<?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 Sooriyakumar on Medium]]></title>
        <description><![CDATA[Stories by Sooriyakumar on Medium]]></description>
        <link>https://medium.com/@sisooriyakumar?source=rss-668662b96dfd------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*2rGyp9jQAfV6wOVZ.jpg</url>
            <title>Stories by Sooriyakumar on Medium</title>
            <link>https://medium.com/@sisooriyakumar?source=rss-668662b96dfd------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 12:23:20 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@sisooriyakumar/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[Your Job Is Not Your Tasks: How to Survive the AI Era]]></title>
            <link>https://medium.com/@sisooriyakumar/your-job-is-not-your-tasks-how-to-survive-the-ai-era-ce119bdfc5d9?source=rss-668662b96dfd------2</link>
            <guid isPermaLink="false">https://medium.com/p/ce119bdfc5d9</guid>
            <category><![CDATA[life-lessons]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[jobs]]></category>
            <category><![CDATA[jenson-huang]]></category>
            <category><![CDATA[purpose]]></category>
            <dc:creator><![CDATA[Sooriyakumar]]></dc:creator>
            <pubDate>Tue, 17 Nov 2026 04:43:03 GMT</pubDate>
            <atom:updated>2026-11-17T04:43:03.151Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jc39DkzjQ14GyxDgrW6KcA.png" /></figure><h3>Two people in one kitchen</h3><p>Picture two people working in the same kitchen. The first has a one-line job: chop vegetables. The second is a cook, and the cook’s job is to send out a meal that people enjoy.</p><p>Now bring in a machine that chops faster and cleaner, and never gets tired. The first person has a problem. The second person just got a better knife.</p><p>That picture comes from a <a href="https://youtube.com/shorts/m4ZV9g4GU5s">46-second clip of Nvidia CEO Jensen Huang</a> on how not to lose your job to AI. The caption sums it up in one line: “If your job is the task, you’re the one at risk.”</p><p>I think this is one of the most useful ideas in the whole AI and Jobs debate. It is also easy to oversell. So this post does two things: it explains the idea with examples, and it tells you where the idea stops working.</p><h3>A task is what you do. A purpose is why anyone pays for it.</h3><p>AI is very good at tasks. Give it a clear input and a clear expected output, and it will produce that output quickly, cheaply and at any hour. Summarize this document. Write this function. Draft this email. Classify this image.</p><p>A purpose is different. It is the result someone else depends on: a patient gets the right diagnosis, a customer’s problem goes away, a business makes a sound decision. Serving a purpose means deciding which tasks are worth doing, noticing when the output is wrong, and answering for the result.</p><p>Here is a quick test. Describe your job in one sentence.</p><ul><li>“I write reports.” That is a task. It is a verb and an object.</li><li>“I help the management team decide with the right numbers in front of them.” That is a purpose. It names who benefits and how.</li></ul><p>If your sentence sounds like the first one, AI is competing with you. If it sounds like the second, AI is working for you.</p><h3>The radiologist who was supposed to disappear</h3><p>In 2016, AI pioneer Geoffrey Hinton said AI would outperform radiologists within five to ten years, and suggested we stop training new ones. Reading medical images is exactly the kind of task AI is good at, so the prediction sounded reasonable.</p><p>It did not happen. According to <a href="https://dc.fortune.com/2026/05/04/godfather-of-ai-geoffrey-hinton-radiologists-future-of-work-tech-ai-job-anxiety/">Fortune</a>, the number of active radiologists in the US grew by about 10% over the past decade, and the average salary reached $571,000 in 2025.</p><p>Huang’s explanation is the task-versus-purpose idea. Reading the scan is the task. Finding the disease and helping the patient is the purpose. AI took over part of the task, and the radiologist was still needed for the purpose.</p><p>The short clip tells this story a little too neatly, so two corrections are worth making.</p><ul><li><strong>AI assists radiologists. It has not taken over the reading.</strong> A licensed physician still does the final read. US public insurers only pay for a study when that happens, and nobody has settled who is liable when AI misses something.</li><li><strong>Purpose was not the only reason.</strong> Case loads rose 25% between 2018 and early 2025, driven partly by an ageing population. Cheaper, faster imaging also meant more scans. Demand grew faster than automation.</li></ul><p>So the accurate lesson is this: radiologists survived because their purpose mattered, their judgement was legally and practically required, and demand for that purpose kept growing.</p><h3>The same split in other jobs</h3><p>The radiologist is one case. The same split shows up almost everywhere.</p><p><strong>Software engineer.</strong> I work in software, so this one is close to home. Writing code is the task, and AI now does a lot of it well. The purpose is a system that solves the customer’s problem and keeps working in production. Someone still has to understand what the customer actually needs, choose the trade-offs, review what the AI wrote, and take the call when it breaks at 2 a.m. An engineer who only turns tickets into code is exposed. An engineer who owns the outcome has just become faster.</p><p><strong>Customer support agent.</strong> Answering “where is my order?” is a task, and a chatbot handles it. The purpose is making the customer’s problem go away. That includes the refund that does not fit the policy and the angry customer who needs someone with authority to say “we got this wrong, and here is what I’ll do.”</p><p><strong>Accountant.</strong> Entering transactions and reconciling accounts are tasks. The purpose is helping an owner make sound money decisions and stay on the right side of the law. Software can produce the numbers. The owner still wants a person to say “don’t sign that lease yet.”</p><p><strong>Teacher.</strong> Delivering content and marking quizzes are tasks, and AI tutors do both. The purpose is making sure a particular child actually learns. That means noticing that a student has gone quiet and working out why.</p><p>In every case, AI removes the routine part and leaves the part that needs judgement, trust and responsibility.</p><h3>Where this idea stops working</h3><p>“Find your purpose and you are safe” is a comforting message. It is not the whole truth, and I would rather you hear the limits from me.</p><p><strong>Purpose does not guarantee headcount.</strong> If AI lets one person do the work of three and demand stays flat, two people are not needed, however purposeful they are. Radiology grew because demand for scans kept rising. Not every field has that.</p><p><strong>Some jobs really are the task.</strong> Huang <a href="https://www.techradar.com/pro/you-have-to-separate-the-task-from-the-purpose-of-the-job-nvidia-ceo-jensen-huang-reflects-on-early-ai-fears-around-job-displacement-and-why-we-shouldnt-be-worried">has acknowledged this himself</a>: roles made up mostly of repetitive work are exposed. If that describes your role, renaming it will not help. You need to move towards work where you own a result.</p><p><strong>Beginners learn through tasks.</strong> Junior people build judgement by doing the routine work that AI now does. If we hand all of it to AI, we need another way to train the next generation. I don’t think anyone has fully solved this yet.</p><p><strong>AI is getting better at goals too.</strong> AI agents can already plan and carry out multi-step work. What they cannot do is be accountable. A person signs the report, faces the client and carries the consequences. That accountability is the durable part of a job.</p><h3>Five things you can do this month</h3><ol><li><strong>Write your purpose sentence.</strong> Who depends on your work, and what do they get from it? If you can’t answer, ask your manager or your customer. Their answer may surprise you.</li><li><strong>List your weekly tasks and mark the ones AI can do.</strong> Then hand those over on purpose. Don’t wait for someone else to do it for you.</li><li><strong>Spend the saved time closer to the person you serve.</strong> Talk to the customer, the patient, the student or the user. AI has the least context there, and you have the most.</li><li><strong>Own the result.</strong> Be the person who checks the AI’s work and answers for it. Being able to say “I reviewed this and I stand behind it” is worth more every year.</li><li><strong>Learn the domain as well as the tool.</strong> Tools change every few months. Knowing how hospitals, banks or supply chains actually work lasts much longer.</li></ol><h3>Back to the kitchen</h3><p>The chopping machine is already here, and it gets better every month. You can’t stop that, and arguing with it is a waste of energy.</p><p>What you can decide is which person in the kitchen you are. The one who chops is competing with the machine. The one who cares whether the guest enjoyed the meal is using it.</p><p>So ask yourself the question the clip asks. Is my job a task, or is it a purpose? If the honest answer is “a task”, today is a good day to start changing it.</p><h3>Sources</h3><ul><li><a href="https://youtube.com/shorts/m4ZV9g4GU5s">Jensen Huang explains how to not lose your job to AI</a>, Hyperbolic AI, YouTube Shorts</li><li><a href="https://dc.fortune.com/2026/05/04/godfather-of-ai-geoffrey-hinton-radiologists-future-of-work-tech-ai-job-anxiety/">A decade after the “Godfather of AI” said radiologists were obsolete</a>, Fortune</li><li><a href="https://www.techradar.com/pro/you-have-to-separate-the-task-from-the-purpose-of-the-job-nvidia-ceo-jensen-huang-reflects-on-early-ai-fears-around-job-displacement-and-why-we-shouldnt-be-worried">“You have to separate the task from the purpose of the job”</a>, TechRadar</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ce119bdfc5d9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[LLM Sampling Parameters Explained: Temperature, Top-k and Top-p]]></title>
