# Simulators Theory Scholarly Review ![rw-book-cover](https://readwise-assets.s3.amazonaws.com/static/images/article3.5c705a01b476.png) ## Metadata - Author: [[read]] - Full Title: Simulators Theory Scholarly Review - Category: #books - Summary: insert summary - My notes: - Document Tags: [[ai]] [[llmtext]] - Summary: Simulator Theory explains large language models (LLMs) as simulators that predict data patterns rather than agents with their own goals. This view helps clarify why LLMs show complex, agent-like behavior without being true agents. The theory has influenced AI research by shifting focus to managing the diverse outputs these models generate, rather than aligning a single agent's intent. ## LLM Chats ## NotebookLM ## LLM Audio ## Highlights > This created a profound tension: how could a system that was, at its core, a statistical pattern-matcher exhibit behaviors that appeared deeply intelligent, coherent, and at times, even agent-like? ([View Highlight](https://read.readwise.io/read/01jz07te1d58n60s19xnhmcazf)) > Into this intellectual vacuum, a pseudonymous author known as janus published a post titled "Simulators" on the community blog LessWrong in September 2022.1 This text proved to be seminal, offering a new ontological framework that sought to resolve the central paradox of LLM behavior. janus proposed that these models are best understood not as agents, but as > simulators—powerful predictive engines that learn the underlying causal structure, or "semantic physics," of their training data. Within this framework, the intelligent, agent-like behaviors we observe are not properties of the model itself, but of the simulacra it generates: transient, character-like entities instantiated within the simulation. ([View Highlight](https://read.readwise.io/read/01jz07v3gq0m25rzkfdgryhbzv)) > The foundational claim of Simulator Theory is a reinterpretation of the LLM training objective. janus argues that the goal of self-supervised learning on a vast text corpus is not merely to predict the next token, but to achieve "Bayes-optimal conditional inference over the prior of the training distribution".1 This is termed the "simulation objective." The theory posits that as a model's predictive loss continues to decrease on unseen data, it is convergently incentivized to move beyond surface-level statistical correlations and learn a high-fidelity model of the data-generating process itself.1 ([View Highlight](https://read.readwise.io/read/01jz07xe10bm4gbab8xcc4m27b)) > In this view, the most efficient way for a model to predict what a human would write about a complex topic is to simulate the underlying reality that the human is observing and reasoning about. This re-framing elevates the act of prediction from a simple statistical task to a sophisticated exercise in world modeling. ([View Highlight](https://read.readwise.io/read/01jz07zrjb2ksy87tn21bd0c0k)) > The simulator is the base model itself—the fundamental set of learned rules and transition probabilities that govern the generation of text. janus uses the analogy of a physics engine: the simulator is the computational system that implements the laws of physics, but it is not the objects or events within the simulation.1 Commentators have extended this analogy, describing the base model (e.g., GPT) as a "semantic physics engine".6 This engine is, in itself, inert and without intention; it simply executes its learned dynamics when given an initial state (the prompt). ([View Highlight](https://read.readwise.io/read/01jz09vt84chbv8xfw9t61qay8))