The concepts behind the tools

Every conversation you have with an AI model happens inside what's called a context window — think of it as the AI's working memory, or the "table" on which it can spread out all the material it's currently reading and thinking about. When a model has a 200,000 token context window, that's how much information it can actively hold in mind at once before it starts "forgetting" earlier parts of the conversation.

Here's the crucial thing: AI models don't have persistent memory like a human brain. Once a conversation ends — or once you exceed the context limit — the model has no access to anything that fell outside that window. It's as if every session starts with a fresh blank slate, and the context window is the only desk space available.

The number "tokens" isn't the same as words or characters. A token is roughly ¾ of an English word — so 1 million tokens is approximately 750,000 words, or about 1,500 pages of text. The leap from 8,000 tokens (where most models started) to 1,000,000 tokens (where frontier models are now) is genuinely transformative, not incremental.

An AI benchmark is a standardised test — a set of questions, tasks, or challenges with known correct answers — used to objectively compare models. Instead of saying "Claude feels smarter than GPT," a benchmark lets you say "Claude scores 82% on SWE-Bench Verified vs GPT-5's 77%." They give the industry a shared measuring stick.

The problem is that benchmarks attract Goodhart's Law: once a measure becomes a target, it ceases to be a good measure. Companies train models specifically to perform well on popular benchmarks — which is why a model can score brilliantly on a test but disappoint you in real-world use.

The word "token" comes up constantly in AI — in context window sizes, in pricing, in model descriptions. A token is the basic unit of text that a model processes. It's not a word and it's not a character — it's a fragment of text, roughly corresponding to 3–4 characters or about ¾ of an average English word.

Common English words are usually 1 token. Longer or unusual words get split into multiple tokens. Numbers, punctuation, and code have their own tokenisation patterns. The model "sees" text as a sequence of tokens, not a sequence of words.

API pricing has two components: input tokens (everything you send to the model — your prompt, any documents, the conversation history) and output tokens (what the model writes back). Output is almost always priced higher than input because generating text requires more computation than reading it.

A hallucination is when an AI model generates text that is plausible-sounding but factually incorrect — and presents it without any indication that it might be wrong. The term is borrowed loosely from psychology, but the AI version is distinct: the model isn't confused or dreaming; it's doing exactly what it was designed to do, just with imperfect underlying information.

Language models work by predicting the most statistically likely next token given everything that came before. They're extraordinarily good at this — but "most likely" and "most true" are not the same thing. A model trained on vast amounts of human text learns patterns of how information is typically expressed, not a perfect factual database.

1. Use RAG (Retrieval-Augmented Generation). When you give the model the actual document to reference, it grounds its answers in real text rather than reconstructed memory. This is why NotebookLM and Claude with uploaded documents are far less likely to hallucinate than a bare chat prompt.

2. Ask for citations and check them. Request that the model cite its sources — and verify the key ones. Tools like Perplexity do this automatically with live web links you can click.

3. Calibrate your confidence. Use AI for first drafts and exploration; use authoritative sources for anything that goes into a published document, legal brief, or medical decision.

In traditional software, open source means the code is publicly available for anyone to read, modify, and run. In AI, open source means the model's weights — the billions of numerical parameters that define what the model "knows" — are publicly downloadable. Closed source means the model runs only on the company's servers, accessible only via their API or products.

The implications are enormous. An open-source model can be downloaded, run on your own hardware, modified for specific tasks, deployed without ongoing fees, and studied by researchers. A closed-source model gives you none of those abilities — but typically represents the absolute frontier of capability, since the company is incentivised to keep its best work proprietary.

The DeepSeek moment illustrates why this matters so much. When DeepSeek released their R1 model in January 2025 as open-source — at a level of capability approaching GPT-4o — it proved that frontier-level AI could be built at a fraction of the cost of Western labs, and then given away freely. This single event erased enormous perceived moats and raised fundamental questions about whether closed-source AI companies could maintain their pricing power long-term.

A neural network is made up of layers of mathematical operations. Each operation involves numbers called parameters (also called weights) — these are the millions or billions of numerical values that the model adjusts during training until it gets good at predicting text. The total count of these numbers is the model's "parameter count."

A model with 7 billion parameters has 7,000,000,000 individual numerical values that were optimised during training. A model with 700 billion parameters has 100 times more. More parameters generally means more capacity to learn patterns, more nuance, better reasoning — but also dramatically more computational cost to run.

Training is when a model is created — the months-long process of feeding billions of examples of text through the neural network and adjusting its parameters until it becomes good at predicting text patterns. This happens once (per model version), costs tens to hundreds of millions of dollars, requires enormous GPU clusters, and permanently "freezes" what the model has learned at a particular point in time.