            <link>https://medium.com/@sisooriyakumar/llm-sampling-parameters-explained-temperature-top-k-and-top-p-a07550e3f6b3?source=rss-668662b96dfd------2</link>
            <guid isPermaLink="false">https://medium.com/p/a07550e3f6b3</guid>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Sooriyakumar]]></dc:creator>
            <pubDate>Mon, 16 Nov 2026 12:26:16 GMT</pubDate>
            <atom:updated>2026-11-16T12:26:16.573Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="Infographic illustrating LLM sampling parameters, including Temperature, Top-k, and Top-p, using probability distribution bar charts, a temperature dial, probability filtering, and dice to demonstrate how large language models select the next word during AI text generation." src="https://cdn-images-1.medium.com/max/1024/1*cOuz30Ohyr3MATk-b9vaJw.png" /></figure><p><em>What temperature, top-k and top-p do to an LLM’s next-word choice, with worked numbers, code and the settings to start from.</em></p><p>Ask an LLM the same question twice and you often get two different answers. That is not a bug. It is the result of a few settings called <strong>sampling parameters</strong>, and the three you will meet most often are <strong>temperature</strong>, <strong>top-k</strong> and <strong>top-p</strong>.</p><p>This post explains what each one does, with one small example carried all the way through, and then shows where you can actually set them: in code through an API or SDK, and in apps like Claude and ChatGPT.</p><h3>What is a sampling parameter in an LLM?</h3><p>A sampling parameter is a setting that controls how a large language model chooses its next token from the list of candidates it has scored.</p><p>To see where that fits, recall how an LLM writes. It never produces a whole sentence at once. It produces one token (a word or a piece of a word), adds it to the text, and repeats. At every step it does two things:</p><ol><li><strong>Score.</strong> The model gives every token in its vocabulary a probability of coming next.</li><li><strong>Pick.</strong> One token is chosen from those probabilities. This step is called <strong>sampling</strong>, or <strong>decoding</strong>.</li></ol><p>The model’s training decides the scores. Sampling parameters decide the pick. They do not make the model smarter or change what it knows. They only change how adventurous the choice is.</p><h3>The example we will use</h3><p>Suppose you type <em>“I like to eat banana”</em> and stop. To keep the numbers readable, pretend the model’s vocabulary has only seven tokens. After the scoring step, the probabilities look like this:</p><figure><img alt="LLM next-token probability distribution showing candidate words such as bread, splits, every, and pudding with their corresponding probabilities, illustrating how a large language model predicts the next word during AI text generation." src="https://cdn-images-1.medium.com/max/1024/1*3lVLGqlZ66scAZYKHjxZxg.png" /></figure><p><em>Bread</em> is the favourite, a few other tokens are reasonable, and <em>carburettor</em> is nonsense. Every sampling parameter in this post is a different way of answering one question: <strong>given this table, which token do we pick?</strong></p><h3>Why do LLMs need sampling parameters?</h3><p>The obvious answer to “which token do we pick?” is “the most likely one”. That approach has a name, <strong>greedy decoding</strong>, and it has three problems.</p><ul><li><strong>It is repetitive.</strong> The same prompt always gives the same answer. Ask for ten product names and you get the same ten every time.</li><li><strong>It is dull.</strong> The most likely next word is usually the safest one. Text built only from safe choices reads flat, and long outputs can get stuck repeating a phrase.</li><li><strong>It throws away good options.</strong> <em>Splits</em> at 21% is a perfectly good continuation. Greedy decoding will never choose it.</li></ul><p>The opposite approach is to roll a weighted dice over the whole table. That brings variety, but it also means <em>carburettor</em> gets picked once in a while. A real vocabulary has around 100,000 tokens, and thousands of unlikely ones together hold enough probability to derail a sentence.</p><p>So there is a trade-off:</p><ul><li>Too little randomness -&gt; Repetitive, predictable, flat</li><li>Too much randomness -&gt; Incoherent, off-topic, error-prone</li></ul><p>Sampling parameters let you choose where to sit between those two ends. The right spot depends on the job. Extracting an invoice number wants almost no randomness. Brainstorming blog titles wants plenty.</p><p>The three parameters attack the problem in two different ways:</p><ul><li><strong>Temperature</strong> reshapes the probabilities.</li><li><strong>Top-k</strong> and <strong>top-p</strong> remove unlikely tokens before the pick.</li></ul><h3>The three main sampling parameters: temperature, top-k and top-p</h3><h3>Temperature: how random should the pick be?</h3><p><strong>Temperature</strong> is a number that makes the probability table sharper or flatter before a token is picked.</p><p>Under the hood, the model’s raw scores (called logits) are divided by the temperature before they are turned into probabilities:</p><figure><img alt="Softmax probability formula used in large language models to convert temperature-scaled logits into a probability distribution for next-token prediction, illustrating how temperature influences AI text generation." src="https://cdn-images-1.medium.com/max/1024/1*A2hYtqIaQZ9BKPnAt_G5Sw.png" /></figure><p>Here z is a token’s raw score and T is the temperature. You do not need the formula to use it. Three cases cover everything:</p><ul><li><strong>T below 1</strong> stretches the gaps between scores. The favourite gets stronger.</li><li><strong>T equal to 1</strong> leaves the probabilities exactly as the model produced them.</li><li><strong>T above 1</strong> shrinks the gaps. The table flattens and the underdogs get a real chance.</li></ul><p>Here is our banana example at three temperatures:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*z8V9JcE6jmS8YRiNNzQT_A.png" /></figure><p>At 0.5, <em>bread</em> wins about half the time and <em>chips</em> almost never appears. At 2.0, the six sensible tokens are close to even, so the output varies a lot from run to run.</p><p>Two details are worth noticing:</p><ul><li><strong>Temperature never removes a token.</strong> It only changes the odds. <em>Carburettor</em> stays near zero here, but in a real vocabulary of 100,000 tokens, a high temperature lifts thousands of weak tokens a little. Together they add up, which is why very high temperatures produce word salad.</li><li><strong>Temperature 0 means “always take the top token”</strong>, which is greedy decoding again. It makes output far more repeatable, but hosted models can still vary slightly between runs.</li></ul><p><strong>When to use what:</strong> low values (0 to 0.3) for extraction, classification and code. Middle values (0.7 to 1.0) for everyday writing. Higher values for brainstorming and creative work.</p><h3>Top-k: keep only the k most likely tokens</h3><p><strong>Top-k sampling</strong> keeps the k tokens with the highest probability and removes everything else before the pick. The survivors are then rescaled so they add up to 100% again.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0CREi5A9N6orBLPn6L7v7w.png" /></figure><p><em>Carburettor</em> can no longer be picked, however high the temperature goes. That is the point of top-k: it cuts off the long tail of junk tokens.</p><p>Its weakness is that k is a fixed number, and the model’s confidence is not fixed:</p><ul><li>When the model is very sure (one token at 95%), k = 40 still lets 39 weak tokens stay in the running.</li><li>When the model sees 50 equally good options, k = 3 throws most of them away.</li></ul><p>Two values are worth remembering. <strong>k = 1</strong> is greedy decoding. Values in the range of <strong>20 to 50</strong> are common defaults in tools that expose the setting.