<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Code Plu on Medium]]></title>
        <description><![CDATA[Stories by Code Plu on Medium]]></description>
        <link>https://medium.com/@info.codeplu?source=rss-5637d316b425------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*AxUQgqPmN-obsfGu4XnABw.jpeg</url>
            <title>Stories by Code Plu on Medium</title>
            <link>https://medium.com/@info.codeplu?source=rss-5637d316b425------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 03:09:46 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@info.codeplu/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[What Are the Main Types of Prompts?]]></title>
            <link>https://medium.com/@info.codeplu/what-are-the-main-types-of-prompts-7b811d4d3af3?source=rss-5637d316b425------2</link>
            <guid isPermaLink="false">https://medium.com/p/7b811d4d3af3</guid>
            <category><![CDATA[generative-ai]]></category>
            <category><![CDATA[types-of-prompt]]></category>
            <category><![CDATA[writing-prompts]]></category>
            <category><![CDATA[prompt-engineering]]></category>
            <category><![CDATA[ai-prompts]]></category>
            <dc:creator><![CDATA[Code Plu]]></dc:creator>
            <pubDate>Thu, 08 Oct 2026 16:55:42 GMT</pubDate>
            <atom:updated>2026-10-08T16:55:42.555Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WNVj39EwCfii489a4nOREg.png" /></figure><h3>Root Concept</h3><p>Prompts come in a few <a href="https://codeplu.com/content/concept/types-of-prompt"><strong>recognisable kinds, </strong></a>and each one suits a different job — the skill is matching the way you ask to what the task actually needs.</p><h3>Why Are There Different Types of Prompts?</h3><p>Think about how you ask people for things. If you want a coffee, you just ask. If you want your handwriting copied onto forty wedding invitations, you hand over a sample and say ‘like this’. And if you want someone to check your tax arithmetic, you ask them to show their working — because the answer alone is not enough, you need to see how they got there.</p><p>Prompts work the same way. There is no single best way to ask an AI; there are a few recognisable kinds, and each earns its keep on a different sort of job. Plain requests handle ordinary tasks. Prompts carrying examples pin down a style or shape. Prompts that ask for the steps expose the reasoning where mistakes hide. And prompts that assign a role quietly set the audience and tone.</p><p>Knowing the kinds matters less than knowing when to reach for each. In the playground above you will build a prompt down into three ways of asking — and the labels on your connections tell you when each one is the right call.</p><h3>Which Kind of Prompt Should You Use?</h3><h3><strong>1. </strong>When is a plain request all you need?</h3><p>For the huge majority of everyday tasks, just asking works: summarise this, translate that, fix the grammar here, explain this word. These are things the AI has effectively done a thousand times while learning, so it already knows what a good answer looks like. Engineers have a name for this — a zero-shot prompt, meaning you gave it zero examples. Do not read ‘plain’ as ‘inferior’. A simple request with a clear task and audience beats an elaborate one most of the time, and it takes seconds to write. Reach for the fancier kinds when a plain request has actually let you down, not in advance out of habit.</p><h3><strong>2. </strong>Why does showing examples work so well?</h3><p>Some things are almost impossible to describe and trivial to demonstrate. Try explaining in words exactly how your product titles are worded, or the particular way your team writes meeting notes — then compare that with pasting two real ones and saying ‘twenty more like these’. The AI is extremely good at spotting a pattern and continuing it, so examples carry style, length, structure, and tone all at once. This is what people mean by a few-shot prompt: a few examples included in the request. Reach for it whenever the answer must match something that already exists, especially when you would struggle to turn that something into rules.</p><h3><strong>3. </strong>When should you ask the AI to show its steps?</h3><p>For anything involving arithmetic, comparisons, logic, or a decision with several moving parts, ask the AI to work through it in order instead of jumping to an answer. The reason is practical: answering in one leap can produce a confident, fluent, completely wrong result — and a wrong answer with no working is undetectable. When the steps are laid out, mistakes become visible and you can spot the bad assumption in the third line. Many beginners skip this because typing ‘work through it step by step’ feels like babysitting the machine. It is the opposite: it is the cheapest quality check you will ever apply to an AI answer.</p><h3><strong>4. </strong>What does giving the AI a role actually do?</h3><p>‘Act as an experienced maths teacher’ is the most famous prompting trick in the world, and it does work — but not for mysterious reasons. Assigning a role is a fast way of supplying context: in four words it sets the expertise level to write at, the audience to imagine, and the tone to adopt. That is genuinely useful shorthand. What it cannot do is make the AI know more than it knows. Calling it a doctor adds no medical knowledge; calling it a lawyer does not make its advice safe to act on. Treat a role as a convenient way to set voice and level, never as a power-up for reliability.</p><h3><strong>5. </strong>How do you pick the right kind — and can you mix them?</h3><p>Start plain, then escalate based on what actually went wrong. If the answer was right in substance but wrong in style, add examples. If it was fluent but the reasoning looked shaky, ask for the steps. If it was pitched at the wrong level, set a role or name the audience directly. And yes, these mix freely: a real working prompt often assigns a role, includes one example, and asks for step-by-step reasoning, all inside the four-part structure of task, context, format, and limits. The types are not rival techniques to choose between — they are tools on one belt, and the job in front of you decides which you pick up.</p><h3>Real World Example</h3><p><strong>Same AI, same afternoon — two different jobs needing two different kinds of asking.</strong></p><h3>How Did One Shop Owner Fix Two Bad AI Answers?</h3><p>Maya runs a small online shop. She tries an AI assistant on two very different tasks, gets poor results on both, and fixes each by changing the kind of prompt rather than the tool.</p><h3><strong>1. </strong>Job one: she just asks, and gets a mess</h3><p>Maya needs twenty product titles for new stock, so she types ‘Write 20 product titles for my handmade candles’. Back come twenty titles in about twenty different styles — some poetic, some shouty, some with prices — none of them matching the fifty titles already live on her shop.</p><h3><strong>2. </strong>Job one, fixed: she shows what good looks like</h3><p>She pastes two of her existing titles — ‘Lavender Soy Candle — 40 Hour Burn, Hand Poured’ — and writes ‘twenty more in exactly this style’. The pattern lands immediately, and all twenty come back consistent enough to publish with barely an edit.</p><h3><strong>3. </strong>Job two: she just asks, and gets a confident wrong answer</h3><p>Next she compares two suppliers with different unit prices, delivery fees, and bulk discounts, asking which is cheaper over two years. The AI answers instantly and decisively — Supplier B — with a total that, when she checks it on paper, is simply wrong.</p><h3><strong>4. </strong>Job two, fixed: she asks for the working</h3><p>She asks again:<a href="https://codeplu.com/content/concept/types-of-prompt"><strong> ‘Work it out step by step — show the unit costs,</strong></a> the delivery, the discount, and the yearly totals before you answer.’ This time the arithmetic appears line by line, she can see exactly where the earlier reply went astray, and the cheaper supplier turns out to be A.</p><h3><strong>5. </strong>What actually changed</h3><p>Nothing about the AI. Maya matched the kind of asking to the job: examples where a style had to be copied, steps where arithmetic could hide a mistake. Each fix took one extra sentence — and both prompts now sit in her notes, ready to reuse next season.</p><h3>Final Words</h3><p>There is no single correct way to prompt — there are a few kinds of asking, and the job decides between them. Plain requests for everyday work. Examples when a style or shape must match. Steps when a mistake could hide inside confident-sounding reasoning. A role when the level and tone need setting quickly. In the playground you matched three of these to the jobs they fit, and the labels showed you why each pairing works.</p><p>Two habits will carry you a long way: start plain, and escalate based on the actual failure rather than out of habit. And remember these kinds stack freely inside the four-part structure you already know. Once you have chosen your kind of prompt and built it well, one thing remains — running it, reading the answer critically, and refining. That is the prompt iteration loop.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7b811d4d3af3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Is the Structure of a Good Prompt?]]></title>
            <link>https://medium.com/@info.codeplu/what-is-the-structure-of-a-good-prompt-73d35acff0fb?source=rss-5637d316b425------2</link>
            <guid isPermaLink="false">https://medium.com/p/73d35acff0fb</guid>
            <category><![CDATA[generative-ai]]></category>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[prompt-structure]]></category>
            <category><![CDATA[prompt-engineering]]></category>
            <category><![CDATA[writing-prompts]]></category>
            <dc:creator><![CDATA[Code Plu]]></dc:creator>
            <pubDate>Wed, 07 Oct 2026 07:31:01 GMT</pubDate>
