Terms Covered
50
Five clusters — basics, architecture, usage, infra, evaluation
Most Cited on Earnings Calls
Infra
HBM, inference, clusters dominate
Most Misunderstood Term
Parameters
More is not automatically better
How to Use This
Ctrl+F
Built for lookup, not for reading end to end
50 AI Terms
In the order you'll meet them in news and earnings calls
📌 How to use this
This is built for lookup, not linear reading — use Ctrl+F (⌘+F). Each entry has two parts: a one-line definition and where you'll actually encounter it.
Five clusters: ① basics ② architecture ③ usage ④ infrastructure ⑤ evaluation and safety.
This is built for lookup, not linear reading — use Ctrl+F (⌘+F). Each entry has two parts: a one-line definition and where you'll actually encounter it.
Five clusters: ① basics ② architecture ③ usage ④ infrastructure ⑤ evaluation and safety.
① Basics (10)
| Term | One-line definition | Where you'll meet it |
|---|---|---|
| AI | Any technology letting machines do judgment, perception, or generation tasks humans did | The broadest umbrella. Most headlines |
| Machine learning | Finding patterns in data instead of hand-coding rules | A subfield of AI. Recommenders, anomaly detection |
| Deep learning | Machine learning with many-layered neural networks | What the post-2012 AI boom actually is |
| Neural network | A computation structure that multiplies and adds numbers layer by layer | The unit of deep learning |
| LLM | A large network trained on vast text to predict the next token | The engine inside ChatGPT, Claude, Gemini |
| Generative AI | AI that produces new content rather than classifying or predicting | The axis of market growth since 2023 |
| Token | The smallest text unit a model handles — smaller than a word | The unit of API pricing and context limits |
| Parameters | The internal numbers tuned during training. "70B" = 70 billion | Model size notation. Bigger isn't always better |
| Training | Building a model by adjusting parameters from data | The source of bulk GPU demand |
| Inference | Generating answers with a finished model. Parameters don't change | Most of the ongoing cost of running a service |
⚠️ The "more parameters is better" myth
Parameter count is capacity, not performance. Data quality, training method, and post-training alignment routinely let smaller models beat larger ones. Recently the trend has clearly shifted toward smaller, faster models for production work. Don't rank models by parameter count alone.
Parameter count is capacity, not performance. Data quality, training method, and post-training alignment routinely let smaller models beat larger ones. Recently the trend has clearly shifted toward smaller, faster models for production work. Don't rank models by parameter count alone.
② Architecture (10)
| Term | One-line definition | Where you'll meet it |
|---|---|---|
| Transformer | The attention-based network architecture that is today's LLM standard | Introduced 2017; the base of every current LLM |
| Attention | Computing how much each token should reference every other token | Why long inputs cost more |
| Embedding | A token or sentence converted into meaning-bearing numeric coordinates | Underpins search, recommendation, RAG |
| Context window | The maximum tokens a model can consider at once — read more | Marketing lines like "1M token context" |
| Pre-training | Stage-one training that builds general language ability from bulk text | By far the most expensive stage |
| Fine-tuning | Further training a finished model for a specific purpose | Enterprise custom models |
| RLHF | Post-training that shapes answer style using human preference ratings | Why answers come out "helpful and safe" |
| Alignment | The broad work of matching model behaviour to human intent and values | Central term in safety discussions |
| MoE | Splitting a large model into experts and activating only a few per query | Big parameter counts with cheap inference |
| Quantization | Lowering parameter precision to shrink size and cost | On-device and low-cost inference — the 14GB→4GB walkthrough |
③ Usage and applications (12)
| Term | One-line definition | Where you'll meet it |
|---|---|---|
| Prompt | The input text you give the model | The starting point of all chat AI use |
| Prompt engineering | Designing input to get the output you want | See our seven principles |
| System prompt | Standing instructions applied across the whole conversation | The hidden config defining a service's character |
| Zero-shot | Giving instructions with no examples | The default way people use it |
| Few-shot | Supplying a handful of examples along with the task | A practical accuracy multiplier — applied to voice matching |
| CoT (chain of thought) | Having the model write intermediate reasoning to raise accuracy — built into reasoning models via training | The "think step by step" prompt |
| RAG | Retrieving documents before answering and grounding on them — read more | The standard architecture for internal-doc chatbots |
| Vector DB | A database that stores embeddings and finds similar ones fast | The core component of RAG |
| Agent | AI that calls tools and executes multiple steps on its own | The hottest application area of 2025–26 — read more |
| Function calling (tool use) | Letting the model invoke external programs and APIs directly | The foundation agents are built on |
| Multimodal | Handling images, audio, and video alongside text | Photo upload and voice conversation features — see how it works |
| On-device AI | Running the model on the device instead of a server | Smartphone and PC launch marketing — how it actually works |
④ Infrastructure and hardware (10)
📈 The most practical cluster for investors
These ten recur constantly in semiconductor and cloud earnings calls. Knowing them makes conference-call summaries far easier to read accurately. See our semiconductor sector analysis and chip competition deep dive.