</p><h3>Top-p (nucleus sampling): keep the smallest set that covers p</h3><p><strong>Top-p sampling</strong>, also called <strong>nucleus sampling</strong>, fixes the weakness of top-k. Instead of keeping a fixed number of tokens, it keeps a fixed share of the probability.</p><p>It works in three steps:</p><ol><li>Sort the tokens from most to least likely.</li><li>Add up their probabilities until the running total reaches p.</li><li>Keep those tokens, remove the rest, and rescale.</li></ol><p>With <strong>top-p = 0.9</strong>, our table becomes:</p><figure><img alt="Top-p sampling probability distribution table showing original token probabilities, cumulative running totals, and adjusted probabilities after applying a 0.9 nucleus sampling threshold." src="https://cdn-images-1.medium.com/max/1024/1*4DznpsLD32uoo9i5FOVD5A.png" /></figure><p>The running total passes 90% at <em>pudding</em>, so five tokens survive. That group of survivors is the “nucleus”.</p><p>The useful property is that the nucleus changes size on its own:</p><ul><li>When the model is sure, one or two tokens already cover 90%, so the choice is tight.</li><li>When the model is unsure, many tokens are needed to reach 90%, so the choice stays wide.</li></ul><p>Common values are <strong>0.9 to 0.95</strong>. Setting <strong>p = 1.0</strong> keeps every token, which switches top-p off.</p><h3>How temperature, top-k and top-p work together</h3><p>The three settings are not rivals. They are stages in one pipeline that runs for every single token:</p><ol><li>The model scores every token.</li><li><strong>Temperature</strong> reshapes the scores.</li><li><strong>Top-k</strong> and <strong>top-p</strong> trim the tail.</li><li>One token is picked at random from what is left, according to the rescaled odds.</li></ol><p>The exact order of steps 2 and 3 differs between libraries, so the same numbers can behave a little differently from one tool to another.</p><p>One practical rule: <strong>change one setting at a time</strong>. If you move two dials together, you cannot tell which one changed the output.</p><h3>How to set sampling parameters</h3><p>Where you set these parameters depends on how you reach the model. In code you pass them as arguments. In a chat app you mostly cannot, so you steer with the prompt instead.</p><h3>In code: setting temperature, top-k and top-p through an API or SDK</h3><p>Sampling parameters are ordinary arguments on the request. Which ones you are allowed to send depends on the provider and on the model, and as of November 2026 the list is getting shorter for the newest hosted models.</p><figure><img alt="Comparison of Temperature, Top-p, and Top-k sampling parameter support across Google Gemini API, OpenAI API, Claude API, and local open-source language models." src="https://cdn-images-1.medium.com/max/1024/1*FUpvdbfcVuio27EBHNqMrg.png" /></figure><p><strong>Google Gemini: all three parameters</strong></p><pre>from google import genai<br>from google.genai.types import GenerateContentConfig<br><br>client = genai.Client()<br><br>response = client.models.generate_content(<br>    model=&quot;gemini-3.5-flash&quot;,<br>    contents=&quot;Suggest five names for a banana bread bakery.&quot;,<br>    config=GenerateContentConfig(<br>        temperature=0.9,<br>        top_p=0.95,<br>        top_k=40,<br>    ),<br>)<br><br>print(response.text)</pre><p>The parameter names follow <a href="https://docs.cloud.google.com/vertex-ai/generative-ai/docs/multimodal/content-generation-parameters">Google’s content generation parameters guide</a>.</p><p><strong>OpenAI: temperature and top-p</strong></p><pre>from openai import OpenAI<br><br>client = OpenAI()<br><br>response = client.responses.create(<br>    model=&quot;gpt-6-luna&quot;,<br>    reasoning={&quot;effort&quot;: &quot;none&quot;},<br>    input=&quot;Suggest five names for a banana bread bakery.&quot;,<br>    temperature=0.9,<br>)<br><br>print(response.output_text)</pre><p>The <a href="https://developers.openai.com/api/reference/python/resources/responses/methods/create">OpenAI API reference</a> accepts a temperature from 0 to 2 and a top-p from 0 to 1. There is no top-k. On reasoning models, sending either one at any reasoning effort other than none returns an &quot;Unsupported parameter&quot; error, and not every model supports none. Check the model&#39;s page before you rely on it.</p><p><strong>A local open model with Ollama: all three parameters</strong></p><pre>import ollama<br><br>response = ollama.chat(<br>    model=&quot;llama3.2&quot;,<br>    messages=[{&quot;role&quot;: &quot;user&quot;, &quot;content&quot;: &quot;Suggest five names for a banana bread bakery.&quot;}],<br>    options={&quot;temperature&quot;: 0.9, &quot;top_k&quot;: 40, &quot;top_p&quot;: 0.95},<br>)<br><br>print(response[&quot;message&quot;][&quot;content&quot;])</pre><p>Running a model yourself gives you the most control. Nothing is hidden, and you can experiment with every sampling parameter the runtime supports.</p><p><strong>Claude: steer with the prompt, not a parameter</strong></p><p>The Claude API is the notable exception. Starting with <a href="https://platform.claude.com/docs/en/about-claude/models/whats-new-claude-4-7">Claude Opus 4.7</a>, any non-default temperature, top-p or top-k returns a 400 error. <a href="https://platform.claude.com/docs/en/api/migration-2026-10-21">API version 2026–10–21</a> removes the three parameters completely, and requests sample at the default temperature of 1.0. Anthropic’s advice is to guide the model with prompting instead.</p><pre>import anthropic<br><br>client = anthropic.Anthropic()<br><br>message = client.messages.create(<br>    model=&quot;claude-sonnet-5-5&quot;,<br>    max_tokens=1024,<br>    system=&quot;Give varied, unexpected ideas. Avoid the most obvious answers.&quot;,<br>    messages=[{&quot;role&quot;: &quot;user&quot;, &quot;content&quot;: &quot;Suggest five names for a banana bread bakery.&quot;}],<br>)<br><br>print(&quot;&quot;.join(block.text for block in message.content if block.type == &quot;text&quot;))</pre><p>The system prompt is doing the job a high temperature used to do. For the opposite effect, ask for consistency: <em>“Answer in one sentence. Use only facts from the text provided.”</em></p><p>This is the direction hosted models are moving in. Reasoning models manage more of their own sampling, and vendors point you towards prompts and reasoning-effort settings. The concepts in this post still describe what happens inside the model. You just have fewer dials to turn.</p><h3>In apps: controlling randomness in Claude and ChatGPT</h3><p>The consumer chat apps do not give you a temperature, top-k or top-p slider. The provider fixes those values for you.</p><p>A common myth is that you can type <em>“set temperature to 0.2”</em> into the chat. That does not change the real parameter. The model simply reads it as a request to be more careful, which is a prompt instruction like any other.</p><p>So in an app, the prompt is your dial:</p><figure><img alt="Guide to controlling LLM sampling behavior, showing how to get more response variety, improve consistency, save custom instructions, and adjust AI generation parameters using developer playgrounds." src="https://cdn-images-1.medium.com/max/1024/1*8ucDgTB_lyNaJYdwlCD8dA.png" /></figure><p>This is the same idea as the Claude API example above. When the dials are hidden, clear instructions do the work.</p><h3>Summary: temperature vs top-k vs top-p</h3><p>An LLM scores every possible next token, then picks one. <strong>Sampling parameters control the pick.</strong></p><figure><img alt="Comparison of LLM sampling parameters Temperature, Top-k, and Top-p, explaining their effects on token probability distributions, common configuration values, and how lowering or increasing each parameter influences AI text generation, response consistency, creativity, and vocabulary diversity." src="https://cdn-images-1.medium.com/max/1024/1*MHDvzWY8PKid8cSRV3pWHw.png" /></figure><p>The points to remember:</p><ul><li><strong>Temperature changes the odds. Top-k and top-p remove options.</strong> They are different tools and they run one after another.</li><li><strong>Top-p adapts and top-k does not.</strong> Top-p keeps few tokens when the model is sure and many when it is not.</li><li><strong>Change one setting at a time</strong>, and start from the provider’s defaults.</li><li><strong>Not every model lets you set them.</strong> Gemini and local open models expose all three. OpenAI exposes temperature and top-p with conditions. The current Claude API exposes none.</li><li><strong>In chat apps, the prompt is the dial.</strong> Ask for variety or ask for precision.</li></ul><p>The next time an LLM surprises you with a different answer to the same question, you will know why. It scored the same tokens, and then it rolled the dice.</p><p><em>If this helped, follow me for more plain-language explanations of how LLMs work.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a07550e3f6b3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Which AI Career Path Should You Choose in 2027? A Practical Guide for Developers and Freshers]]></title>
