**Links to**: [[Coming abstractions]], [[Invention]], [[Science]], [[Semantic attractor]], [[Politics]], [[Narrative]], [[Theory-fiction]], [[Hyperstition]], etc. ### [[Postulate]]: Fiction provides all of the learning, with (almost) none of the partaking risk. **Abstract**: This paper examines how, through the lens of active inference, fictions can be understood as semantic attractors enabling—relatively—risk-free learning. Fiction allows agents to creatively explore counterfactual scenarios and minimize uncertainty without exposure to “real-world” consequences. This mechanism frames both the evolutionary advantage and persistent appeal of narrative engagement. We treat aspects of punctuation and breath, following the work of Margaret Masterman. We also touch upon aspects of more “atomic” semantic attractors in the structures of language, such as rhyme and alliteration, as patterns which aid memory, communication and information retrieval policies. <small>Keywords: Fiction, Active inference, Fictional inference, Semantic attractor, Risk, Learning.</small> ### Introduction Fictional constructs can be understood as enabling explorative cognitive situations where communicating groups (or individual agents holding an internal, although always dialogical, thought process) can learn by simulating responses to high-stakes scenarios while maintaining corporeal integrity/safety, all the while plastically exploring their epistemic condition. Unlike “direct” experience, where prediction errors carry immediate consequences that need to be dealt with there and then, **fiction**—imagine: monsters, sci-fi, and plenty of other speculations—provides a neither-real-nor-impossible speculative space for testing predictive models against hypothetical outcomes. This speculative environment enables learning, in a way that often expresses itself as shared through narratives. The linguistic, transferrable nature of narrative structure therefore also allows for internal “parsing” and “epistemic foraging” at the level of the single agent. Moreover, the “grand narrative” that all fictions come to an end, that they can be understood as _chunks_, provides an additional factor which minimizes uncertainty about an agent’s engagement with a particular cultural product. On the contrary, in the context of experimental artistic practices such as performances where spectators are not informed of an event’s structure (such as when its ending is supposed to take place), we observe situations which lead to higher discomfort and even intolerance (see, e.g., the famous “Riot at the Rite”, people leaving the theater during the premiere of the Rite of Spring by Stravinsky, 1913). Additionally, as traditional kinds of fictions (movies, novels, jokes, etc.) and structures (exposition, climax, resolution) become more commonplace and well-known, these evolve to incorporate more uncertainty: leading to more cultural learning possibilities (respectively, with regard to the previous examples: expanded cinema and VR, experimental novels and publishing practices, ever-novel joke structures, etc.). By drawing on these kinds of examples, we will analyze how narrative structures act as “semantic attractors” by organizing *possible* world-states into explorable patterns. These patterns can be understood to function as probability-possibility landscapes for cultural belief-updating, allowing collectives to refine their predictive models through (repeated, rehearsed, newly imagined, etc.) engagement with fictional scenarios. Considering homo sapiens’ universal tendency towards narrative-sharing, this process of “fictional inference” may be fundamental to human learning capabilities, particularly in domains where direct experience would be prohibitively risky or impossible. After all, the only place where plans can begin, is in speculation about their effects. This analysis extends to all manner of media environments (such as VR), where possible worlds and (sometimes) interactive narratives provide increasingly sophisticated platforms for this type of risk-free learning. The proliferation of fictional modes across human cultures can be understood as the *ever more diverse* exploration of the learning-risk trade-off inherent in active inference. The productive capacity to engage with fictional scenarios represents one of the most interesting evolutionary solutions to the exploration-exploitation dilemma faced by all complex learning systems. The conclusion reached is that fictional inference’s capacity to facilitate learning in relative “safety” may represent one of the most interesting functions or adaptations in human cognitive evolution: from imagining-projecting gods and spirits, to the idea of a “Zeitgeist”, or even collectivity under a nation state (a rigid fiction), to our current understandings of whatever it is we call “AI”. The mechanisms of fictional inference therefore demand particular attention in the context of our futurological prospects. Please get in touch for a copy: [[ｃｏｎｔａｃｔ]]. %% there are several notable examples of audience walkouts and protests at experimental cultural events throughout history. Here are some of the most famous: 1. The Riot at the Rite (1913) - The premiere of Igor Stravinsky's "The Rite of Spring" in Paris caused such an uproar that it devolved into a riot. The avant-garde music and choreography were so shocking to audiences that they began shouting, arguing, and some people walked out. The performance famously continued despite the chaos in the audience. 