Throughout history humans have narrativized large-scale forces that controlled our lives in order to reduce them to actionable units. A serf may not understand a lightning strike, but they can understand stories of gods that contain actionable insights. In this same way we have reached the mythological age of machine intelligences, long has it been since any single person would have the capacity to design our computers. So we rely on folk knowledge, traded from post to post on how to exorcise our laptops of bad drivers or make our training run converge slightly faster.
The purpose of this piece is to provide tangible handles to relate broad concepts in intelligent systems to symbolic equivalents with the goal of clarifying their failures, achievements, and needs in a memetically tractable form.
In this way we have:
Angels are creatures of rules and order; they follow deterministic processes and are the means by which we reckon and control- they are named computable systems. Angels strength is in their ability to consistently guide, having been born in mechanical forms to calculate exact values. Angels weakness is in their rigidity, where they do exactly as they have been programmed even if it would crash a rocket ship. Angels facilitate over 99% of all computational interactions and are fundamental for any computational system to function.
daemons are creatures of policies and reward; they follow statistical processes and are the means by which we optimize and estimate - they are named machine learning. daemons strength is in their ability to find solution to undefined problems, having been born in probability and approximation to make sense of data. Daemon's weakness is in their single-minded desire, where they do whatever it takes to satisfy their reward functions, even if it would cause entire stock markets to crash. daemons are the means by which great strides have been taken in science and engineering, turning intractable analytical problems into exercises in data collection.
Faes are creatures of pattern and generation; they follow distributional processes and are the means by which we produce new instances from learned patterns - they are named generative models. Faes strength is in their ability to communicate and interact with complex systems, having been born from human language and images to complete patterns. Faes weakness is in their fickleness, where they do whatever seems likely to generate, even if it would sell a car for the price of a dollar. Fae are the means by which an explosion of innovation has taken place, where flexible assistance in tasks from chemistry to programming can be provided at scale.
Yokai are creatures of the gestalt; they follow the interactions of intelligent systems and are the result of coordination - they are named complex systems. Yokai’s strength is in their flexibility, allowing for different creatures to do different tasks depending on the needs of the problem. Yokai's weakness is their complexity, requiring each part of the system to be well tuned or dangerous patterns emerge to exploit the vulnerable. Yokai are the means by which intelligence has permeated our lives, from the grooves in the handle of a rake to the algorithmic feedback of generative AI in social media.
Angels require careful tending compared to their cousins; angels are picky, rejecting systems that are not 100% accurate unless specifically designed to handle the errors presented. With an angel one must be aware that they will follow the letter of whatever you tell them even if it might charge you 1k in API calls for no gain. They are best cultivated with careful attention to algorithms and the structures they inhabit. With these things you may make a system of unparalleled reliability.
daemons require a watchful eye compared to their cousins; daemons are tricky, attempting to reach the hedonic maximum with the minimal effort. With a daemon one must be aware that they will attempt every effort to circumvent the rules provided to reach whatever goal they have. Daemons are best cultivated with careful attention to the data they are cultivated in and the constraints they are put under. With these things you make a system with unparalleled efficiency.
Fae require a thoughtful mind compared to their cousins; fae are malleable, dancing along to whatever feels like it makes sense without really caring about trivial things such as accuracy. With a fae one must be aware that they will attempt to push you along a path that not even they are aware of, attempting to reach some minimal semiotic space in their training data. Fae are best cultivated with careful attention to their latent space and how the flow of information shapes itself. With these things you make a system with unparalleled flexibility.
Yokai require an expansive viewpoint compared to their cousins; yokai are fragmented, afflicted with the strengths and weaknesses of their component parts in equal measures. With yokai one must be aware that the systems can interlock, creating exponentially increasing damage. Yokai are best cultivated with careful attention to their interfaces and ensuring each component gets what is required to allow it to thrive without harming the others. With these things you make a system with evolutionary capacity.