            <link>https://medium.com/@sisooriyakumar/which-ai-career-path-should-you-choose-in-2026-a-practical-guide-for-developers-and-freshers-e3199e761386?source=rss-668662b96dfd------2</link>
            <guid isPermaLink="false">https://medium.com/p/e3199e761386</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[generative-ai]]></category>
            <category><![CDATA[career-advice]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[software-engineering]]></category>
            <dc:creator><![CDATA[Sooriyakumar]]></dc:creator>
            <pubDate>Sun, 15 Nov 2026 04:15:33 GMT</pubDate>
            <atom:updated>2026-11-15T04:16:13.265Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*A-6S6zc2mUAAZCA88wECVg.jpeg" /></figure><p>Artificial intelligence is not a finished technology. It is a moving target.</p><p>Every few months a new frontier model, a new agent framework, or a new benchmark shifts what is possible. What counted as cutting-edge AI engineering two years ago is now a weekend tutorial. Tools that didn’t exist last year are already “industry standard”. The field is still being invented while we work in it.</p><p>That instability is exactly why so many people, especially freshers and early-career software engineers, feel stuck. The questions come up again and again: Should I learn machine learning from scratch? Is it enough to get good at using ChatGPT and Claude? Do I need a master’s degree? Will my job even exist in five years?</p><p>The honest answer is that there is no single “AI career”. There are several distinct paths, and they ask for very different things from you. This guide walks through the three main AI career paths, what each one requires, who it suits, and the trade-offs you should know before committing.</p><h3>The Three AI Career Paths at a Glance</h3><p>Broadly, AI-related careers fall into three groups:</p><ol><li><strong>AI-augmented software engineer</strong> — uses AI to build software faster.</li><li><strong>AI engineer</strong> — builds AI-powered software, either by integrating large language models (GenAI engineer) or by building and training models themselves (ML engineer).</li><li><strong>AI researcher</strong> — works on how AI itself works, typically through a PhD and publications.</li></ol><p>The first two are industry paths. The third is academic. Let’s take them one by one.</p><h3>Career Path 1: AI-Augmented Software Engineer</h3><figure><img alt="Infographic of the AI-augmented software engineer career path showing prerequisites, tools like Claude Code and Cursor, who it suits, and pros and cons" src="https://cdn-images-1.medium.com/max/1024/1*eUwb34MlNqZYHbXFaz83pQ.jpeg" /></figure><h4>What this path is about</h4><p>This is the most common path, and the one most developers are already on without naming it. You are still a software engineer. You build web apps, mobile apps, backend systems, and infrastructure. The difference is that AI is now your most powerful tool: frontier models like Claude, GPT and Gemini, AI-native editors, and coding agents help you design, write, review, test and document code far faster than before.</p><p>This is sometimes loosely called “vibe coding”, but that label undersells it. The real skill is directing AI effectively: knowing what to ask, how to verify the output, and how to keep architecture and quality under control while moving fast.</p><h4>Prerequisites and tools</h4><ul><li>Solid software engineering fundamentals: data structures, system design, APIs, databases, testing, version control.</li><li>At least one production-grade stack you know well (for example TypeScript/Node, Python, Java, or .NET).</li><li>Prompting and context-management skills: writing clear specifications, breaking work into tasks, and reviewing AI-generated code critically.</li><li>Typical tools: Claude Code, Cursor, GitHub Copilot, Windsurf, ChatGPT, and the APIs behind them; MCP-style integrations; AI-assisted testing and documentation tools.</li></ul><h4>Who this path fits</h4><ul><li>Developers who enjoy building products and shipping features, not training models.</li><li>Freshers who want the fastest route into a paying job while the AI landscape settles.</li><li>Engineers in startups and agencies where speed of delivery matters most.</li><li>Anyone who wants to stay a “generalist builder” rather than specialise in AI internals.</li></ul><h4>Pros and cons</h4><p><strong>Pros</strong></p><ul><li>Lowest barrier to entry; you can start today with the skills you already have.</li><li>Immediate productivity gains and strong demand across every industry.</li><li>Transferable: the fundamentals remain valuable no matter which AI tools win.</li></ul><p><strong>Cons</strong></p><ul><li>Your edge depends on tools you don’t control; the tools change every few months.</li><li>You don’t understand what’s happening inside the model, so debugging AI behaviour is largely trial and error.</li><li>As AI coding becomes the default, “uses AI tools” stops being a differentiator on its own. Engineering judgement becomes the real currency.</li></ul><h3>Career Path 2: AI Engineer</h3><figure><img alt="Infographic comparing the GenAI engineer and ML engineer career paths, with prerequisites, tools, pros and cons for each" src="https://cdn-images-1.medium.com/max/1024/1*GP8FE-yUw9loUD8RGwfjLg.jpeg" /></figure><p>An AI engineer builds software whose intelligence comes from AI rather than hand-written rules. Think of chatbots that understand intent, systems that read documents, computer vision that inspects products, or recommendation engines that learn from behaviour. This path splits into two very different routes.</p><h4>Path 2A: GenAI Engineer (LLM Integration)</h4><p>What this path is about</p><p>GenAI engineers build applications on top of large language models. They don’t train models; they orchestrate them. The work involves prompt engineering, retrieval-augmented generation (RAG), agents and tool calling, evaluation pipelines, and connecting LLMs to real business data and workflows.</p><p>Prerequisites and tools</p><ul><li>Strong backend development skills (Python or TypeScript are the most common).</li><li>Understanding of APIs, async processing, vector databases, and basic data pipelines.</li><li>Working knowledge of how LLMs behave: context windows, tokens, hallucination, temperature, embeddings.</li><li>Typical tools and frameworks: OpenAI, Anthropic and Google model APIs; LangChain, LlamaIndex, Vercel AI SDK; vector stores such as Pinecone, Weaviate, pgvector; evaluation tools like LangSmith or Ragas; open-weight models via Hugging Face, Ollama or vLLM.</li></ul><p>Who this path fits</p><ul><li>Software engineers who want to move into AI quickly without going back to mathematics.</li><li>Freshers with decent coding skills who want to ship AI products within months.</li><li>Product-minded builders who care more about user value than model internals.</li></ul><p>Pros and cons</p><p><strong>Pros</strong></p><ul><li>Fast to learn: with a programming background, three to four months of focused work gets you job-ready.</li><li>Very high current demand; most companies adopting AI start here.</li><li>You can build impressive, useful products almost immediately.</li></ul><p><strong>Cons</strong></p><ul><li>You learn how to <em>use</em> LLMs, not how they <em>work</em>. When a model behaves unexpectedly, you often lack the depth to explain why.</li><li>Heavy dependence on vendor APIs, pricing and rate limits.</li><li>The ecosystem churns quickly; frameworks that are popular today may be abandoned next year.</li><li>Shallow understanding caps your ceiling. Harder problems (fine-tuning, evaluation design, cost-quality trade-offs) require the deeper path.</li></ul><h4>Path 2B: ML Engineer (Learning AI from the Ground Up)</h4><p>What this path is about</p><p>This is the traditional route into AI engineering. You start with the foundations: statistics and probability, linear algebra and calculus, then classical machine learning, then deep learning, and only after that do you specialise in NLP, computer vision, or LLMs. You learn to collect and clean data, train and evaluate models, and deploy them in production.</p><p>Prerequisites and tools</p><ul><li>Mathematics: statistics, probability, linear algebra, basic calculus and optimisation.</li><li>Python with the scientific stack: NumPy, pandas, scikit-learn, Matplotlib.</li><li>Deep learning frameworks: PyTorch (most common today) or TensorFlow/Keras.</li><li>Data engineering basics: SQL, data pipelines, feature stores.</li><li>MLOps and deployment: MLflow, Weights &amp; Biases, Docker, cloud ML services such as AWS SageMaker, Vertex AI or Azure ML.</li><li>For LLM specialisation: Hugging Face Transformers, PEFT/LoRA fine-tuning, vLLM, quantisation tools.</li></ul><p>Who this path fits</p><ul><li>People who want to understand <em>why</em> a model works, not just <em>that</em> it works.</li><li>Developers who enjoy mathematics and experimentation and are willing to invest a year or more.</li><li>Engineers aiming at roles like ML engineer, data scientist, applied scientist, or MLOps engineer.