Dante Peppermint Glossary: AI Thinking and Brand Strategy Terms

Not a chatbot, not an assistant. A system built to think alongside you, not just respond to you. Distinguished by the presence of a genuine point of view.

What you get when an AI is trained on the entire internet. Statistically correct, creatively inert. The enemy of original work.

Where your brand sits in relation to alternatives, in the mind of the people you're trying to reach. Not what you say about yourself — what they understand without being told.

The act of giving context, constraints, and intent to an AI before asking it to think. Quality of briefing determines quality of output. Most people skip this step.

The specific body of work, voices, and thinking that a model draws from. A corpus is curated, not scraped. It determines the character of the output.

The act of making deliberate choices about what something should feel like and why. Not execution — intention.

What makes something impossible to substitute. Not features. The reason a specific person would choose this over everything else.

AI that produces content based on patterns in training data. Useful for volume. Not the same as thinking.

Work that could not have been produced by averaging. Not novel for novelty's sake — specific in a way that reflects a genuine perspective.

In the context of AI, a genuine perspective encoded through specific source material — not averaged consensus. The difference between a tool that tells you what sounds right and one that tells you what it actually thinks.

Stress-testing an idea, strategy, or brand position against objections, alternatives, and edge cases. What thinking partners are for.

The space between convention and irrelevance. Where the work that actually resonates comes from. Not the center.

A structured approach to reasoning through a problem, rather than producing an answer. Built on architecture, not just parameters.

The specific character of how something communicates. Not tone (which can shift). Voice is structural — it's the same across contexts.

A neural network trained on massive text datasets to predict and generate language. The architecture powering most modern AI writing tools. "Large" refers to parameter count — typically billions. Size correlates with capability, up to a point.

The field in which systems improve performance on a task through exposure to data rather than explicit programming. The foundation beneath most modern AI. It's not magic — it's optimization over a loss function.

The discipline of enabling machines to understand, interpret, and generate human language. Modern LLMs have largely superseded classical NLP methods, but the term still describes the problem space.

A computational system loosely modeled on biological neurons. Layers of nodes transform inputs into outputs through learned weights. The substrate for most modern AI, including language models and image generators.

The dataset used to teach a model. What goes in shapes what comes out. Biases in training data become biases in model behavior. Understanding a model's training data is essential to understanding its outputs.

The learned numerical weights inside a neural network. GPT-4 has an estimated ~1 trillion parameters. More parameters generally means more capacity — not necessarily more intelligence.

Running a trained model to generate output. Training teaches the model; inference is when you use it. Most AI usage in production is inference, not training.

The neural network architecture introduced in the 2017 "Attention Is All You Need" paper. The foundation for virtually every major language model. The key innovation: attention mechanisms that weigh the relevance of different tokens to each other, enabling coherent long-range reasoning.

The mechanism inside transformer models that determines which parts of an input to focus on when generating each output token. It's how the model knows that "bank" in "river bank" means something different than in "savings bank."

A numerical vector that represents a word, sentence, or concept in high-dimensional space. Semantically similar things end up close together. The geometric representation that lets models reason about meaning mathematically.

Adapting a pre-trained model to a specific task or domain by continuing training on a smaller, curated dataset. Cheaper than training from scratch. Used to make general models behave more appropriately in specialized contexts.

A technique where a model is given access to an external knowledge base at query time, retrieving relevant documents before generating a response. Keeps the model grounded in specific, current, or proprietary information without retraining.

The practice of crafting inputs to an AI model to improve output quality, relevance, or behavior. Part craft, part engineering, part psychology. As models improve, the delta between mediocre and excellent prompts stays significant.

The maximum amount of text a model can consider in a single interaction, measured in tokens. Everything outside the window is invisible to the model. Modern frontier models support hundreds of thousands of tokens.

The fundamental unit of text a language model processes. Not quite words — "running" might be one token; "unimaginable" might be two. Models have hard limits on total tokens in their context window.

A parameter controlling randomness in model outputs. Low temperature = more deterministic, predictable responses. High temperature = more varied, creative — and potentially incoherent. Most production tools default around 0.7–1.0.

