CODEAI101 · chapter 1 of 7 · free to read
This is the whole chapter, not a sample of it. Nothing is shortened for the showcase and nothing here is recorded — no account, no cookie, no reading history. Every chapter of this ladder was signed by a named person before it was served; this one by Capt Anil Sharma.
You finished Code101 knowing what software is, how the internet carries a message, why the shape of data decides what you can ask of it, what code is, how identity is proven, why memory forgets and storage does not, and how all of it fails. Today the second half begins, and it begins by taking the most oversold word of the decade apart.
There is nothing to install and no API key to hold. This ladder carries no shell commands at all. It is read and reasoned through, and every paired conversation it asks for can be had in a chat window you already have open, free tiers included.
And the part that is easy to leave unsaid: ANKR provides no machine. There is no VM, no account, no sandbox, no proxy bundled with any course on this Academy, and no arrangement under which you get one. The ladder after this one does ask something of you — CLAUDECODE101 needs a machine you control and your own Anthropic account, billed by them to you, because the fee buys the course and never the tokens. That is stated on its card before you pay and again in its first chapter, along with the fact that its cheaper-model chapter is the twenty-first of twenty-one rather than the starting point. Nothing of the kind is required here.
A language model is a system that, given some words, predicts the next word. That is the whole mechanism, all the way down. Every general-purpose assistant you have heard of, every coding assistant named on Day 1 of Code101, every chat product on the market, is at the bottom doing one thing: guessing the next token — roughly a word, or part of one — from every token that came before it.
It does not think. It does not understand in the way you understand. It has read an enormous quantity of text and learned which words tend to follow which others, in which contexts, with what probabilities. When it produces a reply that sounds like reasoning, that is because human writing already encoded reasoning, and the model is imitating the shape of it.
This is load-bearing for everything that follows. The model will produce a confident sentence about a dose adjustment in pregnancy, about the interaction between two prescriptions, about whether your filing may claim a deduction, about how a clause in a rental agreement should be read. The sentence will sound authoritative. The model has no idea whether it is true. It is the shape of a true sentence. Truth has to be checked against something else — a textbook, a regulation, a doctor, a lawyer, a real source.
A LANGUAGE MODEL IS A FLUENT IMITATOR OF HOW KNOWLEDGE SOUNDS, NOT A SOURCE OF KNOWLEDGE.
Re-read that. The fluent surface is what makes these systems useful and what makes them dangerous, in exactly the same measure. The literacy is to use the fluency without trusting the substance — to check, every time the stakes are real, that the sentence corresponds to something true outside the model.
You will hear all four from vendors, in articles, from colleagues. Most people use them interchangeably, which is wrong.
LLM, a large language model, is trained on text from the whole internet plus books, code and papers. Big, expensive, knows a little about everything. ChatGPT, Claude, Gemini and DeepSeek are LLMs. The size of a small data centre to train, millions or billions to produce.
SLM, a small language model, is trained or tuned for one domain. Cheaper to run, faster to answer, knows one thing well. An SLM for legal contracts knows contracts deeply and cannot help you with chess. An SLM for medical billing knows billing codes and cannot summarise a poem. Some run on hardware as cheap as a phone.
RAG, retrieval-augmented generation, is when the model is handed a stack of documents at the moment of the question and asked to answer from them. The documents are the source of truth; the model is only the translator between your question and what those documents say. Most "chat with your PDF" products are RAG, and Day 5 is spent entirely on it.
LoRA, low-rank adaptation, teaches an existing model a domain without retraining it from scratch. Cheap fine-tuning, useful when you hold a stack of specific data — a firm's contracts, a teacher's notes, a hospital's case files — and want a general model to behave like a specialist in it. Days rather than months.
Why it matters: a vendor can sell you any of the four and call all of them AI. Buying RAG is fine. Buying RAG while believing you bought an SLM that learned your business means your expectations are wrong and nobody told you. And when someone asks whether a system is an LLM or an SLM, the honest answer is usually both, depending on the layer. A mature product uses domain depth where depth is needed and broad knowledge where breadth is needed. The product is the layering, not any one of the four words.
There is a shape that recurs. Someone takes a general model, writes a prompt telling it to be an expert assistant for some domain, puts a login screen in front, and charges seriously for it. That shape has a name — a wrapper — and three tests find it.
THE SAME ANSWERS REGARDLESS OF INPUT. Ask three different domain questions. If the replies sound generically intelligent but cite nothing specific to your situation, you have your answer. Eloquent paragraphs about contract law that cannot cite the actual section of the actual statute; neat sentences about depreciation that cannot show the specific rule or the specific books.
NO DECLARED DATA SOURCE. The system cannot tell you where the answer came from. Press for the document, the regulation, the calculation. A real product points at something specific. A wrapper offers reassurance — the model is trained on industry data — instead of a citation you could check.
NO VERIFICATION PATH. Ask how you would trust the output, and there is nothing beyond the model is very good. No audit trail, no citation, no proof that this particular answer corresponds to anything real. The product is one fluent paragraph at a time, and a paragraph cannot be audited.
Wrappers fail in production for one reason: they cannot survive a dispute. When real money or a real outcome is on the table, "the AI said so" is never an admissible answer. The operator has to point at the document, the calculation, the regulation. A wrapper cannot, because it was never built to — it was built to look fluent. Recognising the shape is half the literacy, and the three questions fit in your head during any demo: does the answer change with the input, where is the source, how would I verify it?
Note what is being judged here. The test is of a product's construction, not of anyone's competence or honesty. Plenty of wrappers are built by capable people solving a real problem quickly, and some of them will earn a substrate later.
If a wrapper is a fluent layer with nothing underneath, the opposite is a substrate: software with its own declared structure — what data it holds, what operations it claims to perform, what it observes, what proof it can produce of any answer it gives. The model sits on top as a translator between human language and what the substrate already knows. The substrate is the source of truth; the model is the speaker.
Ask such a product a question and four things happen. The model works out what you actually want to know. The substrate runs a real operation on real data — a query, a calculation, a lookup, a verification. The substrate returns an answer with a source: a specific document, a specific row, a specific computed result. Then the model puts that answer into ordinary language, with the citation still attached.
This is not theoretical. Your bank's app, your tax software, a hospital records system — all of them are substrates that have since added an AI translator. The substrate was always there. Hold one claim from it: A WRAPPER SELLS YOU THE TRANSLATOR; A SUBSTRATE SELLS YOU THE SOURCE AND THE TRANSLATOR, WITH A CITATION BETWEEN THEM. That is the difference between AI that survives an auditor's question and AI that does not. Day 6 goes deeper. For now it is enough to know the distinction exists, and to ask at the next demo whether there is a substrate at all, or whether the AI is all there is.
The coding assistants you have heard about since Day 1, and the chat assistants you have been talking to in your paired conversations, are language models with wrappers and tool layers around them. They are extraordinarily useful. They are also exactly what this chapter began with: fluent imitators of how code and explanation sound.
Which is why the job, in every paired conversation, is to read what the assistant produces and check that it matches what you actually wanted, rather than accepting it because it reads well. The five-step pattern you used through Code101 — set context, ask small, get the code, ask for the explanation, save it — is the workflow that turns fluent-but-unverified into verified-and-understood. The explain-every-line step is the verification step. Without it you are using a wrapper-grade assistant. With it you are using the same model in a substrate-shaped way, and the substrate is you: your specific knowledge of your own project, checking every sentence the model hands you.
The rest of this ladder is for enrolled learners. This chapter is the whole of what is open — it is not a teaser with the ending removed.