</li><li>Anyone planning to work on problems where off-the-shelf LLMs are not enough: tabular prediction, forecasting, vision on custom data, or domain-specific models.</li></ul><p>Pros and cons</p><p><strong>Pros</strong></p><ul><li>Deep understanding: you know what LLMs and other models can and cannot do, where they fail, and how to work around their limits.</li><li>Far more durable skills; the mathematics doesn’t change when the frameworks do.</li><li>Higher ceiling and typically higher senior-level salaries; you can move into GenAI work easily, but not the reverse.</li><li>You can fine-tune, evaluate and optimise models rather than only calling them.</li></ul><p><strong>Cons</strong></p><ul><li>Slow: realistically 12 to 18 months of serious study before you are job-ready, longer without a technical background.</li><li>Mathematics is a real barrier for many developers.</li><li>Entry-level ML roles are more competitive and often expect a master’s degree.</li><li>Easy to get lost in theory and never ship anything; you must pair study with projects.</li></ul><h3>Career Path 3: AI Researcher (The Academic Path)</h3><figure><img alt="Infographic of the AI researcher career path showing PhD prerequisites, research tools, who it suits, and pros and cons" src="https://cdn-images-1.medium.com/max/1024/1*2dCPivECKneuHFYDYE9zMw.jpeg" /></figure><h4>What this path is about</h4><p>Researchers don’t just use AI; they push its boundaries. They work on new architectures, training methods, evaluation, alignment, efficiency, and the theory behind why models behave as they do. This path usually runs through a master’s degree, then a PhD, with publications at venues such as NeurIPS, ICML, ICLR or ACL. Many researchers later move to industry labs, but the entry ticket is academic depth.</p><h4>Prerequisites and tools</h4><ul><li>Strong mathematical foundation: linear algebra, probability, statistics, optimisation, and often information theory.</li><li>Research skills: reading papers critically, reproducing results, designing experiments, and writing.</li><li>Deep learning frameworks at a low level (PyTorch, JAX), plus experiment tracking and distributed training tools.</li><li>A master’s degree in a related field, followed by a PhD under a supervisor whose interests match yours.</li></ul><h4>Who this path fits</h4><ul><li>People genuinely curious about how intelligence works, not just how to apply it.</li><li>Those who can commit four to six years to a single area with delayed financial payoff.</li><li>Students who enjoy writing, experimentation, and open-ended problems without clear deadlines.</li><li>Anyone aiming at frontier labs, university faculty, or roles that shape where AI goes next.</li></ul><h4>Pros and cons</h4><p><strong>Pros</strong></p><ul><li>The deepest possible understanding of AI, including its hard limits.</li><li>Access to frontier problems and the most influential roles in the field.</li><li>Strong long-term demand for genuine expertise, which is scarce.</li></ul><p><strong>Cons</strong></p><ul><li>Longest and least certain path; a PhD takes years and research outcomes are unpredictable.</li><li>Lower income during study compared with industry peers.</li><li>Highly competitive, and much frontier research now needs compute that only large labs have.</li><li>Not suited to people who want to ship products and see users quickly.</li></ul><h3>How to Choose: A Quick Decision Guide</h3><ul><li><strong>Want a job soon and love building products?</strong> Start as an AI-augmented software engineer. Keep your fundamentals strong and treat AI tools as leverage, not a replacement for judgement.</li><li><strong>Want to build AI features fast with the coding skills you have?</strong> Go GenAI engineer. Plan to deepen your understanding over time so you don’t hit the ceiling.</li><li><strong>Want to truly understand models and have a longer runway?</strong> Go ML engineer. Learn the mathematics, build real projects, and the GenAI side will come easily.</li><li><strong>Want to shape what AI becomes and can commit years?</strong> Pursue research and a PhD.</li></ul><p>These paths are not walls. Many people start with AI-assisted development, move into GenAI work, and then go deeper into ML as their interest grows.</p><h3>Summary</h3><p>AI is still evolving, and no career choice made today is final. The sensible move is not to predict which tools will win but to pick the depth of understanding you want. If you want to build quickly, use AI as a tool and start now. If you want to build AI into products, start with LLM integration and keep going deeper. If you want to understand AI at its core, invest in the mathematics and machine learning fundamentals. And if you want to define the future of the field, take the research path. Whichever you choose, keep learning, because the ground will keep moving under all of us.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e3199e761386" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How LLMs Generate Text: Next-Token Prediction Explained Step by Step]]></title>
            <link>https://medium.com/@sisooriyakumar/how-llms-generate-text-next-token-prediction-explained-step-by-step-cd738c990a6a?source=rss-668662b96dfd------2</link>
            <guid isPermaLink="false">https://medium.com/p/cd738c990a6a</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[transformers]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[deep-learning]]></category>
            <dc:creator><![CDATA[Sooriyakumar]]></dc:creator>
            <pubDate>Sat, 14 Nov 2026 09:07:40 GMT</pubDate>
            <atom:updated>2026-11-14T09:07:40.397Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="How LLM predicts next word" src="https://cdn-images-1.medium.com/max/1024/1*76I3jTr7S3R3QN5ZT9sLeA.png" /></figure><p><em>From “I like to eat banana” to the final token: tokens, embeddings, attention, softmax and the transformer pipeline, explained for beginners.</em></p><p>You type <em>“I like to eat banana”</em> and stop. The model instantly suggests <em>“bread”</em>, <em>“splits”</em>, or <em>“every day”</em>. It feels like magic, but under the hood it is a very specific pipeline: tokens, embeddings, attention, a feed-forward network, one big matrix multiply, and a softmax. This article walks through that pipeline at a high level, using that one sentence as the running example, and clears up two questions that confuse almost every learner.</p><h3>What is next-token prediction in LLMs?</h3><p>A large language model is a <strong>next-token predictor</strong>. Given everything written so far, it outputs a probability for every token in its vocabulary, picks one, appends it, and repeats until it emits a stop token. Everything else is detail about how those probabilities are computed.</p><h3>Step 1: What is tokenization in LLMs?</h3><p>The model never sees letters or whole words directly. Your text is split into <strong>tokens</strong>, which are roughly word pieces. <em>“I like to eat banana”</em> might become five tokens: I, like, to, eat, banana. A modern model has a vocabulary of around 100,000 such tokens, each with an ID.</p><h3>Step 2: Token embeddings explained</h3><p>Each token ID is looked up in an <strong>embedding table</strong> and becomes a vector of numbers, say 4,096 of them. Similar tokens end up with similar vectors. A position signal is added so the model knows that <em>banana</em> came last and <em>I</em> came first.</p><p>At this point we have five vectors, one per token, and none of them knows anything about the others yet.</p><h3>Step 3: How the attention mechanism works in transformers</h3><p><strong>Attention</strong> is where the tokens talk to each other. For every position, the model asks: which earlier tokens matter for understanding this one? The vector for <em>banana</em> pulls in information from <em>eat</em>, so it now “knows” it is food being eaten rather than, say, a banana-coloured car. The result is a new, context-aware vector at every position.</p><p>This is what people mean by “attention vectors”: after this step you still have five vectors, but each one has absorbed context from the ones before it.</p><h3>Step 4: The feed-forward network inside a transformer block</h3><p>The part many explanations skip: after attention, each position’s vector goes through a small <strong>feed-forward neural network</strong>, two linear layers with a non-linearity in between, applied to each position independently. Attention mixes information <em>across</em> tokens; the feed-forward layer processes that information <em>within</em> each token.</p><p>One attention step plus one feed-forward step is a <strong>transformer block</strong>. A real model stacks dozens of these, so every vector is refined layer after layer. The output of the last block is the model’s final internal representation.</p><h3>Doubt 1: Does the LLM use all token vectors or only the last one?</h3><p>Both, but at different moments.</p><ul><li><strong>All positions</strong> go through every layer. The vectors for <em>I</em>, <em>like</em>, <em>to</em>, and <em>eat</em> are all computed, and they feed context into <em>banana</em> through attention.