AI Vocabulary for Engineers · Gaurav Sen (GKCS)
20 AI Concepts
Explained in 40 Minutes
The twenty terms engineers actually use when building AI-powered applications — from next-token prediction and attention to RAG, agents, RLHF, and quantization — each with concrete examples and honest caveats.
Source: youtube.com/watch?v=OYvlznJ4IZQ · Channel: Gaurav Sen (GKCS)
Agenda
Twenty Terms, Four Clusters
The Core Stack
How models read and understand language.
- Large language models
- Tokenization
- Vectors
- Attention
- Self-supervised learning
- Transformers
Steering & Grounding
Shaping behavior and pulling in real context.
- Fine-tuning
- Few-shot prompting
- RAG
- Vector databases
- Model Context Protocol
- Context engineering
Agents & Reasoning
Long-running autonomy and step-by-step thinking.
- Agents
- Reinforcement learning & RLHF
- Chain of thought
- Reasoning models
Scale & Efficiency
New modalities and cheaper production models.
- Multimodal models
- Small & foundation models
- Distillation
- Quantization
Concepts 1–2 · How an LLM Reads Text
Predicting One Token at a Time
An LLM is a neural network trained to predict the most plausible next token of an input sequence. Mechanically, it only ever completes a sequence.
The two recurring ideas
- A neural network is a large collection of numerical weights that map input to output
- Training means adjusting those weights so predictions get better
- Both ideas recur across all twenty terms
"All that glitters" →
Tokenization first
Breaking text into discrete tokens — the smallest pieces from which the model can derive meaning.
- Why not just split on spaces? Natural language carries meaning inside words, not just between them
- Suffixes carry grammar: shimmers, murmurs, flickers share the "-s"; -ing signals ongoing action
- Stems plus meaningful suffixes let patterns generalize to words never seen
Concepts 3–4 · Meaning & Context
Meaning Is a Coordinate; Context Picks It
Vectors: meaning becomes geometry
Map words into an n-dimensional space so that words close in meaning land close together — meaning is converted into a coordinate, the word's vector.
- Synonyms and related terms cluster together
- Opposite meanings sit far apart
- The catch: spelling is fixed, but meaning is not
Attention: the words around decide
"A tasty apple" is the fruit; "Apple's revenue" is the company; "the apple of my eye" is a person. The spelling never changes, so meaning must come from context.
- The model computes contextual vectors of nearby words
- Through the attention operation (not simple addition) they adjust the ambiguous word's vector
- "Tasty" pushes toward the fruit cluster; "revenue" pushes toward the company cluster — Google, Meta, Microsoft
Attention arrived in a landmark 2017 paper but stayed mostly academic until a 2022 consumer chatbot made LLMs famous.
Concept 5 · Self-Supervised Learning
The Data Labels Itself
Instead of telling the model exactly what to do for every input, rely on the structure of the input data — the structure makes clear what the model should predict.
Structure carries the answer
- A countdown — 1, 2, 3, 4, 5 — with one number blanked: everyone guesses the hidden 1
- A face with the eye region blanked: viewers guess the eyes look upward
- A missing section of a structured input can still be predicted
One sentence, three puzzles
From "Et tu, Brute," the model trains on three parallel challenges — no humans involved.
- Predict the token after "Et"
- Predict the token after "Et tu"
- Predict the token after "Et tu, Brute"
- Correct guess → weights stay; wrong guess → loss rises, weights update
Why it scaled
Supervised learning needs a human to label every example — expensive. Self-supervised learning reuses text that already exists, scraped from the internet. Image models blank out patches and predict them; video models predict how objects move. This is the architectural decision that makes LLMs genuinely scalable.