            <atom:updated>2026-10-07T07:31:01.651Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kCEOJJ2WGrBOJj0kZAWlmA.png" /></figure><h3>Root Concept</h3><p>A good prompt answers <a href="https://codeplu.com/content/concept/structure-of-a-good-prompt"><strong>four questions before the AI has to guess them:</strong></a> what to make, who it is for, what shape the answer takes, and what to leave out.</p><h3>What Goes Into a Well-Built Prompt?</h3><p>Walk into a bakery and say ‘chocolate cake, please’. You will get a cake — but probably not the one in your head. Now try it properly: ‘A chocolate cake for my daughter’s eighth birthday, twelve children coming. Round, one layer, with Happy Birthday Zara written on top. No nuts, she is allergic, and please keep it not too sweet.’ Same shop, same baker, wildly different odds of getting what you wanted.</p><p>Notice what the good order contains. What to make. Who it is for and why. What shape it should take. What to leave out. Those four pieces are exactly the structure of a good prompt — and each one exists for a single reason: to answer a question the AI would otherwise have to guess.</p><p>A well-built prompt is not a longer prompt. It is a prompt with no unanswered questions left in it. In the playground above you will assemble one part by part, in the order a sensible briefing naturally follows, and finish with the complete request.</p><h3>What Are the Four Parts of a Good Prompt?</h3><h3><strong>1. </strong>What is the task, and why is it not enough on its own?</h3><p>The task is the job itself — write, summarise, translate, compare, plan, fix. It is the one part no prompt can do without, because without it there is nothing to make. But a task alone is almost always underspecified. ‘Write a summary’ leaves open how long, for whom, and what to keep. ‘Summarise this article in five bullet points for someone who has not read it’ is the same task with the guessing removed. So the useful habit is not just naming the task, but naming it in a way that is hard to misread: prefer ‘compare these two phones on battery and price’ over the vague ‘tell me about these phones’.</p><h3><strong>2. </strong>What does context add that the task cannot?</h3><p>Context is who the answer is for and why it exists. It is the single most powerful part, because it silently changes everything — vocabulary, depth, examples, even what gets left out. Tell the AI ‘for a ten-year-old’ and it reaches for simple words and everyday comparisons. Tell it ‘for my manager, who will decide our budget’ and the same request produces something crisper and more careful. Many beginners assume context is optional decoration for the real instruction. In practice it is the difference between a technically correct answer and a usable one, and it is the part people forget most often — precisely because the audience is so obvious inside their own head.</p><h3><strong>3. </strong>How do you control the shape of the answer?</h3><p>Format is how the answer arrives: how long, and laid out how. Three sentences or two pages. A bulleted list, a table, an email, a numbered set of steps. This is the part that decides whether you can use the answer straight away or have to spend ten minutes reshaping it. It is also the easiest to state and the most satisfying to get right — ‘as a five-row table with columns for pros, cons, and price’ takes four seconds to type and saves the whole edit. And when the shape is fussy or unusual, do not describe it, show it: one small example of the layout you want communicates more than a paragraph of instructions.</p><h3><strong>4. </strong>Why tell the AI what to leave out?</h3><p>Limits are the boundaries: no jargon, under 200 words, do not mention price, no invented statistics, avoid clichés. They feel negative, and beginners often skip them thinking restrictions can only make an answer worse. The opposite is true — limits are what stop the predictable failures. AI writing drifts toward padding, hedging, and confident guesses, so the instructions that ban those behaviours protect the answer’s quality. A useful trick: whenever an answer disappoints you the same way twice, that recurring annoyance is a limit waiting to be written down, and once written it keeps working every time you reuse the prompt.</p><h3><strong>5. </strong>Does the order of the parts matter?</h3><p>A little, and less than people expect. The AI reads your prompt from top to bottom, so leading with the task and the audience frames everything that comes after — which is why the natural briefing order works well: task, context, format, limits. What matters far more is that all four are present. And beware the trap this can lead to: seeing that detail helps, beginners start padding prompts with filler, repetition, and flattery. A bloated prompt buries its own instructions, and the rule that matters gets lost in the noise. Complete is the goal; long is a side effect at best. Four clear lines beat four paragraphs every time.</p><h3>Real World Example</h3><p><strong>One request, assembled part by part — watch the answer become usable.</strong></p><h3>How Do You Build a Prompt to Sell Your Old Bicycle?</h3><p>You want an advert <a href="https://codeplu.com/content/concept/structure-of-a-good-prompt"><strong>for a second-hand bicycle you are selling online.</strong></a> Follow the same four parts you connected in the playground, added one at a time.</p><h3><strong>1. </strong>Start with the task alone</h3><p>‘Write an advert for my bicycle.’ The AI returns something breathless and generic — ‘Amazing bike, excellent condition, must see!’ — that could describe any bicycle on earth. Nothing is wrong with it except that it is useless.</p><h3><strong>2. </strong>Add the context: who is reading, and why</h3><p>‘…a six-year-old city bike with a few scratches, being sold to a student on a small budget. Honest about the wear.’ Now the AI knows the buyer and the tone, and it stops overselling — it starts sounding like a real person with a real bike.</p><h3><strong>3. </strong>Add the format: how it should arrive</h3><p>‘Under 80 words: two short lines of description, then a bulleted list of the specs, then the asking price on its own line.’ The answer arrives shaped like an actual online listing, ready to paste into the form without rearranging anything.</p><h3><strong>4. </strong>Add the limits: what to leave out</h3><p>‘No exclamation marks, no words like amazing or perfect, and do not invent any details I have not given you.’ This kills the salesy tone, and — importantly — stops the AI from cheerfully making up a brand name or a service history.</p><h3><strong>5. </strong>The finished prompt, and why it keeps paying</h3><p>Four parts, roughly forty words, and the answer is a listing you would genuinely post. Better still, the prompt is now reusable: next time you sell something, swap the item and the specs and everything else still holds. That is what structure buys you — not one good answer, but a template for many.</p><h3>Final Words</h3><p>A good prompt is not a magic sentence — it is a briefing with no gaps. Name the task so it cannot be misread, name the reader and the purpose, name the shape you want back, and name what to leave out. In the playground you assembled exactly that, one part at a time, in the order a sensible briefing follows.</p><p>Keep the bakery in mind: the baker was always capable, the order was what changed. And once one of your prompts works well, keep it — swap the details and the structure carries over to the next task. Next comes the natural question: prompts also come in recognisable kinds, each suited to a different job. That is the story of the types of prompts.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=73d35acff0fb" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Is Prompt Engineering?]]></title>
            <link>https://medium.com/@info.codeplu/what-is-prompt-engineering-108efe9ac4ef?source=rss-5637d316b425------2</link>
            <guid isPermaLink="false">https://medium.com/p/108efe9ac4ef</guid>
            <category><![CDATA[generative-ai]]></category>
            <category><![CDATA[ai-explained]]></category>
            <category><![CDATA[prompt-engineering]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[prompt-crafting]]></category>
            <dc:creator><![CDATA[Code Plu]]></dc:creator>
            <pubDate>Tue, 06 Oct 2026 06:02:54 GMT</pubDate>
            <atom:updated>2026-10-06T06:02:54.625Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1E2KmGNQdlKWEiZ33jXD5g.png" /></figure><h3>Root Concept</h3><p>The AI is fixed — your words are the only thing you control. Prompt engineering is the skill <a href="https://codeplu.com/content/concept/what-is-prompt-eng"><strong>of steering the same AI toward the answer you actually wanted.</strong></a></p><h3>What Is Prompt Engineering, in Plain Words?</h3><p>Two people sit down with the exact same AI chatbot on the same afternoon. One walks away calling it overhyped and useless. The other gets work done in ten minutes that would have taken an hour. The AI was identical for both. The only thing that differed was what they typed.</p><p>That gap is what prompt engineering is about. A prompt is simply the request you type. Prompt engineering is the skill of writing that request so the AI gives you what you actually wanted — not something technically related but practically useless. Here is the key realisation: you cannot change the AI. It is already built and trained. Your words are the one and only steering wheel you have.</p><p>So this is not a programming skill, and there are no secret magic phrases. It is the everyday skill of saying clearly what you want — the same skill that separates a helpful set of directions from ‘it’s near the big tree’. In the playground above you will build both paths from one goal: the same question asked in a hurry, and asked with care, ending in two very different answers.</p><h3>How Does Prompt Engineering Actually Work?</h3><h3><strong>1. </strong>Why does the same AI give one person better answers?</h3><p>Because the AI has no idea what you know, what you need, or who the answer is for — unless you say so. Ask ‘tell me about dogs’ and it has to guess whether you want a vet’s explanation, a bedtime story, or a shopping guide. It picks something safely average, which is exactly why average is all you get. Many beginners conclude the AI is simply not that smart. Watch what happens when a skilled user takes the same chatbot and gets a perfect answer in one try, and the truth becomes obvious: the intelligence was always there, waiting for direction. Your words are the steering wheel, and a vague prompt is letting go of it.</p><h3><strong>2. </strong>What can your prompt actually control?</h3><p>Far more than most people realise. You can set who the answer is for — a child, your boss, a beginner — and the AI will change its whole vocabulary. You can set the length: three sentences, one page, five bullet points. You can set the tone: warm, formal, funny, blunt. You can set the shape: a table, a list, an email, a step-by-step recipe. And you can rule things out: no jargon, no invented facts, no mention of price. Every one of these is a dial you either set on purpose or leave to chance. Prompt engineering is nothing more mysterious than turning the dials that matter for the answer you need.</p><h3><strong>3. </strong>Is prompt engineering the same as programming?