2. Minimalist composer Steve Reich's "Four Organs" (1973) - During a performance at Carnegie Hall, the experimental piece caused such a negative reaction that audience members began shouting, mocking the music, and walking out. One woman reportedly banged her head on the front of the stage yelling "stop, stop, I confess!" 3. Antonin Artaud's "The Cenci" (1935) - His experimental theater production in Paris lasted only 17 performances, with audiences regularly walking out due to its violent themes and unconventional staging. 4. John Cage's "4'33"" (1952) - While not exactly a walkout, the premiere of this famous "silent piece" caused significant controversy, with many audience members becoming angry or leaving, not understanding that the silence was intentional. 5. "Miss Julie" by August Strindberg (1889) - The naturalistic play shocked audiences with its frank treatment of class and sexuality, causing walkouts and protests at early performances. 6. The premiere of Alfred Jarry's "Ubu Roi" (1896) - The first word of the play was "merdre" (a modified version of the French word for "shit"), causing immediate outrage and walkouts. The show caused a riot and only ran for one night. 7. Andy Warhol's "Empire" (1964) - This experimental film consisting of eight hours of continuous footage of the Empire State Building caused numerous walkouts during its rare full screenings. 8. Sarah Kane's "Blasted" (1995) - The premiere at London's Royal Court Theatre caused numerous walkouts and sparked intense critical controversy due to its extreme violent and sexual content. These incidents often marked important shifts in artistic expression, with works that were initially rejected later being recognized as groundbreaking or influential. Framed around ActInf: The relationship between experimental art and audience rejection provides a fascinating lens through which to examine active inference and risk-free learning through fiction. The notable cases of Stravinsky's "Rite of Spring," Cage's "4'33"," Jarry's "Ubu Roi," and Artaud's "The Cenci" represent instances where the semantic attractors presented were so radically different from established patterns that they triggered what might be termed "prediction error cascades" in their audiences. Consider Stravinsky's "Rite of Spring" (1913): The audience riot occurred precisely at the point where the work violated multiple predictive models simultaneously. The dissonant harmonies, irregular rhythms, and primitive choreography created a massive prediction error that the audience's existing cognitive models couldn't readily absorb. While the performance offered a "safe space" for experiencing these violations, the magnitude of the prediction errors overwhelmed the usual fictional container that allows for risk-free learning. The audience's physical reaction—shouting, fighting, leaving—suggests that the boundary between "safe" fictional space and genuine threat became momentarily blurred. John Cage's "4'33"" (1952) presents an even more intriguing case. By removing all traditional musical content, Cage forced audiences to confront a complete violation of the concert-going semantic attractor. The piece created what might be called a "null space" where the expected patterns of musical performance were entirely absent. The audience's discomfort and anger reveal how deeply ingrained these predictive models are—the very absence of expected content became a kind of "anti-attractor" that forced confrontation with fundamental assumptions about what constitutes music and performance. Jarry's "Ubu Roi" (1896) deliberately attacked the semantic attractors of polite society and theatrical convention. The opening profanity ("merdre") served as an immediate signal that all predictive models about theatrical propriety would be violated. The subsequent chaos in the audience demonstrates how fictional spaces, when pushed to certain extremes, can rupture their own containment mechanisms. The single-night run suggests that the social body's immune response to radical prediction errors can be swift and absolute. Artaud's "The Cenci" (1935) represents perhaps the most theoretically self-aware attempt to manipulate these dynamics. Artaud's Theater of Cruelty explicitly sought to shatter conventional semantic attractors, creating what he called a "plague" that would infect audiences with new modes of perception. The production's failure with audiences, despite (or because of) its theoretical sophistication, suggests there may be natural limits to how quickly predictive models can be reformed, even in ostensibly safe fictional spaces. These historical examples suggest