The Signal Glossary: AI Terms You Need to Know

Not Wikipedia definitions. What each term means and why it matters to you.

AI moves fast enough that the vocabulary shifts under your feet. This glossary is a living reference not academic definitions, but practitioner-grade explanations of the terms you'll encounter in Signal coverage and in the field. Bookmark it.

Foundation Model

A large AI model trained on broad data to serve as a general-purpose base. Think of it as the foundation of a building it doesn't do anything specific on its own, but everything useful gets built on top of it. GPT-4, Claude, Llama, Gemini: all foundation models. The term matters because the economics of AI increasingly flow through a small number of these base models, making the companies that build them extraordinarily powerful.

Frontier Model

The most capable AI model available at any given moment. "Frontier" isn't a fixed label it's a moving target. Today's frontier model is tomorrow's baseline. What matters: frontier models define the ceiling of what's possible, and the gap between frontier and everything else is where competitive advantage lives.

AGI (Artificial General Intelligence)

AI that can perform any intellectual task a human can. Nobody agrees on the exact definition, which is part of the problem. Some labs claim we're months away; others say decades. The practical question isn't "when is AGI?" it's "at what point does AI capability cross the threshold where your industry gets restructured?" For most industries, that threshold arrives well before anything deserving the label AGI.

ASI (Artificial Superintelligence)

Intelligence that surpasses the best human minds across every domain. If AGI is "as smart as a human," ASI is "smarter than all humans combined." This is where the existential risk conversations live. It's also where the most transformative economic potential lives. The two are not separable.

Superintelligence

Used interchangeably with ASI, but sometimes refers more broadly to any system whose cognitive capabilities exceed humanity's collective capacity in a given domain. The word carries weight because it forces a question: what do humans do when they're no longer the smartest entity in the room?

Inference

The act of running a trained model to get outputs the "thinking" part. Every time you send a prompt to ChatGPT or Claude, that's inference. It matters because inference is where the cost is. Training a model is a one-time expense (albeit enormous); inference is the ongoing meter running every time anyone uses it. The economics of AI are increasingly inference economics.

Training

The process of building a model by feeding it data and adjusting its parameters. Training a frontier model costs hundreds of millions of dollars, requires thousands of GPUs running for months, and consumes enough electricity to power a small city. This is why only a handful of organizations can do it, and why the "who trains the models" question is a geopolitical one.

Fine-Tuning

Taking a pre-trained model and training it further on specialized data. If training builds the foundation, fine-tuning adds the specialization. A medical fine-tune makes the model better at diagnosis. A legal fine-tune makes it better at contract review. Fine-tuning is how general models become useful for specific work and it's dramatically cheaper than training from scratch.

RLHF (Reinforcement Learning from Human Feedback)

A technique where humans rate model outputs and the model learns from those ratings. This is how raw language models become helpful language models the difference between a system that can write anything and one that writes what you actually want. RLHF is also where alignment happens (or doesn't). The humans doing the rating are making value judgments that get baked into the model's behavior.

Agent

An AI system that can take actions, not just generate text. An agent doesn't just tell you what to do it does it. Books a flight. Writes and runs code. Sends an email. The shift from "AI as oracle" to "AI as actor" is the single biggest capability jump in the current generation.

Autonomous Agent

An agent that operates without human approval for each step. It receives a goal, decomposes it into subtasks, executes them, handles errors, and reports back. The "autonomous" part is what makes this transformative and terrifying in equal measure it means humans are no longer in the loop for every decision.

Multi-Agent Orchestration

Multiple AI agents working together, coordinated by a lead agent or framework. One agent researches, another writes code, a third reviews it, a fourth deploys it. Opus 4.6 orchestrating sub-agent teams is a real-world example. This is how AI scales beyond the limits of any single model's context or capability.

Agentic AI

The broad category encompassing agents, autonomous agents, and multi-agent systems. "Agentic" has become the industry's favorite adjective because it captures the shift from passive to active AI. When someone says "agentic AI," they mean AI that does things rather than says things.

Context Window

The amount of text (measured in tokens) a model can process at once its working memory. A small context window means the model forgets the beginning of your conversation by the end. A large one means it can digest entire codebases or legal documents in a single pass. Context window size is a hard constraint on what an AI can reason about, and it's been expanding rapidly.

Tokens

The units models use to process text. Roughly, one token equals about three-quarters of a word in English. Tokens matter because they're the unit of cost (you pay per token for inference) and the unit of capacity (context windows are measured in tokens). When someone says a model has a "200K context window," they mean 200,000 tokens roughly 150,000 words, or a decent-length novel.

Prompt Engineering

The practice of crafting inputs to get better outputs from AI models. Part art, part science, and increasingly part anachronism. As models get smarter, the need for elaborate prompting decreases. But for now, the difference between a mediocre prompt and a great one can be the difference between useless output and genuine insight.

Self-Improving AI

An AI system that can enhance its own capabilities. This is the concept at the heart of "the wall" when AI gets good enough to make itself better, the improvement rate decouples from human research speed. We're in the early stages: AI writes code that improves AI training pipelines, which produces better AI that writes better code.

Recursive Self-Improvement

Self-improving AI taken to its logical conclusion: each improvement enables a faster or better improvement, creating a feedback loop. This is the theoretical mechanism behind an intelligence explosion. Whether it proceeds smoothly or hits diminishing returns is one of the most important open questions in the field.