</h3><p>No — and this is the misconception that keeps people from trying. Traditional programming means writing exact instructions in a special language, where a single missing comma breaks everything. A prompt is written in ordinary human language, and small imperfections are fine. If you can explain a task clearly to a new colleague, you can write a good prompt. What the two skills share is a mindset, not a syntax: being precise about what you want, thinking about edge cases, and expecting to improve your first attempt. That is why people who are good at giving instructions to humans usually turn out to be good at prompting too.</p><h3><strong>4. </strong>Why doesn’t the AI just ask me what I meant?</h3><p>A thoughtful colleague, handed a vague request, would ask a question back: ‘For whom? How long? By when?’ Most AI tools do not do this by default — they take your words at face value and produce something immediately. So the burden of clarity sits with you, up front. It helps to picture the AI as an extremely well-read new assistant on their first hour of work: enormous knowledge, willing, fast — and zero context about your life, your audience, or your standards. Nobody would hand that person a two-word instruction and expect a usable result. A prompt is the briefing you would naturally give them.</p><h3><strong>5. </strong>Do I need secret magic words to get good results?</h3><p><a href="https://codeplu.com/content/concept/what-is-prompt-eng"><strong>No. Collections of ‘magic phrases’ </strong></a>circulate constantly, and beginners often believe the skill is about knowing the right incantations. It is not — most of those phrases work only because they smuggle in real information the prompt was missing. ‘Act as an experienced teacher’ helps not through magic, but because it tells the AI the audience and the tone. Once you understand that, you can skip the tricks and state what you mean directly: who it is for, how long, what tone, what to avoid. Clarity beats cleverness every time, and unlike a list of phrases, clarity keeps working when the AI tools change.</p><h3>Real World Example</h3><p><strong>Same task, same AI, same day — one gets a template, the other gets a message worth sending.</strong></p><h3>Why Did Two Job Applicants Get Such Different Help?</h3><p>Two people have just finished interviews and want to send a thank-you email. Follow what each of them types, and what comes back.</p><h3><strong>1. </strong>Person A types the obvious thing</h3><p>‘Write a thank you email after a job interview.’ Four seconds later they have a polite, colourless template: ‘I am writing to express my gratitude for the opportunity…’ Correct English, zero substance, recognisable instantly as AI filler by any recruiter who reads it.</p><h3><strong>2. </strong>Person A blames the AI and gives up</h3><p>They conclude the tool is useless for real work and go back to writing it themselves. Nothing about the AI was the problem — their nine-word request simply never mentioned the job, the interview, the tone, or the length.</p><h3><strong>3. </strong>Person B briefs the AI like a colleague</h3><p>‘Write a short thank-you email — under 120 words — to Priya, who interviewed me this morning for a junior data analyst role. Mention that I enjoyed hearing about their customer dashboard project. Friendly but professional. No clichés like “express my gratitude”.’</p><h3><strong>4. </strong>Person B gets something they can send</h3><p>The reply names Priya, references the dashboard conversation, sounds like a human being, and fits in a phone screen. Person B changes two words and hits send. Total time: about the same as Person A spent being disappointed.</p><h3><strong>5. </strong>The only difference was the prompt</h3><p>Same model, same day, same task. Person B never learned a secret trick — they simply said who, what, how long, what tone, and what to avoid. That is prompt engineering in its entirety, and it is why the skill is worth ten minutes of your attention.</p><h3>Final Words</h3><p>Prompt engineering turns out to be far less exotic than its name suggests. The AI is fixed and already built; your words are the only lever you hold. In the playground you saw the whole idea in one picture — one goal, asked in a hurry and asked with care, producing a wall of useless facts on one side and three lines you could read aloud tonight on the other.</p><p>Carry the core idea with you: anything you leave unsaid, the AI will guess. Name the audience, the length, the tone, and the limits, and average answers turn into useful ones. Next comes the natural follow-up question — what exactly goes into a well-built prompt, and in what order? That is the anatomy of a good prompt, waiting in the next concept.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=108efe9ac4ef" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How Do LLMs Generate a Response?]]></title>
            <link>https://medium.com/@info.codeplu/how-do-llms-generate-a-response-c465bbc83cdd?source=rss-5637d316b425------2</link>
            <guid isPermaLink="false">https://medium.com/p/c465bbc83cdd</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[generative-ai]]></category>
            <category><![CDATA[how-llms-work]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[text-generation]]></category>
            <dc:creator><![CDATA[Code Plu]]></dc:creator>
            <pubDate>Mon, 05 Oct 2026 06:31:01 GMT</pubDate>
            <atom:updated>2026-10-05T06:31:01.355Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MenhriIS1c3Zw-it0FNcUg.png" /></figure><h3>Root Concept</h3><p>An LLM grows its answer <a href="https://codeplu.com/content/concept/how-llms-generate-response"><strong>one chunk at a time:</strong></a> it re-reads everything so far, scores what could come next, picks one, adds it, and goes around again — with no plan for how the answer ends.</p><h3>How Does an AI Actually Write Its Answer?</h3><p>Imagine writing a sentence under a strange rule: you may add only one word at a time, you must re-read the whole sentence before choosing each next word, and you are never allowed to go back and change anything. No outline, no draft, no editing. Somehow the sentence still has to make sense when you reach the end.</p><p>That is exactly how an AI writes. It does not compose an answer and then show it to you. It grows the answer: read everything so far, score every chunk that could come next, pick one, add it — then read everything again, including the chunk it just added. Round and round, a few hundred times for a single paragraph.</p><p>This one mechanism explains an astonishing amount of everyday AI behaviour: why answers appear word by word, why the same question gives differently worded replies, why an AI sometimes rambles or paints itself into a corner, and why it never takes back a sentence it has already started. Build the loop in the playground above and the rest of this concept clicks into place.</p><h3>What Happens on Each Trip Around the Loop?</h3><h3><strong>1. </strong>Why does the answer appear word by word?</h3><p>Because it is genuinely being made word by word. That typing effect is not a decorative animation designed to look thoughtful — each chunk is sent to your screen the moment it is chosen, while the rest of the answer does not exist yet. When the text pauses slightly, real work is happening: the entire model is running again to choose the next chunk. This also explains why long answers take longer in direct proportion to their length, and why asking for a shorter reply genuinely gets you a faster one. There is no finished paragraph waiting behind a curtain, being revealed slowly for effect.</p><h3><strong>2. </strong>Where does the randomness come from?</h3><p>After scoring every possible next chunk, the model does not always take the top scorer. It usually samples — leaning heavily toward the likely chunks while leaving room for less obvious ones. That small deliberate randomness is why asking the identical question twice gives two differently worded answers. Most tools expose it as a setting called temperature: low temperature sticks to the safest chunks and produces predictable, sometimes repetitive text, while higher temperature takes more chances and produces livelier but riskier text. Many beginners read variation as unreliability. It is simply a dial, usually set to a middle value on purpose.</p><h3><strong>3. </strong>Why can’t the AI plan the whole answer first?</h3><p>Because there is nowhere for a plan to live. Each trip around the loop produces exactly one chunk, chosen from what exists so far, with no separate outline held aside. The model has absorbed the shape of well-organised writing, so answers usually flow sensibly — but it is improvising in a very well-practised way rather than following an agenda. This is why it can begin a sentence it cannot gracefully finish, and why tasks that constrain the ending are surprisingly hard for it. It commits to the beginning long before it knows the end.</p><h3><strong>4. </strong>How does the model know when to stop?</h3><p>Stopping is just another choice. Among all the chunks it can pick sits a special one meaning ‘the answer is finished’, and when that becomes the likeliest option, generation ends. Because the model learned this from countless examples of complete answers, it usually stops somewhere sensible. Two other things can cut it off: the tool imposes a maximum length, or the shared token budget runs out — which is what has happened when an answer stops abruptly mid-sentence. If that occurs, simply asking it to continue usually works, because everything written so far is still in view.</p><h3><strong>5. </strong>Why does it never take back a mistake?</h3><p>Once a chunk is added it becomes part of the input for every later choice, so a wrong turn is not merely left in place — it actively shapes what follows. This is how a small early error grows into a confidently wrong paragraph: the model keeps writing text consistent with what it already said. It also explains why ‘work through it step by step’ improves accuracy, since the reasoning chunks written early become material read later, like showing your working on paper. And it is why the fix for an answer that has gone off the rails is a fresh, clearer prompt rather than an argument.</p><h3>Real World Example</h3><p><strong>A small request that exposes the loop perfectly — no plan, and no going back.</strong></p><h3>Why Does an AI Struggle to End a Sentence on a Given Word?</h3><p>Ask an AI for a sentence about the sea that ends with the word ‘orange’. Follow the loop and watch why this simple-sounding task is genuinely awkward for it.</p><h3><strong>1. </strong>The first chunk is chosen with no ending in mind</h3><p>It reads your request and picks something that fits a sentence about the sea: ‘The’. That is the entire decision. Nothing anywhere holds a plan for how this sentence will arrive at <a href="https://codeplu.com/content/concept/how-llms-generate-response"><strong>‘orange’ — there is only this chunk,</strong></a> chosen because it fits what came before it.