SGI Research Lab · Reference

MUST KNOW
AI TERMS

12 foundational concepts that separate fluent AI practitioners from people who are guessing. The vocabulary you need to move fast in an AI-native world.

12 Core Concepts
TERM 01
01 · Core Architecture
LLM
Large Language Model
  • The AI brain behind ChatGPT, Claude, Gemini, Copilot.
  • Trained on billions of documents.
  • Predicts the most likely next word — at scale.
TERM 02
?
02 · Failure Mode
HALLUCINATION
Confident Wrong Output
  • When AI makes stuff up.
  • Confidently. Completely wrong.
  • Happens when training data has gaps.
TERM 03
unbelievable
03 · Unit of Compute
TOKEN
Chunk of a Word
  • The building block of AI communication.
  • Not a word — a chunk of a word.
  • More tokens = more cost in enterprise settings.
TERM 04
04 · Two Phases
TRAINING
VS
INFERENCE
Teaching vs. Doing the Work
  • Training = teaching the model.
  • Inference = the model doing the work.
  • You can't have the second without the first.
TERM 05
05 · Specialization
FINE TUNING
Specialized Training
  • Taking a general model.
  • Feeding it specialized data.
  • Making it smarter at one specific thing.
TERM 06
+R
06 · Learning Loop
REINFORCEMENT
Reinforcement Learning
  • AI learns by trial, reward, and repeat.
  • Like training a dog — but the dog is a neural network.
  • This is how reasoning models get sharper.
TERM 07
07 · Compression
DISTILLATION
Large → Small Model
  • Teaching a small model to mimic a large one.
  • Faster. Cheaper. Nearly as good.
  • Likely how GPT-4 Turbo was built.
TERM 08
08 · Architecture Pattern
RAG
Retrieval-Augmented Generation
  • AI + your own documents.
  • Retrieves relevant context before responding.
  • Reduces hallucinations dramatically.
TERM 09
123
09 · Reasoning Mode
CHAIN OF
THOUGHT
Step-by-Step Reasoning
  • Breaking a complex problem into steps.
  • Like showing your work in math class.
  • Slower — but far more accurate.
TERM 10
10 · Model Parameter
WEIGHTS
Learned Intelligence as Math
  • Numbers that shape what the model "knows."
  • Start random. Adjust during training.
  • The model's actual learned intelligence — stored as math.
TERM 11
11 · Training Signal
VALIDATION
LOSS
Training Progress Score
  • A score that tells you how well training is going.
  • Lower = better.
  • Helps catch overfitting before it's too late.
TERM 12
> write()> test()
12 · Agentic AI
CODING
AGENT
Autonomous Code Execution
  • An AI that doesn't just suggest code.
  • It writes, tests, and debugs — autonomously.
  • Think: a tireless intern who never loses focus.