Intelligence Explosion

The hypothetical point where recursive self-improvement accelerates beyond human ability to track or control. Coined by I.J. Good in 1965: "the first ultraintelligent machine is the last invention that man need ever make." Whether this is utopia or catastrophe depends entirely on alignment whether the exploding intelligence shares human values.

The Wall / Vertical Innovation Curve

Signal's term for the inflection point where AI progress goes vertical. See our series, "The Vertical Wall." The wall isn't a barrier it's the moment the exponential curve turns steep enough to feel like a cliff. Everything before the wall is preparation. Everything after it is adaptation.

Model Layer

The foundation models themselves GPT, Claude, Gemini, Llama. The model layer is where raw intelligence lives. Historically, the model layer and the application layer were distinct: someone built a model, someone else built a product on top of it.

Application Layer

The products and services built on top of models ChatGPT, Cursor, Copilot, custom enterprise tools. The application layer is where models meet users.

The Collapse Between Layers

What's happening right now: model providers are building applications, and application builders are training models. The distinction between "model company" and "app company" is dissolving. This collapse matters because it determines who captures value. If the model layer eats the application layer, a handful of AI labs own everything.

Compute

The raw computational power required to train and run AI models. Compute is the oil of the AI era whoever controls the most of it has the most leverage. The global competition for compute is driving billions in data center investment and reshaping energy policy.

GPU (Graphics Processing Unit)

The hardware that powers modern AI training and inference. Originally designed for rendering video game graphics, GPUs turned out to be perfectly suited for the parallel math that neural networks require. NVIDIA's dominance in AI GPUs has made it one of the most valuable companies on Earth.

TPU (Tensor Processing Unit)

Google's custom AI chip, designed specifically for machine learning workloads. TPUs are Google's answer to NVIDIA's GPU dominance purpose-built silicon that trades general-purpose flexibility for AI-specific performance. The GPU vs. TPU vs. custom silicon competition is one of the key hardware battlegrounds.

Data Center

The physical infrastructure where AI models are trained and served. Modern AI data centers consume as much power as small cities and require specialized cooling, networking, and power systems. The geography of data centers where they're built, who builds them, what powers them is becoming a national security issue.

Open-Weight

A model whose trained parameters (weights) are publicly released, allowing anyone to run it. Llama and Mistral are open-weight. "Open-weight" means you can use the model, fine-tune it, and deploy it yourself. It does not necessarily mean you can see the training data, the training code, or the full methodology. This distinction matters.

Open-Source

In AI, truly open-source means the weights, the training data, the training code, and the methodology are all public. Almost nothing marketed as "open-source AI" meets this standard. Most are open-weight at best. The difference matters because open-weight gives you a tool; open-source gives you the ability to build the tool yourself. The industry's loose use of "open-source" is a deliberate blurring that serves corporate interests.

Hallucination

When a model generates confident, fluent, and completely false information. Hallucinations aren't bugs in the traditional sense they're a structural feature of how language models work (predicting plausible next tokens, not verified facts). Hallucination rates are declining but not zero, which means every AI output still requires human verification for high-stakes decisions.

Alignment

The challenge of ensuring AI systems do what humans actually want. Easy to state, staggeringly hard to solve. A misaligned superintelligence isn't necessarily malicious it might simply optimize for the wrong objective with superhuman effectiveness. Alignment research is, arguably, the most important work happening in AI today.

Guardrails

Technical and policy constraints placed on AI systems to prevent harmful outputs. Content filters, safety training, usage policies. Guardrails are the practical, imperfect implementation of alignment the fence you build while you figure out how to make the horse not want to run away.

Red-Teaming

Deliberately trying to break, bypass, or misuse an AI system to find vulnerabilities before bad actors do. Borrowed from military and cybersecurity practice. Red-teaming is how labs stress-test their guardrails and discover failure modes. The quality of a model's safety is only as good as the creativity of its red team.

SWE-Bench

A benchmark that tests AI models on real-world software engineering tasks fixing actual bugs in actual open-source repositories. SWE-Bench matters because it measures something practitioners care about: can this model do real work? Not toy problems, not multiple choice, but actual engineering. Scores have been climbing fast.

Coding Time Horizon

How far into the future an AI model can reliably plan and execute a software task. A model with a 4-hour coding time horizon can handle tasks that would take a human engineer 4 hours. At 1.04x improvement per model generation, this metric is compounding and it's one of the clearest signals of approach to the wall.

API (Application Programming Interface)

The interface that lets software talk to an AI model programmatically. APIs are how AI gets embedded in everything every app, every workflow, every product that uses AI does so through an API. When we talk about "the API economy" in AI, we mean the massive ecosystem of products built on top of model APIs.

ARR (Annual Recurring Revenue)

The annualized revenue from subscriptions. In AI, ARR growth rates are the market's primary signal for which companies are winning. OpenAI's ARR trajectory, Anthropic's enterprise deals, startup burn rates ARR is the scoreboard everyone's watching.

MoE (Mixture of Experts)

An architecture where a model contains multiple specialized sub-networks ("experts") and a router that activates only the relevant ones for each input. MoE models can be enormous in total parameters but efficient in practice because only a fraction activates per query. This is how models get smarter without proportionally more compute and it's a key architectural bet for the next generation of frontier models.

This glossary is a living document. As the field evolves, so will these definitions. Suggest additions at thesignal.press.

📡 More signal, less noise → www.thesignal.press