</p><h3><strong>2. </strong>Each new chunk narrows what is still possible</h3><p>‘The’ becomes ‘The sun’, then ‘The sun sank’, then ‘The sun sank into the’. Every addition is sensible on its own, and every addition quietly shrinks the set of endings that could still work. The loop has no way to check the destination while doing this.</p><h3><strong>3. </strong>Sometimes it lands, sometimes it cannot</h3><p>From ‘turning the water a deep shade of’ the constraint is easy to meet. But had it begun ‘Fishermen returned at dusk with their nets’, there may be no graceful path to ‘orange’ left at all. It cannot rewind to try a different opening, so it improvises from where it is.</p><h3><strong>4. </strong>The failure looks like carelessness but is structural</h3><p>You may get a clumsy ending, or the word ‘orange’ appearing in the middle instead. It is not ignoring your instruction or being lazy. It committed to a beginning before knowing the end, exactly as the loop forces it to.</p><h3><strong>5. </strong>How to work with the loop instead of against it</h3><p>Ask it to plan out loud first — ‘suggest three possible endings, then write the sentence’ — and that plan becomes written chunks it can read while composing. You have given the loop somewhere to keep an outline, and the same constrained task suddenly becomes easy.</p><h3>Final Words</h3><p>An AI answer is grown, not written: read everything so far, score what could come next, pick one chunk, add it, and go around again. In the playground you built that loop including the edge that makes it a loop — the new chunk becoming part of the input.</p><p>Almost every quirk you will meet traces back to it. Streaming is real work happening live. Variation is a randomness dial. Rambling and awkward endings come from having no outline. And a wrong turn stays in view, shaping everything after it, which is why the fix is a fresh prompt rather than an argument. With this loop in hand you have the full picture of how a language model turns your words into an answer.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c465bbc83cdd" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How Does an LLM Turn Your Words Into Tokens?]]></title>
            <link>https://medium.com/@info.codeplu/how-does-an-llm-turn-your-words-into-tokens-c306ca571bd7?source=rss-5637d316b425------2</link>
            <guid isPermaLink="false">https://medium.com/p/c306ca571bd7</guid>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[text-preprocessing]]></category>
            <category><![CDATA[how-llms-work]]></category>
            <category><![CDATA[tokenization]]></category>
            <dc:creator><![CDATA[Code Plu]]></dc:creator>
            <pubDate>Sun, 04 Oct 2026 14:31:20 GMT</pubDate>
            <atom:updated>2026-10-04T14:31:20.091Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*f93ceorsf3062t1Qz9Eigw.png" /></figure><h3>Root Concept</h3><p>An LLM never sees your letters — <a href="https://codeplu.com/content/concept/llm-token-generation-pipeline"><strong>your text is chopped into tokens,</strong></a><strong> </strong>turned into numbers, predicted over, and stitched back into words. Tokens are the unit everything runs on.</p><h3>What Happens to Your Words Inside an LLM?</h3><p>Imagine a factory that can only handle bricks from its own fixed catalogue. You bring in a sculpture; before anything can happen, it must be broken down into catalogue bricks. At the end, bricks are glued back together and handed to you as a shape again. Everything the factory actually does happens at brick level — it never touches your original sculpture at all.</p><p>That is what happens to your sentence inside a large language model. Your text is chopped into small chunks called tokens, each token is swapped for a number, and all the machinery works on those numbers. When the answer is ready, the numbers are turned back into chunks and the chunks are glued into text on your screen.</p><p>This sounds like a technical detail, but it quietly explains a lot of AI’s real-world behaviour: why tools charge by tokens, why long chats start forgetting the beginning, and why an AI that can write a poem about strawberries cannot reliably count the letter r inside the word. In the playground above you will connect this journey stage by stage.</p><h3>How Does the Token Pipeline Work?</h3><h3><strong>1. </strong>What exactly is a token?</h3><p>A token is a chunk of text — often a whole common word, sometimes a piece of a longer one, and it usually carries its leading space along with it. In English a token averages roughly four characters, so a rule of thumb is that a thousand tokens is around seven hundred and fifty words. Everyday words like ‘water’ or ‘the’ are single tokens because they appear constantly. Longer or rarer words get split: ‘unbelievable’ might arrive as three chunks, something like<a href="https://codeplu.com/content/concept/llm-token-generation-pipeline"><strong> un / believ / able. </strong></a>Many beginners assume one token equals one word. It does not, and that gap is exactly where the surprising behaviour comes from.</p><h3><strong>2. </strong>Why must tokens become numbers?</h3><p>The model is, underneath, an enormous pile of arithmetic — multiplication and addition on a colossal scale. Arithmetic needs numbers, and letters are not numbers. So every chunk in the model’s catalogue has an ID, and your tokenised sentence becomes a list of IDs before anything else happens. This is the crucial consequence: the model never sees your letters. It sees a sequence of chunk IDs and works with the patterns between them. When we say an AI ‘reads’ your message, what really happens is closer to recognising a familiar sequence of catalogue numbers.</p><h3><strong>3. </strong>How does the model choose what comes next?</h3><p>With your tokens loaded, the model does the thing it was trained to do: score every chunk in its catalogue for how well it would continue the sequence. Not one guess — a likelihood for tens of thousands of possible chunks at once. One of them is then chosen and added to the sequence, and the process repeats to build the answer piece by piece. That piece-by-piece construction has consequences of its own, from why answers appear word by word on your screen to why the same question can come back worded differently — and those belong to the concept of how LLMs generate responses.</p><h3><strong>4. </strong>Why do tokens explain the limits you keep bumping into?</h3><p>Almost every practical limit in AI tools is measured in tokens, not words. A model’s context window is a token budget for everything it can hold at once — your instructions, the documents you paste, the conversation so far, and the answer being written. Exceed it and the oldest material falls away, which is precisely why a very long chat starts forgetting what you said at the beginning. Pricing works the same way: services bill per token in and per token out, so a rambling prompt costs real money. Once you see limits as a token budget, these behaviours stop looking like bugs.</p><h3><strong>5. </strong>Why does AI get confused counting letters in a word?</h3><p>Ask an AI how many times the letter r appears in ‘strawberry’ and you may get a confident, wrong answer. Now you know why: the model was never given letters. It received two or three <a href="https://codeplu.com/content/concept/llm-token-generation-pipeline"><strong>chunks — perhaps str / aw / berry — </strong></a>and inside a chunk the individual letters are simply not visible to it. Asking it to count letters is like asking someone to count the bricks inside a wall they have only ever seen as finished rooms. This is also why spelling puzzles, rhyming edge cases, and reversing words trip these systems up. The practical fix is easy: spell the word out with spaces or dashes — s t r a w b e r r y — and each letter becomes its own token that the model can actually see and count.</p><h3>Real World Example</h3><p><strong>One short question, followed all the way through the pipeline — and a famous failure explained.</strong></p><h3>What Happens to “How Many R’s in Strawberry?”</h3><p>Follow the exact sentence from the playground through every stage, and watch the mistake become inevitable rather than mysterious.</p><h3><strong>1. </strong>Your text is chopped into chunks</h3><p>The sentence is split into tokens: ‘How’, ‘ many’, ‘ r’, “‘s”, ‘ in’, ‘ straw’, ‘berry’, ‘?’. Notice that the fruit did not survive as one piece — and that the exact split differs between AI tools, since each has its own catalogue of chunks.</p><h3><strong>2. </strong>Chunks become a list of numbers</h3><p>Each chunk is swapped for its ID — something like 2118, 1523, 436, 885, 306, 15113, 15717, 30. From this moment on there is no sentence and no spelling, only this list of numbers moving through the arithmetic.</p><h3><strong>3. </strong>The model scores what should come next</h3><p>It has seen countless questions of this shape and knows the answer is a small number word. So ‘two’ and ‘three’ both score highly, plus a short explanation to follow. Nowhere in this process does anything actually inspect the letters of the word.</p><h3><strong>4. </strong>Chunks are chosen and glued back into text</h3><p>A chunk is picked, added, and the process repeats until a full reply exists. The numbers are converted back into chunks, the chunks joined into readable text, and you see: ‘There are two r’s in strawberry.’ Fluent, confident — and wrong, because the word has three.</p><h3><strong>5. </strong>The fix follows directly from the pipeline</h3><p>Ask again with the word spelled out — ‘count the r’s in s t r a w b e r r y’ — and each letter is now its own token, visible to the model. It answers three. Nothing about the AI improved; you simply handed it bricks the right size for the job.</p><h3>Final Words</h3><p>Text in, tokens through the middle, text out. In the playground you connected that whole journey: your words chopped into chunks, chunks swapped for numbers, the model scoring what fits next, chunks chosen one at a time, and everything stitched back into readable text at the end.</p><p>Keep the brick factory in mind, because it demystifies so much: token budgets explain forgetful long chats, token billing explains why bloated prompts cost more, and the invisibility of letters inside chunks explains the strawberry problem — along with the simple fix of spelling words out. Next comes the bigger question of what these models actually are, and how the piece-by-piece choosing turns into a flowing answer.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c306ca571bd7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How Does a Large Language Model Work?]]></title>
