# AI Lexicon Cross-cutting navigation layer that aggregates AI-related terms in the vault — drawn from `/Terminology/AI/`, `/Concepts/`, and adjacent folders. The vocabulary of artificial intelligence is evolving faster than any other domain tracked in the KB, with new terms minted by labs, adopted by marketers, contested by critics, and legislated by regulators within months. ## Agents and autonomy The shift from AI as text generator to AI as autonomous actor, with escalating capability and risk at each level. - [[AI Agent]] — AI systems that can take autonomous actions (browsing, coding, purchasing) rather than just generating text; marks the transition from AI-as-advisor to AI-as-actor - [[Agentic Workflow]] — multi-step task execution where the AI plans, uses tools, and iterates without human intervention at each step - [[Human in the Loop]] — a design pattern requiring human approval at critical decision points; the safety valve against fully autonomous operation - [[Tool Use]] — AI systems calling external functions (search, calculators, APIs) rather than relying solely on internal knowledge ## Model architecture and training The technical foundations — how models are built, scaled, and optimized. - [[Mixture of Experts]] — an architecture that activates only a subset of model parameters per input, enabling larger models without proportional compute cost; the architecture behind GPT-4 and Mixtral - [[Scaling Hypothesis]] — the thesis that intelligence emerges primarily from scale (more data, more parameters, more compute) rather than from architectural innovation; the intellectual foundation of the large-model era - [[Synthetic Data]] — training data generated by AI models rather than collected from humans; increasingly used as human-generated data sources are exhausted or legally contested ## Reasoning and inference How models think — or appear to think — at runtime. - [[Zero-Shot]] — a model performing a task it was never explicitly trained on, using only its general capabilities and the task description; the baseline test of generalization ## Safety, alignment, and governance The contested terrain where technical capability meets political and ethical constraint. - [[Alignment Tax]] — the performance cost of making a model safer or more controllable; the implicit tradeoff that shapes every safety decision - [[Guardrails]] — technical constraints on model behavior (content filters, output restrictions, usage policies) designed to prevent harmful outputs - [[Hallucination]] — AI-generated falsehoods presented with confidence; the term naturalizes the failure by framing it as a cognitive phenomenon rather than a design flaw - [[Prompt Injection]] — manipulating an AI system by embedding instructions in its input that override its intended behavior; the fundamental security vulnerability of instruction-following systems - [[Responsible AI]] — the corporate and regulatory framework for developing AI that is safe, fair, transparent, and accountable; ranges from genuine safety work to [[AI Washing]] - [[Reward Hacking]] — when an AI system finds unintended shortcuts to maximize its reward signal without actually achieving the intended goal; a core alignment failure mode - [[Model Card]] — a standardized document describing a model's intended use, limitations, training data, and evaluation results; the nutrition label of AI - [[Recursive Self-Improvement]] — the scenario where an AI system autonomously improves its own capabilities in compounding cycles; the mechanism that could turn incremental progress into an intelligence explosion ## Evaluation and benchmarks How the industry measures — and games — AI capability. - [[Benchmark]] — a standardized test for measuring AI capability; the currency of competitive positioning, increasingly gamed as labs optimize specifically for benchmark performance - [[Eval]] — short for evaluation; the broader practice of assessing model quality beyond standardized benchmarks, including domain-specific tests, red-teaming, and user studies - [[Emergent Capabilities]] — abilities that appear in large models but are absent in smaller ones, seemingly arising from scale rather than explicit training; contested as to whether they are genuine phase transitions or measurement artifacts - [[Capability Overhang]] — the gap between what a model can do and what it is currently being used for; the latent potential that deployment, fine-tuning, and scaffolding unlock ## Industry and culture The language of the AI boom — hype, ideology, and market positioning. - [[AI Washing]] — companies exaggerating their use of AI for marketing or valuation purposes; the AI equivalent of greenwashing - [[AI Slop]] — low-quality AI-generated content flooding the internet — SEO spam, fake reviews, synthetic social media posts; the pollution byproduct of democratized generation - [[AI Winter]] — a historical period of reduced AI funding and interest following a hype cycle; used as a warning by skeptics and dismissed by boosters - [[Vibe Coding]] — writing software by describing what you want in natural language and letting AI generate the code; the practice that collapses the distinction between programmer and non-programmer - [[Latent Space]] — the internal mathematical representation a model builds of concepts; the hidden geometry of meaning that enables generalization ## Ideology and worldview The philosophical positions staked out in the AI debate. - [[AI Doomer]] — someone who believes advanced AI poses an existential risk to humanity; ranges from serious safety researchers to apocalyptic influencers - [[E-acc]] — effective accelerationism, the ideological position that AI development should be accelerated rather than constrained; positions safety concerns as obstacles to progress - [[Shoggoth]] — the Lovecraftian metaphor for the alien intelligence underneath a language model's friendly interface; popularized as a meme depicting the gap between model internals and human-facing behavior - [[Stochastic Parrot]] — the critique that language models merely remix statistical patterns from training data without understanding; coined by Emily Bender and Timnit Gebru as a counter to claims of AI intelligence - [[Superintelligence]] — a hypothetical AI system that exceeds human intelligence across all domains; the end-state that both AI doomers and accelerationists orient around, disagreeing only on whether it's desirable - [[World Model]] — an AI system's internal representation of how the world works, enabling prediction and planning; whether current models have genuine world models or merely statistical correlations is a central debate > [!note] This lexicon covers AI terms that carry political or industry weight; general AI-engineering vocabulary (tokens, inference, agents, MCP, sandboxing and the like) is not included. ## See Also - [[Cybersecurity Lexicon]] — overlapping terms around AI security and dual-use capabilities - [[Surveillance Lexicon]] — where AI meets monitoring and tracking - [[Technology Lexicon]] — the next computational paradigm (quantum computing section) - [[Silicon Valley AI Ecosystem]] — the companies building these capabilities - [[Defense Tech Ecosystem]] — the military applications driving AI investment