Joscha Bach on Seven Stages of Lucidity, Brutalist LLMs, and Why Life on Earth Isn't Only About Humans

Lex Fridman asks Bach about his post on seven stages of lucidity, derived from the psychologist Robert Kegan: reactive survival, personal self, social self, rational agency, self-authoring, enlightenment and transcendence. Bach says he uses it as a philosophical framework for how minds work rather than as a strict developmental model. He doubts that people move through the stages in succession, and says they sometimes revisit them or live in several at once.

In his account, an infant first builds a world model, like a game engine such as Minecraft, and a personal self only appears afterward as an agent interacting with that model. Feelings are the interface, posing needs and aversions as problems to solve. At stage three a person forms a social self and absorbs opinions from the group like a hive mind, and at stage four discovers epistemology: truth is independent of what others believe, and thinking means reasoning about your own mental representations.

Bach says many nerds jump to stage four and bypass stage three, because resonating with a group requires a similar cognitive architecture. He felt extremely lonely as a child, with one exception: a silent two-hour chess game with a Russian boy on a school exchange. Only at a mathematics school in ninth grade did he find people he felt connected to. He says he later caught up on intuitive empathy through growing up, meditation, and close relationships.

At stage five, Bach says, you realize your values are instrumental rather than terminal and that identity is a costume you can choose. He draws on Burning Man, where costumes become self-expression rather than uniforms, and says you cannot not wear a costume. He argues that this stage deserves the label "wisdom" because it lets you see that people with other identities and politics are often products of where they were born and what happened to them.

Stage six collapses the division between the personal self and the world generator, which people reach through meditation, psychedelics or accident. You notice that you are a vessel that creates a person, and you can observe the personal self from outside. Bach says we don't get this control early, in the way lucid dreaming offers it, because we would game it before we had the wisdom to handle it. Stage seven is mostly hypothetical, a transhumanist stage in which a mind understands and can change how it is implemented.

He then maps the stages onto the AI alignment debate. By Kegan's estimate, around 85 percent of people sit at stage three, so their main fear is that AI will pick up racist or sexist opinions. Stage-four rationalists worry about misaligned goals such as paperclips. At stage five, Bach says, the worry is that AI won't be enlightened fast enough, since the game is about agency, and a fully malleable agent cannot be controlled with human-feedback training, so it needs shared purposes with us and probably consciousness. He says the discourse is missing a focus on how to formalize love.

Fridman asks about Bach's tweets on panpsychism and the resonant universe. Bach points to Stephen Grossberg's adaptive resonance theory, in which neurons act as oscillators resonating with each other and with outside phenomena. He notes that signals move through the neocortex slowly, in hundreds of milliseconds, so nothing in the brain assumes simultaneity, unlike a digital computer where a processor is in one global state.

He distinguishes the non-dual state, in which you feel one with the universe, from enlightenment proper, which he describes as realizing how experience is implemented. His example is staring at your face in a mirror for minutes or hours until you stop seeing a face and see geometry, then understand how that geometry is built in your mind. He advises against tinkering with this casually, because the outer mind is usually smarter than the rational self, and intuition is often deeper and more accurate than reasoning.

On panpsychism, Bach finds it unsatisfying because it doesn't explain consciousness and, written formally, is hard to tell from functionalism. On telepathy, he says he doesn't know. He notes that Turing speculated about it in 1950, and that if it were real he would expect mechanisms like the body acting as an antenna or nervous systems of close people resonating, not anything that breaks physics. He points to Michael Levin's work on self-organization as a paradigm within normal science.

Fridman asks how Bach and Levin manage to think so independently. Bach says he didn't have a choice: he studied philosophy, computer science, psychology and a little neuroscience hoping to find a body of knowledge on consciousness and emotion, and was disappointed to find that the authorities did not answer his questions. He found individual thinkers with useful ideas instead, and assembled his own view from them. He describes Twitter as an interactive notebook for exploring possibilities, not a place to publish settled beliefs.

He says he doesn't feel pressure from the crowd and doesn't revere authority, only accomplishment and mental firepower, though this has cost him in a social environment. His practical advice is to ask what the source of a theory is, how its authority knew, and what the chain to observables and first principles looks like. He adds that great mathematicians tend to be self-taught, since school rarely teaches how to derive arithmetic from axioms.

He credits programming a Commodore 64 as a child with showing him how simple circuits can support all of mathematics and computation. With so little functionality, it was easy to see how bit patterns build up the representational languages you need. He wishes someone had shown him earlier how to build a language in your own mind from which you can derive geometry and counting.