</li><li><strong>Only the last position</strong> is used to predict the next word. When the model decides what follows <em>banana</em>, it reads the final-layer vector at the <em>banana</em> position and ignores the others.</li></ul><p>That is not wasteful: the earlier vectors already did their job by shaping the last one. And because those intermediate results are cached, the next step only needs to compute the new position.</p><h3>Doubt 2: How softmax turns logits into the next word</h3><p>After the last transformer block you hold a single vector of 4,096 numbers for the <em>banana</em> position. The vocabulary has 100,000 tokens. Bridging that gap takes one more matrix.</p><ol><li><strong>The language-model head.</strong> The vector is multiplied by a matrix of shape 4,096 by 100,000, often called the <strong>unembedding</strong> or <strong>LM head</strong>. Each column corresponds to one vocabulary token.</li><li><strong>Logits.</strong> The result is a vector of 100,000 raw scores, one per token. Each score is essentially a dot product measuring how well the hidden vector lines up with that token’s direction. <em>bread</em> might score 9.1, <em>splits</em> 8.7, <em>carburettor</em> -4.0.</li><li><strong>Softmax.</strong> Softmax exponentiates every logit and divides by the total, turning the scores into probabilities that sum to one. Now <em>bread</em> might be 32%, <em>splits</em> 21%, and <em>carburettor</em> effectively zero.</li><li><strong>Sampling.</strong> The model picks a token from that distribution. Greedy decoding always takes the top one; a temperature setting flattens or sharpens the distribution to make output more creative or more deterministic.</li><li><strong>Repeat.</strong> The chosen token is appended to the input, and the whole loop runs again, now predicting from the new last position.</li></ol><p>In short: <strong>one 4,096-dimensional vector, times one matrix, equals 100,000 logits, through softmax, equals a probability per word.</strong></p><h3>The LLM text generation pipeline at a glance</h3><figure><img alt="LLM pipeline glance" src="https://cdn-images-1.medium.com/max/1024/1*iOCS2KQ3glB0Fb4b7weGLg.png" /></figure><h3>Why understanding next-token prediction matters</h3><p>Once you see the pipeline this way, a lot of LLM behaviour stops being mysterious:</p><ul><li><strong>Hallucinations</strong> are just high-probability tokens that happen to be false. The model is scoring plausibility, not truth.</li><li><strong>Context windows</strong> exist because attention has to compare every position with every earlier one.</li><li><strong>Temperature</strong> is nothing more than a knob on the softmax step.</li><li><strong>Fine-tuning</strong> reshapes which directions in that 4,096-dimensional space line up with which tokens.</li></ul><p>The next time a chat assistant finishes your sentence, picture it: five vectors flowing through a stack of blocks, one of them multiplied by a giant matrix, and a softmax quietly deciding that <em>bread</em> beats <em>carburettor</em>.</p><p><em>If this helped, share it with someone who is just starting with LLMs. Questions and corrections are welcome in the comments.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cd738c990a6a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Nvidia Needed Jensen, But Does Jensen Need Nvidia?]]></title>
            <link>https://medium.com/@sisooriyakumar/nvidia-needed-jensen-but-does-jensen-need-nvidia-1e43861e94d0?source=rss-668662b96dfd------2</link>
            <guid isPermaLink="false">https://medium.com/p/1e43861e94d0</guid>
            <category><![CDATA[trending]]></category>
            <category><![CDATA[nvidia]]></category>
            <category><![CDATA[jensenhuang]]></category>
            <category><![CDATA[future]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Sooriyakumar]]></dc:creator>
            <pubDate>Sat, 14 Nov 2026 05:56:52 GMT</pubDate>
            <atom:updated>2026-11-14T05:56:52.338Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VLI1tKSkxv3-LG96LCe3lw.jpeg" /></figure><h3>1. The Affair: What Actually Happened</h3><ul><li><strong>Oct 29, 2026:</strong> The New York Times first revealed that US intelligence suspected Alice Huang (also known as Huang Le, no relation to Jensen), an executive at Megaspeed, of working as a Chinese intelligence asset. <a href="https://www.tweaktown.com/rss/click/news/114589/?c=news-mf">tweaktown</a></li><li><strong>Nov 4:</strong> Three Republican House committee chairs, Brian Mast, John Moolenaar, and Rick Crawford, sent Huang a letter requesting a joint, closed-door briefing on November 18. <a href="https://www.keengamer.com/articles/news/nvidia-board-opens-investigation-into-ceo-jensen-huang-after-affair-admission/">keengamer</a></li><li><strong>Nov 8:</strong> Huang sent a company-wide email admitting the affair. “I was unfaithful to my wife,” he wrote. The email does not say who the affair was with. The board opened an independent investigation and said it has no reason to believe national security or company secrets were compromised. <a href="https://www.nbcnews.com/tech/tech-news/nvidia-ceo-jensen-huang-resigns-reports-extramarital-relationships-rcna608402">Nvidia CEO Jensen Huang resigns after reports of affair and extramarital relationships +2</a></li><li><strong>Nov 9:</strong> Chinese officials described the allegations as unsubstantiated and criticised the reporting. Markets reacted. Shares fell as much as 8.8% during early trading on Monday, November 9, before recovering slightly to close around 4.3% lower. <a href="https://lapaasvoice.com/nvidia-ceo-sends-company-wide-email-to-employees-regarding-chinese-spy-allegations/">lapaasvoice</a><a href="https://www.tweaktown.com/rss/click/news/114589/?c=news-mf">tweaktown</a></li><li><strong>Midweek:</strong> The scandal widened. Further reports said he had multiple extramarital relationships, including one with a suspected agent of China and another with an executive who reported to him. <a href="https://www.nbcnews.com/tech/tech-news/nvidia-ceo-jensen-huang-resigns-reports-extramarital-relationships-rcna608402">NBC News</a></li><li><strong>Nov 11–12:</strong> The Board of Directors of NVIDIA Corporation accepted the resignation of Jen-Hsun Huang, as NVIDIA’s President, Chief Executive Officer and member of the Board, effective immediately. A company spokesperson clarified that Huang will not be working with Nvidia in any capacity going forward. <a href="https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000082/nvda-20261111.htm">sec</a><a href="https://www.nbcnews.com/tech/tech-news/nvidia-ceo-jensen-huang-resigns-reports-extramarital-relationships-rcna608402">NBC News</a></li></ul><p><strong>Where things stand now:</strong> The independent members of Nvidia’s board plan to continue investigating, and Nvidia’s stock has lost $585 billion in value this week. The spying claim itself is unproven. Trump said it’s not yet clear whether Alice Huang was a spy. <a href="https://www.nbcnews.com/tech/tech-news/nvidia-investigation-jensen-huang-continue-departure-sources-say-rcna609136">Nvidia’s board is still investigating Jensen Huang despite his resignation, sources say +2</a></p><h3>2. Who Is Jensen Huang, and How Nvidia Took Over Two Markets</h3><p>Nvidia was founded on April 5, 1993 by Jensen Huang, a Taiwanese-American electrical engineer who was previously the director of CoreWare at LSI Logic and a microprocessor designer at AMD; Chris Malachowsky, an engineer who worked at Sun Microsystems; and Curtis Priem. The three of them planned the company in a Denny’s booth. They wanted to build a chip that would accelerate the use of graphics in computers. <a href="https://wiki.riteme.site/wiki/NVidia">riteme</a><a href="https://bakercityherald.com/?p=193718">bakercityherald</a></p><p>The big milestones:</p><ul><li><strong>1999:</strong> Nvidia releases the GeForce 256, the industry’s first GPU <a href="https://bakercityherald.com/?p=193718">bakercityherald</a></li><li><strong>2006:</strong> Nvidia released its Compute Unified Device Architecture (CUDA) platform, a major innovation in GPU-based parallel computing technology. This was the real turning point. It let GPUs do far more than graphics. <a href="https://www.britannica.com/money/NVIDIA-Corporation">britannica</a></li><li><strong>2016:</strong> Nobody wanted the first AI supercomputer, the DGX-1, except Musk. Huang recalled, “I boxed one up. I drove it up to San Francisco and I delivered to Elon in 2016,” and that machine went to OpenAI. <a href="https://www.benzinga.com/markets/tech/25/12/49246329/nvidias-jensen-huang-reveals-he-helped-elon-musk-build-model-s-computer-fsd-hardware">benzinga</a></li></ul><p><strong>User group 1: everyday users.</strong> Gamers, creators and students run GeForce GPUs in their laptops and desktops for gaming and local AI. This was Nvidia’s original market, and it is still its brand foundation.</p><p><strong>User group 2: the companies building AI.