            <link>https://medium.com/@info.codeplu/how-does-a-large-language-model-work-1314b259f358?source=rss-5637d316b425------2</link>
            <guid isPermaLink="false">https://medium.com/p/1314b259f358</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[how-llms-work]]></category>
            <dc:creator><![CDATA[Code Plu]]></dc:creator>
            <pubDate>Sat, 03 Oct 2026 09:01:02 GMT</pubDate>
            <atom:updated>2026-10-03T09:01:02.134Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CGPurYTTiX_mtWK49OTxyA.png" /></figure><h3>Root Concept</h3><p>Inside an LLM <a href="https://codeplu.com/content/concept/how-llm-works"><strong>there are no rules and no stored facts </strong></a>— only billions of numbers tuned by training, plus a mechanism that lets every word weigh the words around it to settle what it means.</p><h3>What Is Actually Happening Inside an LLM?</h3><p>You already know an LLM writes text by continuing what came before. But how? There is no dictionary inside it, no grammar rulebook, no table of facts. Open one up and you find something almost disappointing: a colossal grid of numbers. Billions of them.</p><p>Those numbers are the dials that training tuned, and they hold everything the model ‘knows’ — not as sentences you could read, but as settings that make good continuations likely and bad ones unlikely. On top of that sits one genuinely clever mechanism, the thing that turned language models from clumsy into convincing: every word gets to look at the other words around it and adjust what it means in this particular sentence.</p><p>That is the idea you will practise above. The word ‘spring’ is in all three sentences, but a mattress, some seeds, and a mountain stream each drag it toward a completely different meaning. No dictionary required — the neighbours settle it. Once you see that, an LLM stops feeling like magic and starts looking like machinery.</p><h3>What Is Inside the Model, and How Does It Decide?</h3><h3><strong>1. </strong>What is actually inside the model?</h3><p>Numbers, and almost nothing else. A large model holds billions of adjustable values — think of them as volume dials — that were set during training and are now frozen. There is no list of rules a person wrote, and no folder of facts it consults. This surprises most beginners, who reasonably expect that something so knowledgeable must store its knowledge somewhere readable. It does not. If you opened the file you would see endless numbers with no labels, and no single dial that means ‘Paris is the capital of France’. The knowledge is spread across millions of them at once, which is exactly why nobody can point at a spot inside a model and read off what it believes.</p><h3><strong>2. </strong>How can numbers hold the meaning of a word?</h3><p>Each chunk of text gets turned into a long list of numbers, and the trick is that similar meanings end up with similar lists. ‘Cat’ lands near ‘kitten’ and not far from ‘dog’, while ‘Tuesday’ lands somewhere else entirely, close to ‘Wednesday’. Nobody sat down and assigned these positions — they emerged from training, because words that appear in similar company end up with similar numbers. This is why an LLM copes with a word you never taught it explicitly, and why it can answer a question phrased in a way it has never seen: what matters is where the meaning sits, not the exact letters you used.</p><h3><strong>3. </strong>How does the model know which word ‘it’ refers to?</h3><p>Read this: ‘The trophy would not fit into the suitcase because it was too big.’ You instantly know ‘it’ means the trophy. Now swap one word: ‘because it was too small’ — and ‘it’ becomes the suitcase. Nothing else changed. No grammar rule gets you there; you had to weigh the words against each other. This weighing is precisely what an LLM does, and it is called attention: at every position, the model looks across all the other words and decides how much each one matters for understanding this one. That mechanism is the reason modern models handle long, tangled sentences that older systems mangled.</p><h3><strong>4. </strong>Why does the same word mean different things?</h3><p>Because meaning is not stored in the word, it is assembled from the company the word keeps. ‘Spring’ beside ‘mattress’ is metal; beside ‘planted seeds’ it is a season; beside ‘drank from’ it is water. The model does not look up three dictionary entries and pick one — attention pulls the numbers for ‘spring’ toward whichever sense the neighbours support, and the result is a version of the word that only exists in this sentence. And this happens in layers, over and over, so each pass refines the picture: first the local clues, then the wider sentence, then the whole conversation.</p><h3><strong>5. </strong>So how does all this turn into an answer?</h3><p>After all that weighing, the model has a rich internal picture of everything you wrote. It uses that picture for one final job: producing a score for every chunk in its vocabulary, saying how well each would continue the text. High score for the sensible continuations, low for the nonsense. That single step is the whole output of the model — one set of scores. Turning those scores into flowing paragraphs takes one more idea, which is choosing a chunk and going around again. That belongs to the concept of how LLMs generate responses, waiting next in this track.</p><h3>Real World Example</h3><p><strong>Two nearly identical sentences, and a demonstration</strong><a href="https://codeplu.com/content/concept/how-llm-works"><strong> you can run in your own head.</strong></a></p><h3>How Does One Word Change What “It” Means?</h3><p>This is the clearest way to feel attention working. Read both sentences slowly and notice what your own mind does — because the model does the same thing with numbers.</p><h3><strong>1. </strong>The first sentence</h3><p>‘She poured water from the jug into the glass until it was full.’ Ask yourself what ‘it’ refers to. The glass, obviously — you did not hesitate. But notice that nothing in the word ‘it’ says so.</p><h3><strong>2. </strong>Change one word</h3><p>‘She poured water from the jug into the glass until it was empty.’ Now ‘it’ is the jug. One adjective flipped, and the meaning of a completely different word moved with it. This is why simple word-by-word rules were never enough for language.</p><h3><strong>3. </strong>What you just did, the model does with numbers</h3><p>You weighed ‘full’ and ‘empty’ against ‘jug’ and ‘glass’, using what you know about pouring. The model performs that same weighing as arithmetic: attention gives ‘it’ a strong link to whichever container the surrounding words support, and the numbers for ‘it’ shift accordingly.</p><h3><strong>4. </strong>It happens everywhere, not just for tricky pronouns</h3><p>Every word gets this treatment at once, in layer after layer. ‘Bank’ checks whether rivers or cheques are nearby. ‘Charge’ checks for batteries, bulls, or credit cards. By the final layer, each word carries a meaning shaped by everything else in your message.</p><h3><strong>5. </strong>Why this is the whole game</h3><p>Older systems that could not weigh context lost the thread after a few words, which is why they sounded robotic. Attention is what let models follow long sentences, track who did what to whom, and keep a conversation coherent — the single change that turned language models from a curiosity into something useful.</p><h3>Final Words</h3><p>Inside a large language model there is no rulebook and no filing cabinet — just billions of numbers tuned by training, and a mechanism that lets every word weigh its neighbours to settle what it means. In the playground you did that weighing yourself: a mattress, some seeds, and a mountain stream each dragged the same word toward a different meaning.</p><p>Hold on to two ideas. Meaning lives in the company a word keeps, not in the word itself — which is why context is so powerful and why vague prompts leave the model guessing. And one pass of the model produces only scores for what might come next. How those scores become a flowing, streaming answer is the story of the next concept.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1314b259f358" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Are Large Language Models?]]></title>
            <link>https://medium.com/@info.codeplu/what-are-large-language-models-f54883b18f93?source=rss-5637d316b425------2</link>
            <guid isPermaLink="false">https://medium.com/p/f54883b18f93</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[chatbots]]></category>
            <category><![CDATA[ai-assistant]]></category>
            <dc:creator><![CDATA[Code Plu]]></dc:creator>
            <pubDate>Fri, 02 Oct 2026 09:51:01 GMT</pubDate>
            <atom:updated>2026-10-02T09:51:01.507Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NbHdiPnDaB1rv18yyUIz4w.png" /></figure><h3>Root Concept</h3><p>A large language model<a href="https://codeplu.com/content/concept/what-are-llms"><strong> is not a search engine or a database</strong></a><strong> </strong>— it is a machine that learned the patterns of language from an enormous amount of text, and writes fresh text from those patterns.</p><h3>What Is a Large Language Model, in Plain Words?</h3><p>Picture someone who has read an entire library — millions of books, articles, conversations, manuals. Now take all the books away and ask them a question. They will answer fluently, in whatever style you like, on almost any subject. But they are answering from what stuck in their memory, not by looking anything up. Some of it will be sharp and precise. Some of it will be confidently half-remembered.</p><p>That is a large language model. The ‘language model’ part means it models how language fits together, so given some text it can continue it convincingly. The ‘large’ part refers to sheer scale: trained on an enormous amount of text, with billions of internal settings tuned during that training.</p><p>The single most useful thing to understand is what an LLM is not. It is not a search engine looking through pages, and not a database fetching a stored record. It writes fresh text every time, from patterns rather than from stored answers — which explains both its remarkable flexibility and its most annoying flaws. In the playground above you will build all three paths down from the same typed text, and follow the language model one step further than the others — to where it can be confidently wrong.</p><h3>How Should You Think About an LLM?</h3><h3><strong>1. </strong>What does the ‘language model’ part actually mean?