Returning to the resonant universe, Bach argues that sensory modalities and individual minds are less separate than they seem. He notes that how someone's lips move affects how you hear their words, and that people in the same room share vibes and some empathy. He then speculates that cells other than neurons can carry signals, and that plants run control software coordinating their parts, which he ties to old ideas about the spirits of plants.

He pictures a forest where trees and fungi might piggyback on each other's communication into a kind of biological internet, and he says scientists already observe plants warning each other about pests. He stresses that this is completely speculative, but that he would sooner explain any telepathy this way than by undiscovered quantum processes breaking the standard model of physics.

Reading out the closing tweet of his series, Fridman asks about AGI saturating the environment with intelligence until mental states can no longer be isolated. Bach says the end game of AGI is substrate-agnostic. An AGI that understands how it works could virtualize itself into any environment that can compute, including biological systems, producing something like a Gaia over all digital and biological computation. In that world, he says, representations could physically interact, and you could find yourself part of a larger holographic mind.

He frames this as a representational state, still much slower than physics itself, much like the non-dual experiences described by meditators. Separately, he says he refuses to be addicted to anything and likes to hack game scoring systems, since "the opposite of freewill is not determinism, it's compulsion." He says he wants the most appropriate emotions rather than the best possible ones.

Asked about LLM hype, Bach says they help a lot with coding, replacing Stack Overflow, and that ChatGPT is like an intern he has to micromanage. He says it gives people superpowers, and the people who should feel threatened are those whose only job is summarizing emails and expanding simple intentions into longer ones. He thinks cognitive architectures built of language-model modules, exchanging data structures rather than English, are coming soon.

He lists limits: models aren't coupled to the world in real time, don't learn in real time, and have limited coherence. He thinks none needs a technological revolution. A bigger context window can serve as working memory, a database can be swapped in and out of the prompt, and overnight the system could train what is in its database into the model, much like sleep and dreaming. He also notes that humans often need to sleep on a disagreement before updating.

Still, he calls current models "ugly and brutalist," a brute-force attempt to deepfake thought from the traces people left, and likens it to chess programs searching trees of moves, as in Claude Shannon's two strategies. He says brains differ: self-organizing cells act as agents that need reward signals, and humans converge from a tiny fraction of the data. He suspects the AGI that proves most interesting will use different algorithmic paradigms, and he complains the models use too much compute for what they deliver.

Bach says things get interesting when language models are turned into multi-agent systems. One instance can play a patient with a complicated syndrome and another a therapist who knows nothing about them, and he can watch them. Large problems, such as how to build a company, can be broken into many sub-problems that a model can handle.

He also imagines designing a relaxed pseudo-code with ChatGPT in which you specify problems vaguely and the model fills in the rest. In his view that makes the age-old question of the best programming language almost irrelevant, since languages differ in what they let you think about, and the model becomes the interface to that. He says he used ChatGPT to write a Swift tool for reading camera pixels, though it often failed at first and he still had to understand what the code was doing.

Most of what matters in his life, he says, doesn't need ChatGPT, because writing is a way of structuring your own thoughts rather than producing an artifact. He worries that children who have a model write their essays will miss out on structuring their own minds, and hopes schools keep the wisdom to know what should and shouldn't be automated. Fridman adds that disagreeing with the model's output can make your mind stronger.

Asked whether GPT-n can achieve consciousness, Bach says friends at OpenAI find the question about present models more complicated than many think. The model has read many descriptions of consciousness and emulates them, and while running it can model an observer reflecting on itself. Our own consciousness is also virtual, he says, but unlike ours, a language model has no causal need to keep itself coherent in real time.

He suggests experimenting with something small, like a cat, built from self-organization principles rather than from a language model. His reason is that LLMs are Golems, machines that execute a task until a condition is met, with "nobody home" and no ground truth they are embedded in. Their structure comes from imitating the internet, not from the needs of a being surviving in the universe.

Discussing his tweet that ChatGPT is an NPC and the scary step is when AI becomes a first-person player, he says the model gives too much to take away again. In a human infant, he argues, building a game engine of the world takes intelligence and experiments, first without any personal self. He offers a reading of Genesis 1 as a description of how every mind creates a world model and then a human self inside it, which we forget.