These ten recur constantly in semiconductor and cloud earnings calls. Knowing them makes conference-call summaries far easier to read accurately. See our semiconductor sector analysis and chip competition deep dive.
| Term | One-line definition | Where you'll meet it |
|---|---|---|
| GPU | A chip that runs the same computation massively in parallel; the standard for AI training | Central to Nvidia and AMD results |
| HBM | Stacked memory attached to a GPU to feed it data at very high speed | A key swing factor for SK hynix and Samsung |
| Bandwidth | How much data can move per unit time | The real bottleneck in AI performance |
| TPU / NPU | Chips purpose-designed for AI computation | Google, Apple, smartphone SoCs — vs. CPU/GPU |
| Interconnect | The high-speed links joining GPUs to each other | Make-or-break for large training clusters |
| Cluster | Thousands to tens of thousands of GPUs assembled for training | "100,000-GPU cluster" announcements |
| Data center | The physical facility with servers, cooling, and power | What AI capex actually buys |
| CAPEX | Capital expenditure on facilities and equipment | The intensity gauge for Big Tech AI spending |
| Inference cost | What it costs to serve one user query | The core of AI service profitability |
| Foundry | Contract manufacturing of chips designed by others | TSMC, Samsung Foundry |
⑤ Evaluation and safety (8)
| Term | One-line definition | Where you'll meet it |
|---|---|---|
| Hallucination | Generating false content as though it were fact | See our guide to handling it |
| Benchmark | A score from running the model against a standard problem set | The bar charts in every model launch deck |
| Data contamination | Test questions leaking into training data, inflating scores | Why benchmark numbers deserve suspicion |
| Knowledge cutoff | The last date covered by the model's training data | The limit on questions about recent events |
| Jailbreak | Circumventing safety measures to extract prohibited output | Security and safety discussions |
| Prompt injection | Hijacking an AI via instructions hidden in external documents — read more | The signature security risk for agents and RAG |
| Guardrails | Safeguards that block harmful or inappropriate output | A prerequisite for enterprise adoption |
| Red teaming | Deliberately probing a model for weaknesses before release | Pre-launch safety review |
💡 Three filters for reading AI news
① Coverage that leads with "N billion parameters" is thin on performance evidence — look for benchmarks and real-use evaluation too.
② Because of data contamination, independent evaluation matters more than the absolute benchmark number.
③ Whether an earnings call separates "training demand" from "inference demand" tells you where that company sits in the cycle.
① Coverage that leads with "N billion parameters" is thin on performance evidence — look for benchmarks and real-use evaluation too.
② Because of data contamination, independent evaluation matters more than the absolute benchmark number.
③ Whether an earnings call separates "training demand" from "inference demand" tells you where that company sits in the cycle.
※ This glossary reflects usage as of July 2026. Terminology in AI shifts quickly — check primary sources for current definitions. Companies named here are illustrative for the definitions and this is not investment advice.
※ This guide is provided for general educational purposes and simplifies technical details for readability.
New guides, when they land
We publish AI literacy guides twice a week. Subscribe and the next one comes to you — free, unsubscribe anytime.