LLMs in Seven Minutes

Maya: If you've ever used a chat tool and thought, "Okay, what are you, exactly... a search engine, a robot intern, or just very confident autocomplete?"...same. Rafi: The robot intern comparison is brutal, and also, uh, sometimes accurate. Helpful, fast, occasionally a little too sure of itself. Maya: Exactly. So today, quick crash course. What a large language model, or LLM, actually is, how it works, what it can do, what it can't, and a few habits that make it way more useful. Rafi: And hopefully answer the big question, which is, why does this thing sound smart? Maya: Right. Simplest version, an LLM is a prediction engine for text. You give it some text, and it predicts what token should come next. Rafi: Okay, pause on token, because if you say "word," some very online person will appear in the doorway. Maya: Fair. A token is often a word piece, not necessarily a whole word. So a long word might get split into chunks, and punctuation can be its own token too. The model is working with these pieces of text. Rafi: So under the hood it's not going, "Ah yes, I truly understand your feelings." It's going, "Given these text pieces, what usually comes next?" Maya: That's the core idea, yes. And also, important, it's generating, not just searching and pasting. It has learned patterns of language well enough to produce fresh text one piece at a time. Rafi: Which sounds almost disappointingly simple until you remember how much of language is patterns. Maya: Exactly. If a model has seen huge amounts of text, it learns that certain tokens tend to show up near others, in certain tones, formats, situations. That's where embeddings help as a mental model. Rafi: The map thing, right? Maya: Yep. Think of tokens, and really concept-like patterns, as getting coordinates on a giant map. Things used in similar contexts end up near each other. So "recipe" and "ingredients" live in the same neighborhood. "Lawyer" and "contract" are hanging out nearby. Rafi: Tiny conceptual neighborhoods. I like this. Little cul-de-sacs of meaning. Maya: Exactly. Not because the model has a little city in its head, but because language puts related things together often enough that the model learns their positions relative to each other. Rafi: Okay, now the famous word. Transformer. It sounds like a toy and a power station at the same time. Maya: Honestly, yes. The big idea in transformers is attention. I think of attention like a spotlight. When the model is deciding what token to generate next, it can shine that spotlight back over the earlier text and weigh which parts matter most right now. Rafi: So if I write, "The trophy doesn't fit in the suitcase because it's too small," attention helps the model figure out what "it" probably refers to. Maya: Right. It can weigh different parts of the sentence instead of treating everything as equally important. That's a big reason these models handle longer prompts and multi-step instructions much better than older approaches. Rafi: So how does it get all this in the first place? How does it learn the map and the spotlight behavior and all that? Maya: Pretraining. Imagine reading millions of books' worth of text, articles, websites, code, conversations, all kinds of language, and constantly practicing, "What's the next token?" Over and over across huge text. That teaches the model grammar, style, structure, and a lot of the patterns people encode in writing. Rafi: Not memorizing every page, exactly, but absorbing the statistical habits of language. Maya: Right. Then fine-tuning is specialization. After that broad training, you tune it for a narrower job or behavior. Maybe better customer support, better coding help, safer chat behavior, that kind of thing. Rafi: And then when I open a chat window at, like, 11:43 p.m. and ask for an email that sounds calm but not weirdly passive-aggressive... Maya: Very specific hypothetical. Rafi: Entirely fictional. Maya: That's inference. Real-time use. Your prompt gets broken into tokens, the model maps them into that learned space, uses attention to weigh the context, and generates an answer token by token. Rafi: Let's do a real example, because this is where people actually live. Maya: Great. Say your prompt is: "Summarize this email and suggest a polite reply." And the email says, "Hi Rafi, just checking whether you can deliver the draft by Friday. If not, let me know a realistic timeline. Thanks." Rafi: A plausible model reply might be: "Summary: The sender is asking whether you can deliver the draft by Friday and wants an updated timeline if not. Suggested reply: Hi, thanks for checking in. I can deliver the draft by Friday and will keep you posted. Best, Rafi." Maya: Which sounds smooth... but could be a problem if you cannot deliver by Friday. Rafi: Right, the model just confidently volunteered my future. Maya: Exactly. So a small prompt change helps a lot. Try: "Summarize this email in 2 bullet points. Then draft a polite reply that does not commit to Friday, asks for an extension until Tuesday, and uses a warm professional tone. Base the reply only on the email and my instruction. If anything is unclear, say what needs clarification instead of guessing." Rafi: That's instantly better. You gave it the format, the boundary, the tone, and a little fence around the guessing. Maya: Yes. Being specific is one of the best habits with these tools. Rafi: So let's hit the limits, because this is where people either get way too impressed or way too cynical. Maya: Totally. LLMs are great for drafting, summarizing, brainstorming, reorganizing notes, rewriting in a different tone, and explaining things in simpler language. But they're not authorities. They can be wrong in a very polished voice. Rafi: The world's most articulate "maybe." Maya: Exactly. They can invent details, miss nuance, or sound more certain than they should. And they don't have lived experience or human judgment. They're working from patterns in text. Rafi: So smart habits. Quick list. Maya: First, be specific. Say what you want, what you don't want, and what format you want. Rafi: Second, verify important output. Especially anything medical, legal, financial, safety-related, or high-stakes for work. Maya: Third, treat the model like a helpful drafting partner, not a final authority. Rafi: And fourth, if you're sharing something sensitive, pause. Maybe remove personal details, or ask whether this even needs to go into a chat tool. Maya: Yes. And if the first answer is vague, ask a follow-up. "Give me three options." "Use plainer language." "What are you uncertain about?" The second round is often much better. Rafi: So the big picture is, an LLM is a very capable text prediction system. It works with tokens, learns relationships through something like a map, uses attention like a spotlight on context, and generates answers in real time. Maya: And once you see that mental model, the behavior makes more sense. Why it's fluent, why it's useful, and why it can still be confidently wrong. Rafi: Which, honestly, describes a surprising number of humans too. Maya: True. On that note, use the tools, stay curious, and keep one hand on the verify button. Rafi: That's the show. See you next time.