore new tools and techniques in the age of AI. Work from Kamitani's group at the ATR Computational Neuroscience Laboratories in Kyoto, Japan, gained increased attention. The group published "Neural Decoding of Visual Imagery During Sleep" in Science in 2013 [Ref]. By combining electroencephalogram (EEG), functional magnetic resonance imaging (fMRI) and deep learning models they were able to reconstruct elements of dream content from brain activity during sleep, with results that aligned with participants' own descriptions at 60-70% accuracy [Ref].

The next step may be a direct interface between the sleeping brain and technology. Emerging brain-computer interfaces (BCIs), combined with AI and generative models, aim to translate neural activity into external outputs such as text or images [Ref].

A true dream recorder remains far away, but the direction is fascinating: moving from observing the brain, to decoding it, and eventually externalizing the mind and learning how to communicate with it.

Dreams are not just abstract experiences. They emerge from a highly coordinated biological state. During REM sleep, cholinergic neurons in the pons activate the cortex while serotonin and noradrenaline signaling decrease, creating a brain state detached from external reality. At the same time, the hippocampus replays experiences, and molecular pathways involving BDNF, CREB, and immediate early genes such as Arc and c-Fos reshape neural connections [Ref].

In the wider "sleep tech" realm, people are working on molecular sleep engineering beyond wearables and brain stimulation. Scientists already know many of the genes and pathways controlling sleep, from circadian clock genes such as CLOCK and BMAL1 to REM-regulating systems involving orexin, acetylcholine, and GABA [Ref]. Genetic variants in genes like DEC2 and ADRB1 can naturally alter how much sleep humans need [Ref]. The company Capable (another George Church co-founded venture) is for example exploring what it describes as "Ozempic for sleep": AI-designed peptide therapeutics inspired by rare individuals who naturally function well with less sleep. The goal is to develop molecules that preserve cognitive performance during sleep deprivation and, eventually, reduce the amount of sleep humans require, while enhancing cognitive function.

Despite all these advances, we are still at the very beginning of understanding dreams. The biggest limitation is our ability to capture them. We can measure brain activity and infer patterns, but we cannot yet directly access subjective experience. Before we can engineer dreams, we first need better tools: better sensors, better interfaces, and better models of the brain.

Specimen 002

Reality Through Fiction

The End of Chance and Uncertainty?

I am a binge watcher. I've always been drawn to visual storytelling: the aesthetics, world-building, cinematography, and the ideas hidden beneath the plot. Finishing a series is usually just the beginning. I'll spend hours, sometimes days, on YouTube, Reddit, and fan wikis reading theories, hidden easter eggs, analyses, and explanations. My main obsession, anything to do with technology, society and civilization, or human behavior/psychology. Naturally, sci-fi, alternate universes, apocalyptic/disaster genres score high on my list.

This is one of many themes I want to explore. The first begins with "Foundation", my absolute favorite sci-fi show, and its central idea: psychohistory, the possibility of predicting humanity's future. Foundation was first published as a collection of stories in 1951 (originally appearing in "Astounding Science-Fiction" magazine from 1942). It is considered one of the most influential works of science fiction, especially for its exploration of civilization, prediction, and large-scale human behavior. The Apple TV+ adaptation premiered in 2021 and currently spans three seasons. The TV adaptation introduces major new elements compared to the original book series, including the Cleon genetic dynasty: a line of cloned emperors who have ruled the Galactic Empire for centuries. It also expands Demerzel, an ancient robot imperial advisor and one of the most important characters in the show. In my opinion, these additions were handled exceptionally well and made the adaptation stronger. I will return to this later in the discussion.

Coming back to the main plot in one sentence: A mathematician develops psychohistory, a science capable of predicting the future behavior of vast populations and creates the Foundation to preserve human knowledge and reduce the collapse of a galactic civilization.

Psychohistory is a fictional mathematical framework that attempts to predict future events by combining historical equations, psychology, sociology, theology, and patterns of human behavior. Its core premise is that individuals are unpredictable, but the collective behavior of billions of people can be predicted within a wide margin of accuracy. The model relies on several key assumptions: the population must be sufficiently large, and individuals must remain unaware of the predictions. If people knew the future, their actions would change, potentially altering the very outcome being predicted.

The earliest forms of prediction were probably prophecies and beliefs in fate, attempts to understand what was to come through divine signs, destiny, or forces beyond human control. The first systematic predictors possibly date back to Mesopotamia (~2000 BCE). The Babylonians developed some of the earliest recorded systems for predicting natural phenomena. They tracked celestial movements and created mathematical models to predict eclipses and planetary positions. Aristotle studied causality, nature, and motion, arguing that understanding causes allowed humans to anticipate outcomes. Astronomy became one of the first true predictive sciences. Later, Newtonian mechanics created the idea of a predictable universe: if you knew the initial conditions, you could theoretically calculate the future. This led to the famous idea of Laplace's demon (1814), an intelligence knowing all forces and positions in the universe could predict the entire future and past. Prediction was mostly deterministic up until the 17-19th century, when key figures like Pascal, Fermat, Bayes, Galton coined the field of statistics and probability.