</strong> The GPU’s parallel design turned out to be exactly what deep learning needed. In this market Nvidia sells whole AI factories, not single cards.</p><ul><li><strong>Microsoft, Amazon, Google, Meta, Oracle:</strong> Amazon, Microsoft, Alphabet, Meta Platforms, and Oracle are among its known biggest buyers. All of them are deploying Blackwell systems (GB200/GB300 NVL72 racks). <a href="https://www.webull.com/news/13056480449766400">webull</a></li><li><strong>OpenAI, Anthropic, xAI:</strong> These privately held labs also buy at huge scale. Musk’s xAI startup built its Colossus supercomputer cluster using one hundred thousand Nvidia H100 GPUs in just nineteen days in 2024. <a href="https://www.soapcentral.com/entertainment/news-their-twisted-souls-get-twistier-kara-swisher-reacts-elon-musk-backing-jensen-huang-potential-job-offer-amid-nvidia-ceo-s-affair-scandal">soapcentral</a></li><li><strong>DeepSeek (China):</strong> DeepSeek-V3 is trained on a cluster equipped with 2048 NVIDIA H800 GPUs. The H800 is a China-specific version of the H100 that Nvidia originally designed for the China market to comply with US export restrictions. <a href="https://arxiv.org/pdf/2412.19437">arxiv</a><a href="https://www.scmp.com/tech/big-tech/article/3310639/deepseek-paper-offers-new-details-how-it-used-2048-nvidia-chips-take-openai">scmp</a></li></ul><p><strong>Dominant, though not a legal monopoly:</strong> NVIDIA holds ~80% of the AI accelerator market by revenue with $193.7B in FY2026 data center sales, versus AMD’s estimated 5–7% share. A big part of that lock-in is software. Every major AI framework (PyTorch, TensorFlow, JAX) is optimized for CUDA. <a href="https://siliconanalysts.com/analysis/amd-vs-nvidia-ai-gpu-market-share-2026">siliconanalysts</a><a href="https://blog.pebblous.ai/story/nvidia-story-pb/en/">pebblous</a></p><h3>3. What’s Next for Jensen Huang and for Nvidia</h3><p><strong>Jensen:</strong> He has not announced any plans. Huang has not yet commented on his resignation, though the New York Times reported he’s retained attorneys from law firm Paul Weiss. The Musk story is real but small. Late Wednesday, financial blog Zerohedge wrote on X, “Elon please hire Jensen,” to which Musk replied, “That would be an honor.” However, no offer, role, or formal discussions have been confirmed by either party. <a href="https://www.forbes.com/sites/alisondurkee/2026/11/13/jensen-huang-ousted-timeline-of-the-nvidia-scandal/">Jensen Huang’s Bad Week Leading To His Resignation, Explained +2</a></p><p><em>My view (opinion, not news):</em> Huang would be a huge asset to xAI or Tesla. He knows chip roadmaps, supply chains, TSMC relationships and how to scale data centers. But any move faces real hurdles. Huang could face contractual, fiduciary, or confidentiality issues depending on the terms of his agreements. He also faces an open congressional inquiry. Florida Representative Brian Mast says that Huang indicated he is willing to appear. An advisory role is more realistic in the near term than an operating role. <a href="https://finance.yahoo.com/technology/ai/articles/elon-musk-says-hiring-jensen-182821692.html">Yahoo Finance</a><a href="https://www.pcgamer.com/hardware/live/news/nvidia-jen-hsun-scandal/">pcgamer</a></p><p><strong>Nvidia:</strong> Colette Kress, Executive Vice President and Chief Financial Officer, was appointed Interim CEO. The Board is initiating a search for a permanent Chief Executive Officer and will consider internal and external candidates. Analysts are worried that it will be challenging for NVDA to find a successor with the same leadership and strategic vision as Huang. <a href="https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000082/nvidiaannouncesceotransiti.htm">NVIDIA CORP — Form 8-K — FY2026 +2</a></p><p><strong>Why this matters: AI is a hardware story too.</strong> We often talk about AI as if it were only models and code. But a frontier LLM can’t be trained on CPUs. It needs thousands of GPUs linked as one computer, in rack-scale systems filling data halls the size of buildings. Those systems also need high-bandwidth memory, fast interconnects and huge amounts of power. Whoever leads Nvidia now inherits the job of keeping that stack ahead every year. The demand is there: NVDA’s disclosed backlog for Blackwell and Rubin doubled to $1 trillion through 2027. <a href="https://daloopa.com/blog/analyst-pov/nvidia-customer-concentration-a-big-4-earnings-preview">daloopa</a></p><h3>4. My Takeaways</h3><ol><li><strong>One lapse can undo decades.</strong> Thirty-three years of building ended in one week. Talent and track record don’t protect anyone from the consequences of their own judgment. Character is part of competence.</li><li><strong>Serve every user, not one segment.</strong> Nvidia never gave up gamers to chase hyperscalers, or the other way round. The same CUDA platform runs on a student’s laptop and on a 100,000-GPU cluster. Tomorrow’s AI engineers learn on today’s consumer GPUs, so each market feeds the other.</li><li><strong>Consistency beats short-term wins.</strong> CUDA launched in 2006, years before deep learning made it valuable. The DGX-1 had no buyers at launch. Huang kept betting through quiet years, and the payoff came a decade later. The dark periods are when the foundations get built.</li><li><strong>How you handle a mistake matters, but timing and completeness matter too.</strong> Huang put the admission in writing, apologised to his family and staff, and backed the investigation instead of fighting it. That deserves credit. But to be fair, the admission came only after the press reported it. A second relationship surfaced days later. And the email didn’t address the spying question. The lesson: accountability is strongest when it comes early and covers everything, not when it is forced and partial.</li><li><strong>Build relationships before you need them.</strong> Huang personally drove a DGX-1 to Musk in 2016. Ten years later, Musk publicly called hiring him “an honor.” Opportunities often come through relationships you invested in long before you needed them. Be visible, deliver value, and keep your network alive.</li><li><strong>Grow until your value is portable.</strong> It’s partly true that Nvidia needed Jensen more than Jensen needed Nvidia: the market wiped hundreds of billions off its value in a week. But the board still let him go, so no one is irreplaceable. The real lesson for us in software is this: if your value depends only on your title, a layoff, health issue or crisis can reset you to zero. If your value is in your skills, judgment and reputation, you can always move.</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1e43861e94d0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Datawrapper: The Charting Tool Newsrooms Swear By]]></title>
            <link>https://medium.com/@sisooriyakumar/datawrapper-the-charting-tool-newsrooms-swear-by-c88364df65f5?source=rss-668662b96dfd------2</link>
            <guid isPermaLink="false">https://medium.com/p/c88364df65f5</guid>
            <category><![CDATA[no-code]]></category>
            <category><![CDATA[data-analytics]]></category>
            <category><![CDATA[storytelling]]></category>
            <category><![CDATA[data-visualization]]></category>
            <category><![CDATA[datawrapper]]></category>
            <dc:creator><![CDATA[Sooriyakumar]]></dc:creator>
            <pubDate>Thu, 12 Nov 2026 11:55:40 GMT</pubDate>
            <atom:updated>2026-11-12T11:55:40.148Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ob8sopFCn13_s6spNJwPvA.png" /></figure><p><em>How a small tool built for overworked journalists became the quiet standard for clean, honest charts on the web.</em></p><p>Open almost any major news site and scroll to the chart in the middle of the story. There’s a good chance it was made with Datawrapper. You won’t see a big logo or a flashy dashboard, just a clear chart that loads fast, reads well on a phone, and gets out of the way of the data. That restraint is the whole point.</p><h3>What is Datawrapper?</h3><p>Datawrapper is a browser-based tool for turning data into charts, maps and tables without writing a line of code. You paste or upload your numbers, pick a visualization, tweak the labels and colors, and publish. The result is an interactive, responsive graphic you can embed on any website, or a static image you can drop into a report or slide.</p><h3>Why it exists</h3><p>In 2012, making even a basic bar or line chart could tie up a newsroom graphics department for hours. Most reporters had no designer on call and no coding skills, so stories that needed a chart often went without one. Lorenz wanted any journalist to be able to make a correct, good-looking chart in a few minutes.</p><p>That newsroom DNA still shapes the product. Datawrapper favors clarity over decoration, puts text and annotation front and centre, and bakes in good data-visualization practice so that the default chart is usually a good chart. Today it is used well beyond journalism, by universities, governments, NGOs and companies, but its users still include outlets such as <em>The New York Times</em>, Reuters, Vox Media and <em>The Times</em>.