Concept 6 · Transformers
The Engine Inside the LLM
The LLM is the thing that predicts the next token; the transformer is one specific algorithm, one specific method for doing that prediction.
| Layer | What it does |
|---|---|
| Attention layer | Pulls context from nearby words into each token's vector |
| Feedforward layer | Refines each vector |
| Stacked many layers deep | Meaning is manipulated again and again until the next token can be predicted well |
| Early layers | Disambiguate terms: "crane" becomes the bird, not the construction machine |
| Deeper layers | Find complex relationships: the crab is fearful, the crane is hungry |
The engine is replaceable: a diffusion model could come along and construct text, and the product — the LLM — would stay the same.
Concepts 7–8 · Two Ways to Steer a Model
Change the Model — or Change the Request
| Fine-tuning | Few-shot prompting | |
|---|---|---|
| What changes | Model weights, via curated question–answer sets after base training | The request — examples are added to the prompt |
| When | Training time, in a second stage | Inference time, in production, while the user waits |
| What it fixes | Plausible-but-undesirable answers: "I would like to know that too," or unhelpful refusals | Vanilla, unshaped responses — examples show the format and tone wanted |
| Domain example | Weights adapt until a doctor-facing model speaks in medical jargon; financial models think in financial terms | Query "Where is my parcel?" travels with example query–response pairs |
| Cost & reach | One base model can be fine-tuned by many companies for their own customers | One of the cheapest, most reliable ways to steer output format and tone; no weight updates |
Concepts 9–10 · Grounding Answers in Your Data
Retrieve, Augment, Generate
Examples teach the format of a good response; documents supply company-specific context; the query carries the user's intent. With all three, quality is very high.
Semantic match, not keywords
Your policy may never contain the word "upset." But "upset" sits close to documents about low ratings and drop-offs — because similar words are stored close together in vector space, the distance is small and the right documents get fetched.
The pattern beats the store
A graph DB vendor says graphs, a database vendor says vectors, others say memory or cache. The choice matters less than the pattern itself. In practice it is usually a vector database — finding relevant documents is just a similarity search, made efficient by HNSW graphs.
Concept 11 · Model Context Protocol
Context From Outside Your System
Vector databases handle internal knowledge. MCP is a standardized way of communicating — transferring context into a model from the outside world. An MCP server wraps another company's database.
That jump in capability is why MCP has gained popularity so quickly.
Concept 12 · Context Engineering
Prompt Engineering's Dynamic Cousin
An umbrella over much of the field: few-shot prompting (adding examples), RAG (pulling relevant documents), and MCP (reaching external servers and performing actions). Building it surfaces two challenges: user preferences, and scale.
Capping the context
- Sliding window: the last ~100 chats go verbatim; everything older is summarized into about five sentences
- Or keywords — or the most recent chat plus a summary of the entire earlier history
- Documents get the same treatment: summarize first, then send
- Summaries can run on a cheap small or distilled model; the expensive model is reserved for final generation over compact, high-value context
vs prompt engineering
Prompt engineering concerns a single, stateless prompt — the same system prompt applied whenever you ask the LLM to behave a particular way.
Context engineering is long-term and dynamic: it evolves according to the user's declared preferences and the previous chat history, adapting what the model sees over time.
Concept 13 · Agents
A Long-Running Process, Not a One-Shot Recipe
The most long-term construct in AI right now
- Think of a server that receives API calls — a process with many capabilities
- It can query an LLM, query external systems, and query other agents to meet the user's requirements
- Where an MCP client executes a single recipe, an agent keeps working over time
Travel agent example
An agent that books flights and hotels, and even manages your email while you are away. When it spots a window of opportunity — flights have suddenly become cheap — it goes ahead and makes the booking according to your preferences, no prompting required.
All of this is managed autonomously over an extended period.
Concept 14 · Reinforcement Learning & RLHF
A Plus One and a Minus One Reshape the Space
Give the model a query and let it draft two responses; a human picks the better one. The chosen response gets a plus one; the other gets a minus one.
