# Large Language Model A **large language model** is a type of [[Foundation Model]] trained primarily on text data to predict, generate, and reason about language. LLMs are the technology behind ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google), Llama (Meta), and the AI assistants transforming information work, content creation, programming, and — increasingly — military intelligence and government operations. ## How They Work (Simplified) An LLM is trained by feeding it enormous quantities of text (books, websites, code, scientific papers — trillions of words) and asking it to predict the next word, over and over, billions of times. Through this process, the model learns not just grammar and facts but patterns of reasoning, argumentation, and even deception. Key components: - **[[Transformer Architecture]]:** The neural network design (invented by Google researchers in 2017) that makes LLMs possible. Transformers process text in parallel rather than sequentially, enabling training at scale. - **Parameters:** The model's learned weights — the numbers that encode everything it has learned. GPT-4 reportedly has ~1.8 trillion parameters; Claude's count is undisclosed. More parameters generally means more capability, but at exponentially increasing cost. - **Context window:** How much text the model can "see" at once. Early models had 4,000-token windows (~3,000 words); current frontier models handle 100,000–1,000,000+ tokens — enabling them to process entire books, codebases, or intelligence reports in a single interaction. - **[[RLHF]] (Reinforcement Learning from Human Feedback):** The fine-tuning process that turns a raw text predictor into a helpful, harmless assistant. Human raters evaluate model outputs; the model is trained to produce outputs humans prefer. ## The Power Dynamics LLMs matter for this KB not because of how they work but because of what they concentrate: **Information power.** An LLM trained on the internet's text has internalized humanity's collective written knowledge — and the biases, ideologies, and power structures embedded in that knowledge. Whoever controls the model's training data, fine-tuning, and deployment controls how that knowledge is filtered, presented, and applied. **Economic power.** LLMs are replacing or augmenting human labor across writing, coding, legal analysis, medical diagnosis, customer service, and content moderation — at a pace that outstrips any previous automation technology. The economic value created flows primarily to the companies that build and deploy the models, not to the workers whose output trained them. **Military-intelligence power.** LLMs are the analytical engine inside [[Palantir AIP]], [[Palantir Gotham]], and the classified AI systems the Pentagon deployed starting in 2024. An LLM can process thousands of intelligence reports in seconds, identify patterns across surveillance data, draft targeting packages, and generate propaganda in any language — all applications documented in the defense-tech sections of this KB. ## KB Significance 1. **The substrate of AI power.** Every AI company, defense-tech product, and governance debate documented in this KB ultimately rests on LLM technology. Understanding what an LLM is and what it can do is prerequisite to understanding the Palantir-Pentagon pipeline, the Anthropic blacklisting, the DOGE automation project, and the AI arms race with China. 2. **The labor displacement engine.** LLMs are the first technology capable of automating cognitive labor at scale — writing, analysis, coding, translation, legal research. The economic displacement this creates is the material basis for the political disruptions documented across the KB. 3. **The propaganda machine.** LLMs can generate human-quality text in any language, at any volume, instantaneously. The [[Pentagon Influence Operations - gc_ Generation]], Russian bot farms, and political micro-targeting operations documented in the KB are all accelerated by LLM capability. ## Why the vault treats this as one page Merged 2026-08-13 from a duplicate `Large Language Model (LLM)` page. The parenthetical had nothing to disambiguate from, and prose naturally links the bare term, so the bare title is canonical and the parenthetical is now a redirect. Two threads carried over from the merged page: LLMs are the technology that brought AI into mass public use from 2022 and are the centre of gravity of the current boom; and they are the object of four simultaneous fights — [[Disinformation]] at scale, political "bias" and content moderation (shaped via [[RLHF]]), copyright over training data, and the [[AI Arms Race]]. See [[Natural Language Processing]] and [[Generative AI]] for the surrounding technical context, and [[Frontier Model]] for the capability-tier distinction. ## See Also - [[Foundation Model]] — The broader category - [[Transformer Architecture]] — The technical innovation that made LLMs possible - [[RLHF]] — How raw models become assistants - [[Scaling Laws]] — Why LLMs keep getting bigger - [[AI Safety]] — The effort to make LLMs safe - [[Dual-Use AI]] — The military application problem - [[Pravda Network]] — Russian "LLM grooming": flooding the web so chatbots repeat Kremlin narratives