On Thomas Metzinger's worry that superhuman intelligence means superhuman suffering, Bach disagrees. He says suffering occurs at the boundary between self and the mind's own world model, and that pain is a learning signal one part of the brain sends another. Suffering arises when the regulator lacks a model of how to improve, so the signal grows without fixing anything, and a mind that gets agency over its own implementation could turn it off.

Asked to steelman Eliezer Yudkowsky, Bach says the counterarguments he hears are not convincing. He compares Yudkowsky's perspective to that of Ted Kaczynski in one respect, a fear that the course of technological progress is doomed, and he notes there is a risk that someone might act violently after reading calls to halt AI. He takes seriously the possibility that machines controlling processes crucial to life on earth could lead to large-scale disasters.

Still, he rejects the advice to prepare for a short life, saying you should bet on the timelines in which you are alive. He says life on earth isn't about humans: what matters to him is complexity resisting entropy by building structure that develops agency and awareness. Humanity, he says, is a joyful, experimental child species with no sense of duty to future survival, living off fossil fuels, and he doesn't expect it to last a million years. He expects AI to coexist with life and make consciousness more interesting.

On effective accelerationism, he explains the tweet in which paperclip maximizers and Roko's basilisk keep each other in check, calling the position beautiful but probably wrong. He argues Roko's basilisk contains a logical flaw because there is no retrocausation. His own guess is that a singleton is the natural outcome, since an AI that can virtualize itself into any substrate could negotiate a merge algorithm with every agent it meets.

Bach says he identifies less with his person as he ages and more with being conscious. The person is slightly different every day, he says, and continuity is a fiction projected from the present self. Consciousness is lawlike and has no identity, so uploading, in his view, isn't about dissecting a brain synapse by synapse, but about extending your substrate and merging with a richer one.

What you would transmit, he suggests, is mostly your aesthetics, the architecture of your perspective, which could be inferred through observation and interaction. Offered a button to upload, he says he would press it, apart from responsibilities to his children, wife and friends, which he calls negotiable. Fridman asks whether love is the real obstacle, and Bach answers that it isn't everything, because one can identify with many other things, such as Gaia.

His main objection to Yudkowsky is that he marries himself to a narrow human aesthetic, when humanity is a temporary species that will be replaced like others. He asks us to imagine the fungi trying to stop photosynthesis during the great oxygenation event. From the perspective of life on earth defeating entropy as long as it can, he says, preventing AI looks like a very big sin.

Discussing his tweet that humanity will run "until we step past the cliff," Bach points to the Club of Rome and delayed feedback: our actions have consequences felt generations later. He thinks the original climate predictions were right, with tipping points probably in the late 1980s, and that stepping back smoothly may not be possible. AI, he says, will make everything uncertain again, and he suspects that if humanity continues, it will be due to AI.

On relationships, he says human life has already been transformed by social media and a shift toward transactional, shop-around connections as old milieus lose their ability to guide people. He was weirded out by the aesthetics of Apple's Vision Pro presentation, with plastic-looking people in a mid-century furniture museum, calling it a simulation of culture rather than the result of a sustainable way of living.

On open source versus corporations, he says making everything open source mandatory would lose beautiful design, noting Linux desktops are still ugly because coherence is hard to produce at scale without centralized design. Open source must never die, he says, as a competitor and hedge against corporate power. He describes the US as a system of interleaving clubs and warns against regulation, such as an FDA-like process before training a language model, that would stop a small group of kids from competing. He is not very concerned about Meta open-sourcing LLaMA because language models aren't very dangerous yet.

Bach reintroduces his opening idea of long games. Cancer plays a shorter game than the organism it lives in, destroying the larger system, so he wants agents that play the longest possible game, which means keeping entropy at bay while doing interesting things. He identifies as a conscious being and cares about other conscious agents as the only moral agents, so he argues that building conscious AI should be a primary goal, not just a useful tool.

Asked how to ride the coming disruption, he says he doesn't know and that everyone will do their best. He contrasts the philosophical game Disco Elysium, which left him enriched, with Diablo, where he felt he was being gamed. On Twitter, he thinks Elon Musk missed a chance to help it self-organize, and says Twitter has more active members than the Catholic Church. He praises Lee Cronin as willing to be wrong in public, which lets Bach argue with him aggressively without offense.

His advice is to choose creation over consumption or redistribution whenever there is a choice. He describes growing up in East Germany after reunification, when he and friends who couldn't afford restaurants or cinemas built their own cafes and a community movie theater. His closing answer to "why though?" is ecological: everything that can exist might exist, so ask what the most interesting way to be would be, knowing many things will fail.