A class of generative models that learn to reverse a noise-adding process. The architecture behind most modern image generators (Stable Diffusion, DALL-E, Midjourney). Starts with random noise, iteratively refines toward a coherent image.

A model that can process and generate across multiple input/output types — text, images, audio, video. GPT-4V and Gemini are examples. The direction most frontier models are moving.

An AI system that can take actions autonomously — browsing the web, writing and executing code, calling APIs — rather than just responding to a single prompt. Agents operate in loops: observe, plan, act, repeat.

OpenAI's consumer AI interface, launched November 2022. The product that made large language models a mainstream phenomenon. Built on GPT models. Not a model itself — a product wrapping one.

OpenAI's line of large language models. The "pre-trained" part matters: the model is trained on broad internet data first, then fine-tuned. GPT-4 is as of 2024 one of the most capable models available.

Anthropic's large language model, designed with a focus on safety, interpretability, and helpfulness. The model powering Dante Peppermint. Known for strong long-form reasoning and following nuanced instructions.

Google DeepMind's multimodal large language model. Integrated across Google's product suite. Competes directly with GPT-4 and Claude at the frontier model tier.

An open-source diffusion model for image generation. One of the most widely used image AI tools, notable for running locally without cloud dependency.

AI models whose weights and/or training code are publicly released. Enables transparency, local deployment, and community modification. Llama (Meta), Mistral, and Stable Diffusion are notable examples. Contrasted with proprietary closed models.

When a model generates confident, plausible-sounding content that is factually wrong. The model doesn't "know" it's wrong — it's pattern-matching toward a probable next token, not verifying against ground truth. A fundamental limitation of current LLMs.

The challenge of ensuring AI systems reliably pursue goals that are genuinely beneficial to humans. Not just making AI safe by restriction — but ensuring its values and objectives stay coherent with human values as capability scales.

The study and practice of ensuring AI development and deployment is fair, accountable, transparent, and non-harmful. Covers bias in training data, surveillance applications, labor displacement, environmental costs, and access equity.

Systematic skew in model outputs resulting from imbalanced or unrepresentative training data, or from human labelers encoding their own assumptions. Manifests as demographic disparities, stereotyping, and uneven performance across populations.

An attack where malicious content in a model's input overrides its intended instructions. Example: a webpage with hidden text instructing a browsing AI agent to exfiltrate the user's data. An active security concern in agent-based AI systems.

A hypothetical AI system capable of performing any intellectual task that a human can, at human level or above. Not yet achieved. The definition is contested. The timeline is vigorously debated. The implications are significant.

Constraints and policies built into AI systems to prevent harmful outputs. Can be hard-coded (refusal logic) or soft (fine-tuning toward cautious behavior). Always a tradeoff against capability.

A database designed to store and efficiently search high-dimensional embeddings. The infrastructure layer that makes RAG work at scale. Pinecone, Weaviate, and Chroma are common examples.

The time between a prompt submission and the first token of a response. In AI products, latency is a meaningful quality signal. Streaming responses mask latency — you see tokens appear as they're generated, not all at once.

The absence of genuine point of view is the defining failure mode of most AI tools. A model trained to maximize agreement produces content that hedges, qualifies, and centers the middle. Dante was built on a different premise: that AI can be configured to have — and hold — a perspective.

The consistent, identifiable way an organization communicates — its tone, cadence, diction, and what it refuses to say. Not a style guide. A personality. The difference between "we help businesses grow" and "growth is not the strategy. It's the result."

Where a brand sits in the mind of a customer relative to alternatives. Not a tagline. A structural claim about distinctiveness. Good positioning defines what you are by defining what you are not.

In the context of Dante Peppermint, the six curated bodies of thinking that define the system's perspective. Not a training dataset — a philosophical architecture. The model didn't consume these voices to mimic them; it uses them as a lens.

A structured document that defines the objective, audience, constraints, tone, and non-negotiables for a creative project. The quality of a brief determines the quality of the output. Garbage in, garbage out — even with a powerful AI.

Searching by meaning rather than exact keyword match. Powered by embeddings. Ask "what does the CEO think about AI regulation?" and a semantic search will surface relevant content even if those exact words never appear together.