</h3><p>A model, in this sense, is a working representation of how something behaves — and here the something is language. The model has absorbed how words follow words: that ‘thank you very’ is usually followed by ‘much’, that a question about a recipe expects ingredients and steps, that a complaint letter sounds different from a birthday card. Give it any piece of text and it continues that text in a way that fits. That is its one and only job. Everything an LLM appears to do — answering, translating, summarising, writing code — is that same continuation skill wearing different clothes, because all of those tasks can be expressed as ‘here is some text, what should come next?’</p><h3><strong>2. </strong>What makes it ‘large’?</h3><p>Two things, both enormous. First the reading material: modern models are trained on quantities of text no person could read in a thousand lifetimes — a large slice of the public internet, books, and reference works. Second the internal settings, those adjustable dials tuned during training, which now number in the billions. Scale is not just bragging rights, and this is the genuinely surprising part of the story: as these models grew, they started handling tasks nobody built them for — translating, writing working code, explaining jokes — simply because those patterns exist in language too. Nobody added a translation feature. It came along with the reading.</p><h3><strong>3. </strong>Why is an LLM not a search engine?</h3><p>A search engine keeps an index of pages and hands you the ones matching your words: it finds. A database stores exact records and returns them unchanged: it retrieves. An LLM does neither — it generates, composing text that never existed before, guided by patterns rather than by any stored copy. This is why an LLM has no links to give you by default, why two people asking the same question get differently worded answers, and why it cannot tell you today’s weather from its own knowledge. Many beginners assume the AI is quietly searching behind the scenes. It is not, unless the product has deliberately bolted a search tool onto it.</p><h3><strong>4. </strong>Why does it sometimes make things up?</h3><p>Once you accept there is no lookup, invented answers stop being mysterious. The model’s job is to produce text that fits — and text that fits can be entirely false. Asked for a book about an obscure topic, it may produce a plausible title by a plausible author that does not exist, because that shape of answer is exactly what should follow such a question. The name for this is hallucination, and the crucial detail is that it feels identical to a correct answer from the inside: same fluency, same confidence, no flicker of doubt. So the practical habit is simple. For anything that must be true — names, numbers, dates, quotes, laws, medical or legal specifics — verify elsewhere.</p><h3><strong>5. </strong>Why does it not know what happened last week?</h3><p>Training happens once, over a fixed pile of text collected up to a certain date, and afterwards the model’s internal settings are frozen. It does not quietly keep reading the news. That date is the training cutoff, and everything after it is simply absent — a model can be unaware of an election, a product, or a death that everyone else has known about for months. This also means it does not remember you between conversations unless the product deliberately saves notes. When an AI tool does answer with current information, something extra is happening: it has been given search or a document to read, and that fetched text is placed in front of it as part of the conversation.</p><h3>Real World Example</h3><p><a href="https://codeplu.com/content/concept/what-are-llms"><strong>One shopping problem, two tools, and a clear line between them.</strong></a></p><h3>Search Engine or LLM — Which Should You Ask?</h3><p>You need a birthday present for your aunt, who is turning seventy and loves gardening. Watch what each tool gives back, and where each one fails.</p><h3><strong>1. </strong>The search engine hands you the raw material</h3><p>Typing ‘gifts for gardeners’ returns pages of listicles, shop pages, and ads. Everything is real and current — someone genuinely wrote it and the prices genuinely exist — but you now have twenty tabs and half an hour of reading before you have decided anything.</p><h3><strong>2. </strong>The LLM hands you a shaped answer</h3><p>Ask an LLM instead, mentioning she is seventy, has arthritis in her hands, and already owns every tool going. It returns six tailored ideas — lightweight ergonomic tools, a kneeling stool, seeds for a scented windowsill — in the tone you asked for, no tabs required.</p><h3><strong>3. </strong>Where the LLM genuinely shines</h3><p>It weighed the details you gave it and shaped language around them. Ask it to turn the best idea into a warm gift-card message and it does that too. This is its home ground: understanding a messy request and producing well-formed text for it.</p><h3><strong>4. </strong>Where it quietly fails</h3><p>Now ask ‘how much does that kneeling stool cost at my local garden centre today?’ It has no lookup and no access to the shop, so it may still answer — with a confident, invented-looking figure. Nothing warns you, because a made-up price reads exactly like a real one.</p><h3><strong>5. </strong>The line that keeps you safe</h3><p>Use search for facts that live in the world and change: prices, stock, opening times, news. Use an LLM for shaping language: brainstorming, drafting, explaining, summarising, rewriting. And when you need both, use a tool that combines them — one that searches first and hands the results to the model to write up.</p><h3>Final Words</h3><p>A large language model is a machine that learned how language works by reading an unimaginable amount of it, and now writes fresh text from those patterns. In the playground you drew the distinction that matters most: search engines find, databases retrieve, language models generate. Every strength and every irritation follows from that third verb.</p><p>So expect fluency everywhere and truth only sometimes. Use an LLM for shaping language, verify anything that must be factual, and remember that its knowledge stops at a fixed date unless the tool fetches material for it. Next comes the machinery behind the curtain — what is actually inside the model that makes such convincing continuation possible.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f54883b18f93" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Happens Inside the Machine Learning Training Loop?]]></title>
            <link>https://medium.com/@info.codeplu/what-happens-inside-the-machine-learning-training-loop-5588ec226d74?source=rss-5637d316b425------2</link>
            <guid isPermaLink="false">https://medium.com/p/5588ec226d74</guid>
            <category><![CDATA[ml-explained]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[model-training]]></category>
            <category><![CDATA[ml-training-loop]]></category>
            <dc:creator><![CDATA[Code Plu]]></dc:creator>
            <pubDate>Thu, 01 Oct 2026 06:36:01 GMT</pubDate>
            <atom:updated>2026-10-01T06:36:01.808Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5wqZniffbRGrJh1SKeNKzw.png" /></figure><h3>Root Concept</h3><p>Inside training, a model improves through<strong> </strong><a href="https://codeplu.com/content/concept/machine-learning-model-training-loop"><strong>one repeating cycle: </strong></a>take an example, guess the answer, check the miss, adjust — millions of tiny corrections that quietly add up to skill.</p><h3>What Is the Training Loop, in Plain Words?</h3><p>Think about learning to throw darts. Your first throw lands nowhere near the bullseye. But you see exactly where it landed, so your next throw corrects a little — too far left becomes a little more right. Throw after throw, each miss teaches your arm a tiny lesson, and an hour later you are hitting close to center. Nobody explained dart physics to you. The loop of try, see the miss, and adjust did all the work.</p><p>That loop is precisely how a machine learning model learns. Earlier concepts showed you the big picture — collect data, prepare it, train, evaluate, deploy. This tutorial zooms into the step everyone treats as a black box: training itself. Inside it, one small cycle repeats millions of times: take a practice example, guess the answer, check how far off the guess was, and adjust a little. Then the next example, and around again.</p><p>In the playground above you will build this cycle yourself — all four beats, plus the edge that makes it a loop rather than a line. Once you see it, ‘the model is training’ stops being magic and becomes something wonderfully ordinary: practice.</p><h3>How Does One Trip Around the Loop Work?</h3><h3><strong>1. </strong>What are the four beats of the training loop?</h3><p>Every trip around the loop has the same four beats. First, take an example — one practice question whose correct answer is already known, like a photo already labeled ‘cat’. Second, guess: the model produces its answer, and early on these guesses are almost comically wrong. Third, check the miss: the guess is compared with the known answer to see how far off it was. Fourth, adjust: the model’s internal settings get nudged slightly in the direction that would have made the guess better. Then the loop grabs the next example and goes around again. One trip teaches the model almost nothing — the power is in repetition, exactly like one dart throw versus a thousand.</p><h3><strong>2. </strong>How does the model know how wrong it was?</h3><p>The miss is not a feeling — it is a number. If the model guessed a house costs 500,000 and the real sale price was 300,000, the miss is 200,000. If it guessed ‘dog’ with high confidence and the answer was ‘cat’, the miss is large; a hesitant wrong guess counts as a smaller miss. Engineers call this number the error — the measured size of the miss. Here is the misconception worth dropping now: many beginners imagine the model ‘understands’ it made a mistake, maybe even feels bad about it. It does not. The entire drama of learning boils down to one goal a machine can follow: make this number smaller on the next trip around.</p><h3><strong>3. </strong>How does adjusting actually work?</h3><p>Inside every model live thousands or even billions of adjustable settings — think of them as tiny volume dials that together decide the model’s guess. Adjusting means turning each dial a tiny bit in the direction that would have shrunk the miss. Why tiny? Imagine correcting your dart throw by hurling the next dart twice as hard the other way — you would overshoot wildly and never settle. Small, patient corrections let the dials drift steadily toward values that work. A common beginner picture is one big ‘aha’ fix, like a student suddenly getting it. Real training has no aha moment — just millions of nudges so small that each one is invisible, adding up to skill.