that while fiction generally provides a safe space for exploring new predictive models, there exists a threshold beyond which the magnitude of prediction errors overwhelms the protective frame of "fiction." This threshold appears to be related to both the number of simultaneous prediction errors and their fundamental nature. When art violates too many core predictive models simultaneously, audiences may experience what we might term "cognitive vertigo"—a state where the usual mechanisms for incorporating new information through fiction break down. The paradox is that these "failed" experiments often became crucial semantic attractors for future audiences, suggesting that the social processing of radical prediction errors requires time and multiple exposures. What was initially rejected as "impossible" gradually becomes integrated into the cultural repertoire of possible patterns, expanding the range of what subsequent audiences can process without rejection. This pattern suggests that fiction's role in risk-free learning operates not just at the individual level but as part of a larger cultural process of predictive model evolution. These historical cases thus provide valuable insight into both the mechanisms and limits of fiction as a space for safe prediction error processing, suggesting that while fiction generally provides an ideal environment for updating predictive models, there exist thresholds beyond which this protective function can temporarily fail. Ah, I apologize for the misinterpretation. Let me explore how the original abstract's thesis about fiction as risk-free learning environments and semantic attractors could inform novel approaches to chain-of-thought architectures in LLMs: The concept of fiction as a "safe space" for exploring self-information minimization suggests an intriguing parallel for how we might restructure chain-of-thought processes in language models. Rather than pursuing purely logical or deterministic reasoning chains, we might consider implementing what could be called "narrative reasoning layers" – intermediate processing spaces where models can explore multiple potential reasoning paths simultaneously without committing to them, similar to how humans use fictional scenarios to test different possibilities. The idea of semantic attractors in fiction could inspire a new approach to prompt engineering and model architecture. Instead of treating chain-of-thought as a linear process, we might design systems that maintain multiple parallel "possibility spaces" – each functioning as a semantic attractor that pulls different aspects of the reasoning process into coherent patterns. This would allow models to explore multiple potential solution paths simultaneously while minimizing the risk of getting stuck in local optima. The abstract's emphasis on risk-free learning suggests we might benefit from implementing what could be termed "hypothetical reasoning buffers" in language models. These would be specialized layers or modules that allow the model to explore speculative chains of reasoning without immediately affecting its core prediction mechanisms. Like fiction for humans, these buffers would provide a protected space for testing novel logical connections and inferential leaps. This framework could be particularly valuable for handling edge cases and novel problems. Just as humans use fiction to explore scenarios they've never directly experienced, language models could use these narrative reasoning layers to approach problems that fall outside their training distribution. The model would treat such cases not as direct logical problems to solve, but as "fictional scenarios" to explore through multiple potential reasoning paths. The concept of self-information minimization through fiction suggests a possible new training objective for chain-of-thought systems. Instead of optimizing solely for accuracy or likelihood, models could be trained to maintain and explore multiple coherent "story-worlds" of reasoning, gradually converging on solutions through a process more akin to narrative resolution than logical deduction. This might help models develop more robust and flexible reasoning capabilities, particularly for complex or ambiguous problems. There's also an interesting implication for how we might handle uncertainty in language models. Just as fiction allows humans to maintain multiple potentially contradictory scenarios simultaneously, we might design chain-of-thought architectures that explicitly maintain multiple competing reasoning paths without forcing early resolution. This could help prevent the kind of premature convergence that often leads to reasoning failures in current systems. These insights suggest a potential shift from current chain-of-thought implementations that often mimic formal logical reasoning toward architectures that more closely resemble human narrative thinking. Such systems might be better equipped to handle ambiguity, uncertainty, and novel situations, much as human engagement with fiction builds capacity for handling complex real-world scenarios.