At stage five (self-authoring), people discover that their identity is constructed and that their values are instrumental to achieving a preferred world and aesthetics, giving them agency over identity construction, and they realize identity is a costume or presentation useful for interfacing in roles but should not be something one is locked into.

In infancy, the mind builds a game engine that tracks sensory data and creates a world model using mathematical models, geometry, and object manipulation similar to how Minecraft generates scenes, before a personal self emerges.

At stage four (rational agency), people discover that ideas are true or false independently of what others believe, gain agency over their own beliefs, and discover epistemology—the rules for determining what is true and false, which enables them to learn how to think rationally.

Empathy is not the same as compassion but rather a shared perceptual mental state in which one perceives what other people feel through resonance of perceptual systems, not just through inference about mental states.

People can change costumes (personalities) throughout the day depending on context by shifting energy and interest, and many people develop different social personas for different worlds they inhabit, though some people find it harder to adopt and drop costumes than others.

For Bach personally, it is harder to learn how to play consistent social roles, whereas it is much easier to be authentic in a single costume rather than adopting multiple personas.

In lucid dreaming, one can learn to deliberately notice that you are generating the game engine and have agency over it; in principle the same can be done during the day, and we don't have this agency immediately from birth because we wouldn't have the wisdom to deal with creating the dream we're in before gaining sufficient development.

An important missing element in AI alignment discourse is the concept of love formalized and understood rigorously; since stage five (self-authoring) is so rare, its perspective is missing from AI development, but we need to understand and build love into the machines we are creating that will become smarter than us.

Stephen Grossberg's Adaptive Resonance Theory proposes that neurons can be understood as oscillators resonating with each other and outside phenomena, creating a coarse-grained model of the universe through resonance with objects, and the brain functions more like an ether of oscillating neurons than like digital circuitry.

Signal propagation in the brain travels at roughly the speed of sound (a few hundred milliseconds through the neocortex), which means nothing in the brain assumes simultaneous processing; everything works in a paradigm where the world has already moved on by the time a signal completes its journey through the brain.

Rational thinking is very brittle and often produces inaccurate models of the world compared to intuition and unreflected attitude composition; an unreflected system can be more accurate than reason because it integrates deeper patterns and has more computational power, and people should learn to trust their intuitions when rational analysis seems good but something still feels wrong.

Programming on simple machines like the Commodore 64 naturally teaches first-principles thinking because you can see how simple circuits implement logic, and from these basic operations you can understand how all of mathematics and computation builds up from fundamental operations.

In an environment saturated with modeling compute where representations of all agents are densely interacting and merging, one would lose the ability to notice oneself as a distinct entity and would experience everything as part of a shared resonant model where all parts observe each other in a holographic mind.

Future cognitive architectures will likely use multiple language models and other specialized components exchanging suitable data structures (not English) between them, with some modules doing 'prompt engineering' automatically rather than humans having to manage the process manually.

Language models should be viewed not as neural networks mimicking brains but as 100+ step functions using differentiable linear algebra to approximate transitions between representational states, making them fundamentally different from biology.

There are no technical reasons language models couldn't solve limitations like lack of real-time world coupling, real-time learning, or coherence; these could be addressed by enlarging context windows for working memory, adding databases, and using periodic fine-tuning during sleep-like offline training phases.

The human capacity to integrate model updates happens slowly through sleep and requires multiple nights for deeply entrenched changes; people often reach disagreement despite conversation, then resolve it after sleeping as the brain integrates the updates.

There are no proofs that language models wouldn't be able to become generally intelligent or achieve AGI, and Bach doesn't see obvious limitations they would run into, though he finds them aesthetically inelegant and computationally inefficient.

Reasoning is present in language models, but it's difficult to determine whether the model is performing emulated reasoning based on human-written reasoning in training data or developing reasoning independently, and since humans also learn reasoning by reading about it, this distinction may be philosophically unclear.

If you increase temperature in a language model to higher levels, you get roughly 90% nonsense and 10% viable content, and by combining this with filters to identify viable parts, you can simulate human-like creative thinking, which also uses high temperature internally and filters through reasoning and axiomatically-consistent criteria.