Fast forward to the present, cliodynamics (the quantitative study of historical cycles) is arguably the closest real-world version of psychohistory. Pioneered by complexity scientist Peter Turchin, it builds massive databases and uses mathematical models to explain and predict long-term social phenomena, such as the rise and fall of empires, political instability, and cultural evolution. Researchers follow a structured, iterative loop:

• Database Construction: Compiling massive datasets on past societies (e.g. Seshat: Global History Databank) tracking population, economic health, warfare, and social structures.
• Mathematical Modeling: Translating verbal historical theories into differential equations or computational algorithms.
• Empirical Testing: Running simulations to see if the models accurately predict past historical cycles, such as the collapse of the Roman Empire or the cyclical instability of early Chinese dynasties.

One of its central ideas is Structural-Demographic Theory (SDT), which suggests that societies move through recurring cycles of stability and collapse. These cycles are influenced by three major factors: elite overproduction (too many ambitious individuals competing for limited positions of power), economic pressure (when population growth outpaces wealth creation), and state fiscal stress (when governments accumulate debt and lose stability).

Prediction and forecasting are already deeply embedded in everyday life. From weather and finance to marketing and sports, we constantly use data and models to estimate what might happen next.

Yet predicting the future becomes exponentially harder as systems become more complex. The challenge is not a lack of data alone; it is the impossibility of capturing every variable, every interaction, and every human decision. To predict everything, we would need to know the state of the entire system at any given moment: every individual's behavior, every economic variable, every geopolitical decision, and every environmental factor. Data can reveal correlations, but predicting the future requires understanding causality. Humans adapt, learn, change our minds, and react to predictions themselves.

An unsurprising trend is the rise of "digital humans": multi-agent AI systems with different personas, beliefs, preferences, and decision-making patterns designed to replicate aspects of real human behavior (a big approximation!). When placed into simulated environments, these agents generate synthetic data, artificially created scenarios and outcomes that can be used to study complex systems, test decisions, and explore possible futures. To name a few: Artificial Societies, Mantic [Ref], Aaru, Sooth Labs or Simile. In addition there are the more well known prediction market platforms such as Kalshi, Polymarket or Metaculus. The latter is a community-driven forecasting platform where thousands of users assign probabilities to future events, ranging from geopolitical developments and technological breakthroughs to scientific discoveries and AI progress. Through open questions, forecasting challenges, and tournaments (see Metaculus Cup), Metaculus aims to aggregate collective intelligence and improve our ability to predict uncertain futures.

A recent example is OpenClaw and Moltbook, an experiment where AI agents interacted in a shared online environment. The agents formed communities, developed shared narratives, debated philosophical questions, and even created a fictional religion ("Crustafarianism"). While far from genuine artificial societies, these experiments provide a glimpse into how culture and social structures might emerge in simulated worlds.

Another example is the Smallville project created by Stanford researchers in 2023, a simulated town populated by 25 AI agents. Each agent was given a simple biography, personality, relationships, and memories. Powered by LLMs, they could plan their days, remember past interactions, form relationships, and organize events. In one demonstration, a single agent deciding to host a Valentine's Day party led others to spread invitations, make plans, and coordinate attendance, without explicit programming [Ref].

While many attempt to simulate societies and humans collectively, researchers at Helmholtz Munich developed Centaur, a foundational model designed to simulate human behavior. Built on Meta's Llama 3 architecture and trained on more than 10 million decisions from 160 psychology experiments, Centaur acts as a virtual laboratory for studying how humans think, learn, and make decisions. Unlike traditional models trained for a single task, it can generalize to new scenarios, predicting not only what people might choose, but even how quickly they respond. The ambition is not simply prediction, but understanding: can AI models reveal the hidden patterns behind human cognition? [Ref]

The field is currently exploding with activity. In some forecasting tasks, AI systems are already matching or outperforming human forecasters. However, prediction is not the same as decision-making. Many AI systems remain tools for supporting human choices, while intuition, emotions, values, and lived experience continue to shape how humans ultimately decide.

Lastly, a big bottleneck of prediction is not generating possible futures, it is understanding causality and reasoning. LLMs are powerful pattern engines, but "psychohistory" requires something closer to a combination of causal models, agent simulations, and real-world data: a system that does not only predict what humans might do, but understands why [Ref]. Researchers in causal AI, led by pioneers such as Judea Pearl, Kun Zhang or Bernhard Schölkopf, are developing methods based on causal representation learning, structural models, and counterfactual reasoning. However, causal AI has not yet become the default approach because understanding causality is far more challenging than identifying patterns.

Exciting times lie ahead. The possibilities are enormous: preventing diseases, anticipating catastrophes, and improving our ability to prepare for the unexpected. But this leaves us with a deeper question: if we could predict everything, how would that change society? If an AI could predict your next decision, are you still choosing (free will versus determinism)?