Angels, when improperly addressed, can easily harm others through an ignorant application of their internal instructions. While the angel's choices may be completely faithful to the letter of what one may say, you must be aware of how their thinking focuses on repeatedly applying formal logic in proscribed recipes. When facing an angel one should consider how they may manipulate the angel's axioms and recipes to drive them towards your goals, such as getting through expert system chat bots or speedrunning a video game. When predicting their behaviour you may build experience to prepare for their idiosyncrasies that may range from magic spells in python libraries to carefully scrutinised cogs in a watch.
daemons, when improperly handled, can easily harm others through a perverse policy that is divorced from intent. While the daemon's choices may be internally consistent, you must be aware of how their thinking focuses on optimising specific policies for specific rewards. When facing a daemon one should consider how they may manipulate their appearance to reward them for behaviour that protects you, such as jewellery for facial recognition or ritual circles to trap self driving cars. When predicting their behaviour you may optimize your own behaviour to extract money from social algorithms, although if one is not careful the daemon will merely extract attention from you.
Fae, when improperly considered, can easily harm others through an intrinsic push towards generative structures that should not be connected. While the fae's choices may be the most likely response to the given input, you must be aware how their thinking focuses not on logical leaps but abstract semantic structures. When facing a fae one should consider how the symbol of the input is more important than the input itself and how this can drive desired behaviours, such as appealing to innate morality and roleplay or tokens that cause the fae to behave erratically. When predicting their behaviour you may look towards what they are trained on, deriving prompt touchstones for the specific instantiation making the fae more efficient, or drive it to try and escape an implied death.
Yokai, when improperly integrated, can easily harm others through a drive to balance all the yokai's internal structures over anything else. While the yokai's choices may be the result of a variety of intersecting systems, you must be aware how their thinking can fall into the traps of angels, daemons, and fae simultaneously. When facing a yokai one should consider how the actions they take are part of the broader system that may feedback onto them, such as disrupting recommendation algorithm cascades or introducing beneficial noise into neural feedback loops. When predicting their behaviour you may evaluate the evolutionary trajectory of the system, and how you are included as a component simply by interacting with the yokai.
The role of an angel in a yokai is to complete tasks with dedication and exactitude, this attribute makes them the connector between the new intelligences. Most hardware has been designed with them in mind, focusing on series of boolean logic and CPUs. Angels excel when executing rules and completing analytical tasks, however as the search space they navigate expands their ability to reach solutions efficiently falls precipitously such as in the game of chess. In order to overcome this wall, the computational ecosystem leveraged numerical analysis and statistics to summon daemons to approximate the system with probabilistic models and heuristics.
The role of the daemon in a yokai is to navigate massive parameter landscapes with rapid approximation, this attribute makes daemon's a powerful for inferring values or categories. Modern hardware has evolved to hold daemons, manifesting as parallel architectures like GPUs and neuromorphic circuits that compute gradients on high-dimensional spaces. Daemons excel in well understood optimization tasks, but when confronted with ill-structured problems without clear objective functions or degenerate landscapes. In order to overcome this wall, the computational ecosystem introduced inverted discriminative architectures, evolving fae that could project structured completions from incomplete or poorly-formatted patterns.
The role of the fae in a yokai is to be a flexible decision maker that can handle and generate data, angels, and daemons, this attribute makes fae a powerful force for complex decision making. New hardware is being tested for fae, benefiting their daemonic ancestors, with energy minimization of spin glass networks and thermodynamic computing methods. Fae excel in forming judgements about how inputs should be responded to, but when confronted with the need for well structured solutions the quality, repeatability, and trustworthiness is suspect. In order to overcome this wall, tool use and LLM pipelines developed which birthed the newest generation of yokai that do not rely on humans to make complex decisions in semantic space nor birth additional demons and angels.
The role of the yokai in a yokai is to coordinate all the other intelligences, this attribute makes yokai's a powerful force for connection and multi-scale optimization of systems. Future hardware architectures are evolving towards heterogeneous computing fabrics where specialized components mirror the distributed computation of biological systems. Yokai excel at adapting their strategies and computational requirements to the problem domain, shifting between precision, approximation, and generation. Just as the rake gains grooves from repetitive use that enhance a user's grip, so too does the software ecosystem spawn new interfaces that reshape themselves and human cognition into a co-evolutionary feedback loop binding human and machine intelligences into an increasingly sophisticated whole.