</p><h3><strong>4. </strong>When does the loop stop?</h3><p>The loop does not run forever. The model works through all its examples — one full pass is called an epoch, simply a round of practice — and then starts another round. Watchers track the misses: early rounds shrink them dramatically, later rounds barely move them. When the misses stop improving, training stops. And stopping matters more than beginners expect, because a loop that runs far too long starts memorizing its practice questions instead of learning the pattern — like a student who reheats the same past papers until they know the answers by heart but not the subject. That failure has a name, overfitting, and it is exactly what the final exam on unseen data — evaluation — is designed to catch.</p><h3>Real World Example</h3><p><a href="https://codeplu.com/content/concept/machine-learning-model-training-loop"><strong>Follow one model around the loop — from absurd guesses to trustworthy estimates.</strong></a></p><h3>How Does a Food App Learn to Predict Delivery Time?</h3><p>A food delivery app wants to show ‘your order <a href="https://codeplu.com/content/concept/machine-learning-model-training-loop"><strong>arrives in 32 minutes’</strong></a> — and be right. Its model learns that skill by looping over thousands of past orders whose real delivery times are known.</p><h3><strong>1. </strong>Take an example: one past order</h3><p>The loop picks a solved practice question from history: an order placed on a rainy Friday evening, from a restaurant <a href="https://codeplu.com/content/concept/machine-learning-model-training-loop"><strong>3 kilometres away,</strong></a> that actually took 41 minutes door to door. The real answer — 41 — is the truth the guess will be checked against.</p><h3><strong>2. </strong>Guess: the untrained model embarrasses itself</h3><p>The fresh model’s dials start at random, so its first prediction is nonsense — it guesses 7 minutes. That is fine. Every learner starts terrible; what matters is what happens with the miss.</p><h3><strong>3. </strong>Check the miss: 34 minutes off</h3><p>The guess is compared with the truth: 41 minus 7 means the model was 34 minutes too optimistic. That single number — the size of the miss — is all the feedback the model gets. No lecture, no explanation. Just: you were this far off, in this direction.</p><h3><strong>4. </strong>Adjust: every dial gets a tiny nudge</h3><p>The settings that pushed the guess too low — the ones underweighting rain, distance, and Friday-evening traffic — each get turned slightly the other way. Not fixed in one stroke: nudged. The next rainy-Friday guess will be a little less wrong.</p><h3><strong>5. </strong>Around again — a million orders later</h3><p>The loop grabs the next order, and the next, looping through months of history round after round. The misses shrink from 34 minutes to 10 to 3. When they stop improving, training ends — and the model faces its real exam: predicting orders it has never seen. That story continues in the evaluation step of the ML workflow.</p><h3>Final Words</h3><p><strong>You have opened the black box of training and found something refreshingly simple inside: a four-beat cycle — take an example, guess, check the miss, adjust — spinning millions of times. No understanding, no aha moment, just patient practice where every miss nudges thousands of tiny dials a little closer to right. You built that cycle in the playground, including the edge that makes it a loop.</strong></p><p><strong>Carry two ideas forward: skill comes from repetition plus feedback, and more repetition is not always better — memorization lurks past the point where misses stop shrinking. That is why the training loop always hands its graduate over to an exam on unseen questions. How that exam fits into the full journey is the story of the machine learning workflow.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5588ec226d74" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Is the Machine Learning Workflow?]]></title>
            <link>https://medium.com/@info.codeplu/what-is-the-machine-learning-workflow-24005864f5d1?source=rss-5637d316b425------2</link>
            <guid isPermaLink="false">https://medium.com/p/24005864f5d1</guid>
            <category><![CDATA[machine-learning-process]]></category>
            <category><![CDATA[machine-learning-pipeline]]></category>
            <category><![CDATA[machine-learning-workflow]]></category>
            <category><![CDATA[model-training]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Code Plu]]></dc:creator>
            <pubDate>Wed, 30 Sep 2026 08:31:01 GMT</pubDate>
            <atom:updated>2026-09-30T08:31:01.664Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QNxms5Z7041TyOqNwPHxrQ.png" /></figure><h3>Root Concept</h3><p><strong>Every machine learning system is built in </strong><a href="https://codeplu.com/content/concept/ml-work-flow"><strong>five disciplined steps — collect, prepare, train, evaluate, deploy — with a quality gate in the middle:</strong></a><strong> fail the evaluation, and you loop back to improve and retrain.</strong></p><h3>What Is the Machine Learning Workflow, in Plain Words?</h3><p>Imagine opening a small bakery. You would not invent a cake and start selling it the same afternoon. You would gather good ingredients, prepare them properly, bake a trial batch, taste it — and only if it tastes right would you put it in the shop window. If the trial batch tastes wrong, you would not sell it anyway; you would fix the recipe and bake again.</p><p>Building a machine learning system works exactly like that. There are five steps: collect the data, prepare it, train the model, evaluate it, and finally deploy it into the real world. And there is one rule that separates professionals from beginners: evaluation is a quality gate, not a formality. A model that fails its exam never goes live — the work loops back, the data gets improved, and the model is trained again.</p><p>In the playground above you will build this whole workflow yourself, including the loop-back edge that most beginners forget. By the end, you will know not just the five steps but also why real projects almost never walk through them in a straight line.</p><h3>How Is a Machine Learning System Actually Built?</h3><h3><strong>1. </strong>Where does every machine learning project start?</h3><p>Every project starts with collecting data — real examples of the thing you want the model to learn. A model that should spot spam needs real emails; one that should predict house prices needs real past sales. The examples come from databases, websites, sensors, or simple records people have kept. Here is the misconception to drop early: beginners often think a clever algorithm is what makes a project succeed. In practice, the amount and quality of the data matter far more. A simple model fed with plenty of good examples beats a fancy model fed with poor ones — just as a skilled baker with fresh ingredients beats a genius baker with spoiled flour.</p><h3><strong>2. </strong>Why does the data need cleaning before training?</h3><p>Raw data from the real world is always messy — missing entries, duplicates, typos, dates written five different ways. Preparation means cleaning all of this up and converting it into the tidy numbers a model can digest: filling or removing gaps, dropping duplicates, making every record follow the same format. It is the least glamorous step and, surprisingly, the biggest one — ask people who build these systems for a living and most will tell you data work eats more of their time than everything else combined. The golden rule explains why: garbage in, garbage out. A model trained on messy data will faithfully learn the mess.</p><h3><strong>3. </strong>What actually happens during training?</h3><p>Training is where the learning happens, and the rhythm is simple: guess, check, adjust. The model takes an example, guesses the answer, compares its guess with the real answer, and nudges its internal settings to do slightly better next time. Repeat this thousands or millions of times, and the guesses stop being random and start being skillful. Note what the humans are doing during this step: mostly waiting and watching. Nobody types in rules. The patterns come out of the examples — which is exactly why the two steps before this one, collecting and cleaning, decide most of the final quality.</p><h3><strong>4. </strong>Why is evaluation the most important checkpoint?</h3><p>After training, the model takes a final exam — and the exam questions must be examples it has never seen before, held back on purpose from the start. Why so strict? Because a model can score perfectly on its own training examples by simply memorizing them, the way a student who stole the answer key aces the practice test and fails the real one. Testing on unseen data is the only way to tell whether the model truly learned the pattern or just memorized answers. Many beginners believe a high score during training means the model is good. It means nothing of the sort — only the unseen-data exam counts.</p><h3><strong>5. </strong>What happens when the model fails the exam?</h3><p>Here is the step that turns the straight line into a loop: when a model fails evaluation — and first versions usually do — it does not go live, and the team does not give up. The work loops back. Usually the cure is better data: collect more examples, fix gaps, add the situations the model got wrong. Then prepare, train, and take the exam again. Real machine learning is this cycle running until the gate opens. Only then comes deployment: the model goes live and starts predicting on brand-new cases. Even then the workflow never fully ends — the world keeps changing, so teams keep watching the model’s accuracy and loop back whenever it starts to slip.</p><h3>Real World Example</h3><p><a href="https://codeplu.com/content/concept/ml-work-flow"><strong>One team, five steps — and one failed exam that saved them from disaster.</strong></a></p><h3>How Does a Bank Build a Loan-Approval Model?</h3><p><a href="https://codeplu.com/content/concept/ml-work-flow"><strong>Follow a bank’s team building a system that predicts whether a loan application is safe to approve. Notice where the straight line becomes a loop.</strong></a></p><h3><strong>1. </strong>Collect: dig out the bank’s own history</h3><p>The team gathers thousands of past loan applications — the applicant’s income, their credit history, and most importantly, whether they actually paid the loan back. Every past customer becomes one solved example for the model to learn from.