Inference is when a trained model is actually used — the milliseconds-to-seconds process of running your prompt through the frozen model to get a response. This is what happens every time you send a message on Claude.ai or ChatGPT. The model is not learning anything during inference; it is applying what it learned during training to your specific input.

A unimodal model understands one type of input — typically text. A multimodal model understands multiple types — text, images, audio, video, and code — and can reason across them together. Ask a multimodal model "what's wrong with this chart?" and it reads the image visually, understands the axes and data, and gives you a thoughtful critique. Ask "does this code match the architecture diagram?" and it reads the diagram and the code simultaneously.

Every frontier model in 2026 is multimodal by default. The question has shifted from "can it handle images?" to "how well does it handle each modality and how seamlessly does it reason across them?"

Traditional AI (2020–2023) was fundamentally reactive: you ask a question, it answers. Agentic AI is proactive: you give it a goal, and it figures out the steps, uses tools, takes actions, monitors outcomes, and iterates — all without needing you to approve every step.

An agent might be given "research and write a report on our top three competitors." It then decides to search the web multiple times, extract data from websites, analyse the results, draft sections, check for inconsistencies, and produce a final document — all autonomously. Each of those sub-steps might itself call another AI model or tool.

Retrieval-Augmented Generation (RAG) is the architecture behind NotebookLM, Perplexity, and enterprise AI search systems. The idea is simple: instead of asking the model to recall information from its training, you retrieve the actual relevant documents first, then feed them into the model's context window alongside the question. The model is then grounded in real text rather than reconstructed memory.

The retrieval step is where it gets interesting. You can't just keyword-search across thousands of documents in real time — you need a smarter way to find relevant content. The solution is embeddings: a numerical representation of text's meaning.

Every time an AI model generates the next word (token), it doesn't pick the single most-likely option — it samples from a probability distribution. The word "sun" might have a 40% probability, "moon" 25%, "sky" 20%, "rain" 10%, "cheese" 0.01%. Temperature controls how much the model "spreads" that distribution before sampling from it.

At temperature 0, the model always picks the highest-probability token — completely deterministic and reproducible. At temperature 1, it samples from the natural distribution. At temperature 2, lower-probability options become more likely — the output becomes more varied, creative, and occasionally strange.

Prompting is the simplest form of customisation: you write careful instructions that tell the model how to behave for your use case. A "system prompt" at the start of every conversation can set the tone, role, constraints, and output format. With modern frontier models, a well-crafted prompt can achieve remarkable domain adaptation — this is why prompt engineering became its own discipline.

Fine-tuning is training the model further on a curated dataset of examples specific to your domain — essentially updating the model's weights to internalise patterns from your data. A medical company might fine-tune a model on hundreds of thousands of clinical notes to improve its medical terminology, tone, and reasoning style.

A model trained purely to predict the next token on internet text will reproduce whatever patterns dominate the internet — including misinformation, harmful content, and persuasion techniques. The "alignment" challenge is shaping the model's behaviour to be helpful, honest, and harmless without breaking its underlying capabilities.

RLHF (Reinforcement Learning from Human Feedback) is the dominant technique. After base training, human raters compare pairs of model outputs and indicate which is better. A reward model learns from these preferences. The main model is then optimised to produce outputs the reward model rates highly. The result is a model that's better at following instructions, more helpful, and less likely to produce harmful content.

Constitutional AI (CAI), developed by Anthropic, takes a different approach: rather than relying entirely on human feedback, it trains the model on a set of written principles (a "constitution") and has the model critique and revise its own outputs against those principles. This reduces reliance on expensive human labelling and makes the model's values more explicit and auditable.

The term "vibe coding" was coined by AI researcher Andrej Karpathy to describe a new workflow: you describe what you want in plain English, the AI generates the code, you run it, describe what's wrong, and iterate — treating the codebase as something that materialises through natural language rather than manual typing. "Vibe" refers to communicating intent and aesthetic direction rather than technical specifications.

This isn't magic. Under the hood, a very specific sequence of events is happening — and understanding it helps you use the workflow far more effectively.

MCP (Model Context Protocol) is an open standard, published by Anthropic in November 2024, that defines a common way for AI models to connect to and interact with external tools, data sources, and services. Before MCP, every AI tool had its own proprietary way to connect to things — Cursor had its own plugin format, Claude had its own API format, and getting an AI to use a database meant custom code every time.

MCP is analogous to what USB did for peripheral devices, or what HTTP did for web communication. Before USB, every device had its own connector and driver. USB made any device plug into any computer. MCP aims to make any tool pluggable into any AI model — write the integration once, use it everywhere.