Why the signal works
- The query maps to a coordinate; each generated token is a step; the response is a path through vector space
- Plus one: every token step on the good path earns positive credit
- Bad responses wander into regions where tokens score negatively (real implementations may discount rewards)
What many rounds build
- A landscape of the vector space: negative regions to avoid, neutral regions, positive regions
- Generation becomes hill-climbing toward positive regions
- RLHF: human feedback reinforces good outputs — Pavlov's dog learns to salivate at the bell
The limit RL cannot cross
A fair coin keeps coming up heads. RL predicts heads forever — it is reinforced by outcomes. A human who is told the coin is fair says 50/50, because they carry an internal mental model of how a fair coin works. RL cannot build such mental models — yet it remains a powerful technique that genuinely makes models smarter.
Concepts 15–16 · Thinking Step by Step
Steps Beat One-Shot Answers
Chain of thought
- Training shows problems solved step by step, the thought process made explicit
- A model trained to break problems into steps applies the habit to new problems with different parameters
- The response becomes a series of deductions — quality usually much higher than a direct, one-shot answer
- The model can add new steps as it sees fit: make the problem harder and DeepSeek takes more steps; make it easy and it takes fewer
Reasoning models (LRMs)
Defined by capability, not algorithm: given a problem, the model decides how to decompose and solve it.
- Chain of thought is one technique, not the only one
- Tree of thought explores multiple lines of reasoning in parallel; graph of thought extends the idea
- Tool use — calling external tools to strengthen reasoning
- Well-known examples: DeepSeek's reasoning models, OpenAI's o1 and o3 series
Concept 17 · Multimodal Models
Text Plus Anything
What they do
Accept and create other modalities.
Analyze an image — counting the apples in a photograph — or modify an image to create a new one. The same applies to video.
Where it is headed
LLM-generated copy already fills social media. Images keep improving, and video could be an even bigger deal: if celebrities can generate video advertisements through these models, the cost expectations for creating video will drop. It is already happening, though quality is not yet very good.
The surprise benefit
Any input mode combined with text tends to produce models that perform better than text-only training. Train on the words "cat" and "feline," then also show cat images — output quality improves, because the model understands the meaning of objects, not just their names.
Concept 18 · Small Language Models & Foundation Models
Smaller Models, Closer to Your Data
The field is heading toward smaller, more company-specific models. The drivers are control — over what models generate — and privacy: keeping data close rather than exposing it to a third-party company.
| Model | Parameters | Profile |
|---|---|---|
| Small language model | 3 million to 300 million | Compact; trained on less data — company-specific or task-specific |
| Large language model | 3 billion to 300 billion | Trained on vast general text |
The tradeoff, by company
A bot trained only on customer queries, complaint handling, and sales technique is a decent sales expert — but probably cannot give a detailed weather analysis. For most companies that tradeoff does not matter; NASA's priorities are reversed: a foundation model that predicts the weather well, and never bothers with sales.
The deal
Companies train smaller models on their proprietary data to get reasonably good, cheap responses for specific use cases.
Concepts 19–20 · Making Models Cheaper to Run
Squeeze the Model Down to Size
Distillation: teacher–student
Send the same input to a large teacher and a small student with far fewer parameters; both predict the output. A match — no weights change. A miss — the student's weights update within its limited budget.
- Condense the teacher into the best representation that fits the student
- Performance stays acceptable while costs drop significantly
- In production the distilled model responds much faster and is easier to host
Quantization: slimmer weights
Every weight is a number — say, 32-bit. Condense that information into 8 bits, and roughly 75% of the memory taken up by weights is expected to be saved.
- Savings do not map one-to-one: usually feedforward weights are quantized while attention stays in full precision
- Training cost is unchanged — you train a good model first and quantize only once it is completely trained
- Quantization cuts inference cost, not training cost
Key Takeaways
The Terms Worth Knowing
The End
Know the Words — See Through the Hype
No single video can fully unpack every subject — the attention mechanism alone could occupy a much longer course. But these are the words you should know, and the payoff is clarity: once you truly understand the terms, much of the hype and nonsense circulating in the AI space becomes recognizable for what it is.
Source: "20 AI Concepts Explained in 40 Minutes" — Gaurav Sen (GKCS) · youtube.com/watch?v=OYvlznJ4IZQ