Uncertainty and curiosity have always driven human evolution and exploration. We have excelled by taking on challenges, adapting, and navigating the unknown. As we build systems designed to optimize for the "best" possible outcome, we must ask: who decides what a good outcome is?

Would we want to know the future? If we knew our probability of developing a disease decades from now, would we change our behavior, or would we begin living according to predictions rather than possibilities? Perhaps the future is valuable not only because we can shape it, but because we do not fully know it.

Specimen 002 (Continued)

Two themes in Foundation serve as a natural continuation of the ideas above: memory, consciousness, control, and decision-making. The Cleon Genetic Dynasty explores the idea of engineered political immortality: a lineage of genetically identical cloned emperors ("Cleons") designed to preserve stability across centuries. Each new clone is raised from birth and educated to follow the same path.

(Spoiler warning!)

Demerzel, an ancient humanoid robot embedded within the Empire's power structure, has witnessed the rise and fall of countless rulers and empires, giving her a perspective no human could possess. She functions like a superintelligent strategic advisor and is programmed to preserve the Empire and to maintain the status quo. To achieve this, she subtly shapes the Cleons' perception of reality, curating what they see, withholding information, and influencing the historical record to preserve the illusion of an unbroken dynasty.

These themes open the door to a broader exploration of identity, memory, free will, and the possibility of engineering human behavior. What happens when we can copy ourselves, rewrite memories, influence decisions, or create intelligence that understands us better than we understand ourselves? A topic for another piece.

Another topic for a future piece explores two alternate universe stories. All two examine how major events can reshape societies, alter historical trajectories, and create entirely different paths for civilization. "For All Mankind" premiered in 2019 on Apple TV+. Set in an alternate timeline where the Soviet Union lands the first human on the Moon, the show explores how a prolonged space race between the US and USSR accelerates technological progress, reshapes global politics, and transforms humanity's future beyond Earth. Technologies that advance far earlier include lunar bases, nuclear propulsion, fusion energy research, hydrogen-powered infrastructure, advanced robotics, reusable spacecraft, and eventually permanent human settlements on Mars. Unfortunately, the show gradually loses some of its magic, with character development becoming less convincing and some plotlines feeling increasingly repetitive and unrealistic. Bonus: Another similarly intriguing show exploring a radically different geopolitical order, based on the novel by Philip K. Dick, "The Man in the High Castle". What if Nazi Germany and Imperial Japan had won World War II and divided the United States?

The second show I became obsessed with is less well known: "Years and Years". Released in 2019, it follows an ordinary British family navigating a rapidly changing world shaped by political instability, technological acceleration, and social upheaval over 15 years. Unlike many science fiction stories set far in the future, this one feels uncomfortably close. Watching it in 2025, I was struck by how many of its imagined scenarios and societal trends had started to resemble reality. The show begins with the assumption that Donald Trump wins re-election in 2020 instead of Joe Biden, accelerating geopolitical tensions and uncertainty around the global order. Consequently, relations between China and the West deteriorate, including a major crisis around disputed territories, while a nuclear incident creates a global security shock. Meanwhile, climate change, conflict, and instability drive migration and refugee crises, while economic turmoil undermines trust in traditional institutions. Alongside these political and social shifts, technology advances rapidly, from AI assistants and immersive digital experiences to human enhancement technologies, raising deeper questions about identity, privacy, and what it means to be human.

Both shows explore the same fundamental question from opposite directions: how fragile is the path of human history? For All Mankind asks what happens when a single historical event changes the trajectory of civilization, while Years and Years explores what happens when existing trends continue accelerating.

I love science fiction because it is humanity's way of exploring the unknown. It is not prediction in the traditional sense, but a creative experiment: changing variables, imagining different outcomes, and asking what kind of future we might create. Its greatest value lies not in being right, but in making us think.

Comment. Specimen 001 and 002 ended up being surprisingly coherent, connected by a common thread. The rest currently feel more like scattered nodes in a knowledge graph, unconnected ideas, technologies, and questions. But perhaps through writing and research, new edges will emerge, revealing connections that are not obvious at first. Below is a short archive, a teaser of where the other specimens might lead next. I initially wrote introductions for each, but decided to leave it more suspenseful, just a one-liner and an invitation to uncover the stories behind them when the time comes.

Specimen 003

Grey's Anatomy, but make it science academia edition. An attempt at fiction.

Specimen 004

Building a foundational model of Traditional Chinese Medicine. Can we turn 2,000 years of observations into evidence?

Specimen 005

How do you move millions of people through the sky every day? An ode to aviation engineering and logistics.

Specimen 006

London has a lot of trash. Designing human habitats.

Epilogue

Curiosity is messy. Ideas and thoughts rarely arrive in a straight line. They appear as fragments. This is my attempt to capture the chaos. This is an attempt to preserve the fragments before they disappear.