</p><h3><strong>2. </strong>Prepare: clean up years of messy records</h3><p>The records span years and systems: salaries in different currencies, missing entries, duplicate applications. The team spends weeks cleaning and structuring — unglamorous work, but they know a model trained on messy records learns messy rules.</p><h3><strong>3. </strong>Train: the model studies who repaid and who didn’t</h3><p>The clean examples flow into training. Guess by guess, the model tunes itself until it can spot the quiet patterns that separate reliable borrowers from risky ones — combinations no employee ever wrote down as a rule.</p><h3><strong>4. </strong>Evaluate: the exam catches a dangerous flaw</h3><p>On held-back applications the model has never seen, the truth comes out: it approves far too many risky loans from recent years, because most of its training examples came from older, calmer economic times. The exam fails. The model does not go live.</p><h3><strong>5. </strong>Loop back, retrain, pass — then deploy</h3><p>The team loops back: they add thousands of recent applications, rebalance the data, and retrain. This time the exam scores hold up. Only now does the model go live — scoring new applications in seconds, while the team keeps watching its accuracy month after month.</p><h3>Final Words</h3><p><strong>You now have the full blueprint in your head: collect, prepare, train, evaluate, deploy — plus the edge that beginners forget and professionals live by, the loop back from a failed exam to better data. In the playground you built it exactly as real teams run it: a straight line with a quality gate, and a circle whenever the gate stays shut.</strong></p><p><strong>Keep two lessons above all: the data steps decide most of the final quality, and no model goes live without passing an exam on examples it has never seen. With the workflow mastered, you are ready to meet the different kinds of learning a model can do — the next concept in this track.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=24005864f5d1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Are the Types of Machine Learning?]]></title>
            <link>https://medium.com/@info.codeplu/what-are-the-types-of-machine-learning-99bc9eba04f1?source=rss-5637d316b425------2</link>
            <guid isPermaLink="false">https://medium.com/p/99bc9eba04f1</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[unsupervised-learning]]></category>
            <category><![CDATA[reinforcement-learning]]></category>
            <category><![CDATA[supervised-learning]]></category>
            <category><![CDATA[deep-learning]]></category>
            <dc:creator><![CDATA[Code Plu]]></dc:creator>
            <pubDate>Tue, 29 Sep 2026 08:11:01 GMT</pubDate>
            <atom:updated>2026-09-29T08:11:01.696Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*oFy4G9fJhQfQUTrt8_QSBQ.png" /></figure><h3>Root Concept</h3><p><strong>There are three ways a machine can learn: from examples </strong><a href="https://codeplu.com/content/concept/types-of-ml"><strong>where the answer is given, from data where no answers exist, </strong></a><strong>or by trying actions and being rewarded — and the kind of data you have decides which is possible.</strong></p><h3>How Many Ways Can a Machine Learn?</h3><p>Think about three ways you have learned things. At school, you did practice questions and a teacher marked them right or wrong — you learned from examples with answers attached. Sorting a drawer of loose photographs into piles, nobody told you what the piles should be; you noticed which pictures belonged together. And learning to ride a bicycle, nobody could mark your answers at all — you tried, wobbled, fell, adjusted, and gradually got further before falling.</p><p>Machines learn in exactly those three ways, and each has a name: supervised learning for the marked practice questions, unsupervised learning for the photo piles, and reinforcement learning for the bicycle. Same underlying idea of improving through practice, but three quite different situations.</p><p>Which one you use is not really a preference — it is decided by what you have. If your examples come with answers attached, one path opens. If they do not, another. And if the machine can act and receive a score for the result, a third. In the playground above you will build machine learning down into its three kinds, each one settled by what came with the data.</p><h3>What Are the Three Kinds of Learning?</h3><h3><strong>1. </strong>What is supervised learning, in plain words?</h3><p>It is learning from worked examples: every example comes with the correct answer attached, so the machine can check itself and improve. Photos labelled cat or dog, past loans marked repaid or defaulted, houses with the price they actually sold for. This is the most common kind in real products because it is the most straightforward — you know what right looks like, so you can measure how wrong the model is. Its catch is the labels. Somebody, usually a person, had to mark all those examples, and that work is slow and expensive. Many promising projects stall not for lack of data but for lack of labelled data.</p><h3><strong>2. </strong>What can unsupervised learning do without any answers?</h3><p>It finds structure nobody pointed out. Give it your customers with no labels and it can group the ones who behave alike; give it transactions and it can flag the handful that look unlike everything else. Notice what it cannot do: it cannot tell you what the groups mean. It might split your customers into four clusters, and interpreting them — ‘these are bargain hunters, these buy gifts’ — is a human job. That is the trade. You need no labels, which is liberating, but you get patterns rather than answers, and you have to decide whether they are useful.</p><h3><strong>3. </strong>How does reinforcement learning work?</h3><p>By trying things and being scored. There is no dataset of right answers, only an environment the machine can act in and a signal telling it how well that went — points for progress, penalties for failure. Over enormous numbers of attempts it learns which actions tend to lead to good scores. This is how systems learn to play games and how robots learn physical movement, and its distinctive requirement is that failure must be cheap. A machine can fall over ten million times in a simulator; it cannot crash ten million real cars. That is why so much of this work happens in simulation first.</p><h3><strong>4. </strong>How do you tell which type a problem needs?</h3><p>Ask what you actually hold. Do you have examples with the answer attached? Then supervised learning, and it is usually the simplest route. Do you have data but no answers, and you are hoping to discover structure rather than predict something specific? Then unsupervised. Can the machine act repeatedly, cheaply, and get scored on the result? Then reinforcement. Beginners often pick the type that sounds most impressive, which is backwards — reinforcement learning is the most glamorous and the least applicable to ordinary business problems, while unglamorous supervised learning quietly runs most of the AI you use every day.</p><h3><strong>5. </strong>Do real systems use only one type?</h3><p>Often not, and the modern AI you use daily is a good example. Large language models begin with a huge amount of self-supervised training, where the answer is hidden inside the data itself — cover up the next word and the text supplies the correct answer, so no human labelling is needed. Then people rate the answers, and a reinforcement-style step nudges the model toward the replies humans preferred. So one product can involve several kinds of learning stacked in sequence. The three types are best understood as tools rather than tribes, and knowing which is which tells you what a system needed in order to exist.</p><h3>Real World Example</h3><p><a href="https://codeplu.com/content/concept/types-of-ml"><strong>Three problems in the same shop, each needing a different kind of learning.</strong></a></p><h3>How Would One Supermarket Use All Three?</h3><p><a href="https://codeplu.com/content/concept/types-of-ml"><strong>A supermarket chain has three quite different questions. Watch how the data available decides the approach in each case.</strong></a></p><h3><strong>1. </strong>Problem one: how much bread will sell tomorrow?</h3><p>They hold years of daily sales with the actual number sold each day — examples with the answers attached. That makes this supervised learning: the model learns from history and its predictions can be checked against what really happened, so accuracy is measurable from day one.</p><h3><strong>2. </strong>Problem two: what kinds of shoppers do we have?</h3><p>Nobody has ever defined the customer types, so there are no labels to learn from. Unsupervised learning groups shoppers whose baskets look alike, and it produces five clusters. The machine does not name them — the marketing team looks and recognises ‘big weekly shops’, ‘lunchtime grabs’, and so on.</p><h3><strong>3. </strong>Problem three: how should the warehouse robots move?</h3><p>There is no dataset of correct movements, but a robot can attempt a route and be scored on time taken and collisions avoided. That is reinforcement learning, and it is trained in simulation first — because a real robot cannot afford millions of failed attempts.</p><h3><strong>4. </strong>Why the choice was never really a choice</h3><p>Each type was forced by the situation. Labelled history made the first one supervised. The absence of labels made the second unsupervised. The ability to try and be scored made the third reinforcement. Asking ‘which type should we use?’ before looking at the data gets the order backwards.</p><h3><strong>5. </strong>What ties them together</h3><p>All three still follow the same underlying loop you already know: try, check how it went, adjust, repeat. What differs is only where the feedback comes from — a labelled answer, the shape of the data itself, or a score from the environment.</p><h3>Final Words</h3><p><strong>Machines learn in three ways: from marked practice questions, from unlabelled data where they must find their own groupings, and by trying actions and being scored. In the playground you matched three real jobs to the learning each one made possible — labelled loans, unlabelled customers, and a robot that can try and be rewarded.</strong></p><p><strong>The habit to keep is looking at your data before choosing an approach, because the data decides and prestige misleads. And remember that all three still run the same underlying loop of try, check, adjust — only the source of the feedback changes.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=99bc9eba04f1" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>