</p><h3>How to use Datawrapper for visualization</h3><p>Every Datawrapper visualization is made in the same four steps, and you can try them <a href="https://app.datawrapper.de/create/chart">without even signing up</a>.</p><ol><li><strong>Upload your data.</strong> Paste a table straight from Excel or Google Sheets, upload a CSV or XLSX file, or link to a Google Sheet or an external CSV URL. Linked data can refresh automatically, which is how newsrooms keep election results or COVID dashboards up to date without touching the chart.</li><li><strong>Check and describe.</strong> Datawrapper shows your data as a spreadsheet so you can spot errors, transpose rows and columns, change number formats, or add computed columns.</li><li><strong>Visualize.</strong> Pick a chart, map or table type, then refine it: titles, descriptions, colors, axis ranges, labels, tooltips, and annotations that point readers to what matters. Live previews show how it looks on desktop and mobile as you edit.</li><li><strong>Publish and embed.</strong> One click gives you a responsive embed code and a shareable link. You can also download the chart as a PNG on the free plan, or as PDF and SVG on paid plans for print and design work.</li></ol><h3>What you can make</h3><ul><li><strong>Charts:</strong> <a href="https://www.datawrapper.de/charts">more than 20 types</a>, including bar, stacked bar, column, line, area, scatter, dot, range, arrow, pie, donut, election donut and small multiples. Business plans add waterfall and dual-axis charts for financial reporting.</li><li><strong>Maps:</strong> three <a href="https://www.datawrapper.de/maps">map types</a>. Choropleth maps colour regions by value, symbol maps place sized circles or icons on locations, and locator maps show where a story happens. Thousands of ready-made base maps cover countries, states and districts, or you can upload your own.</li><li><strong>Tables:</strong> interactive tables with search, sorting, pagination, heatmap colouring, mini bar charts, sparklines and even flag icons inside cells.</li></ul><h3>Beyond the basic chart</h3><p>Datawrapper is more than a chart maker. Teams use it for:</p><ul><li><strong>Live, auto-updating graphics</strong> fed by Google Sheets or a published CSV.</li><li><strong>Automated reporting</strong> through its <a href="https://www.datawrapper.de/developer">REST API</a>, which can create, update, publish and export charts from code. It is increasingly used by AI agents too, and Datawrapper offers an MCP server for them.</li><li><strong>Presentations and print</strong>, with PowerPoint integration and print-ready PDF or SVG export.</li><li><strong>Brand-consistent publishing</strong>, through shared team workspaces, folders and custom themes with your own fonts, colours and logo.</li><li><strong>Accessible, multilingual content</strong>, with alt text, screen-reader descriptions, full keyboard navigation, dark mode and localized number and date formats.</li><li><strong>Inspiration and reuse</strong> via <a href="https://river.datawrapper.de/">Datawrapper River</a>, a gallery of charts other users have shared for anyone to reuse.</li></ul><h3>The benefits of Datawrapper</h3><p><strong>1. No code, almost no learning curve.</strong> If you can copy a table from a spreadsheet, you can make a publishable chart in five minutes. There is nothing to install, and you can start without an account.</p><p><strong>2. Good design by default.</strong> Sensible colour palettes, readable fonts, clean gridlines and smart label placement come built in. The defaults reflect years of data-visualization research, so beginners avoid most classic charting mistakes without trying.</p><p><strong>3. Responsive and fast.</strong> Every chart, map and table adapts to the reader’s screen, from a wide monitor to a narrow phone. Embeds are lightweight, so they don’t slow down the page around them.</p><p><strong>4. A genuinely generous free plan.</strong> The <a href="https://www.datawrapper.de/pricing">Free plan</a> lets you create, publish, embed and export PNGs with no limits on charts or on how many people view them. The only catch is a small “Created with Datawrapper” credit.</p><p><strong>5. Privacy built in.</strong> Datawrapper doesn’t sell your data or track your readers. Embeds set no cookies and are Do Not Track by default, which makes them easy to use under GDPR.</p><p><strong>6. Charts that stay online.</strong> The company has pledged to keep every published chart live, even if you cancel your plan, and to let you download everything for self-hosting if it ever shut down.</p><p><strong>7. Accessible to more readers.</strong> Alt text, screen-reader descriptions, colour-blind-friendly checks, keyboard navigation and dark mode help every reader get the message.</p><p><strong>8. Built for teams and scale.</strong> Shared workspaces, folders, custom themes and an API make it easy to keep hundreds of charts consistent and on-brand.</p><p><strong>9. Affordable when you do pay.</strong> The paid tiers are simple per-seat plans:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*H9Ok9ul97hNR1Z6E97SXKg.png" /></figure><p>Non-profits get 20% off, and teachers can get the Business plan free for their classes. A custom brand theme is an add-on from $499 per month. <em>Prices from the </em><a href="https://www.datawrapper.de/pricing"><em>Datawrapper pricing page</em></a><em>, as of November 2026.</em></p><h3>How Datawrapper compares to other tools</h3><p>Datawrapper isn’t trying to be everything. It is a publishing tool for clear, static-to-lightly-interactive graphics, not an analytics platform. Here’s how it stacks up against the tools people most often weigh it against.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4cIU-N5-oH6SzX2O1dtRHg.png" /></figure><p><strong>The short version:</strong> pick Datawrapper when the chart is going in front of an audience and needs to be right, readable and quick to make. Pick Tableau or Power BI when the job is exploring data or building dashboards for internal decision-makers. Pick Flourish when movement and storytelling effects matter more than speed.</p><p>Many teams use more than one. A common pattern is analysing in Excel, Python or Tableau, then rebuilding the final chart in Datawrapper for publication.</p><h3>When to use Datawrapper: real use cases</h3><p><strong>Journalism and news:</strong> This is Datawrapper’s home turf. Reporters chart inflation, election results, weather and sports data on deadline. Live-updating charts linked to a Google Sheet keep breaking-news graphics current without a designer.</p><blockquote><strong><em>Case study: Aftenposten.</em></strong><em> Norway’s largest newspaper shows what this looks like in practice. As it moved from an ad-funded print paper to a digital-first, subscription business, Aftenposten made Datawrapper its standard charting tool. A single graphics desk used to be the bottleneck for every chart; now more than 100 journalists have each made over 10 charts, and a quarter of them have made more than 100. The team says charts make stories more engaging and more credible, summed up in their line: “An article with a chart feels like a better article.” </em><a href="https://www.datawrapper.de/case-studies/aftenposten"><em>Read the Aftenposten case study</em></a><em>.</em></blockquote><p><strong>Government and official statistics:</strong> Statistical offices and public agencies use it to publish official figures that citizens can actually read, with accessible, privacy-friendly embeds.</p><p><strong>NGOs and advocacy:</strong> Non-profits turn survey results and impact data into charts for campaigns, annual reports and websites, often on the free plan or the 20% non-profit discount.</p><p><strong>Business and finance:</strong> Companies, consultancies and banks use it for client reports, investor updates and market commentary. Waterfall charts for revenue-to-profit bridges and dual-axis charts make the Business plan a fit for finance teams.</p><p><strong>Research and academia:</strong> Researchers make clean figures for papers, posters and project sites, and export them as SVG or PDF for publication. Teachers use it to teach data literacy because students can learn good chart design in a single class.</p><p><strong>Bloggers, writers and creators:</strong> If you write on Medium, Substack or your own site, Datawrapper is one of the easiest ways to add a credible, interactive chart to a post. Paste the chart link and it embeds; or download a PNG.</p><p><strong>Automated and AI-driven reporting:</strong> Developers use the API to generate hundreds of localized charts, such as one per city or country, or to refresh weekly dashboards. AI agents can now build and publish charts through the API and Datawrapper’s MCP server.</p><h3>Reasons to choose it</h3><ul><li>You need a chart <strong>today</strong>, not next week.</li><li>Your audience is the <strong>general public</strong> or busy stakeholders, not analysts.</li><li>You care about <strong>accuracy, readability and accessibility</strong> more than flashy effects.</li><li>You publish on the <strong>web</strong> and want charts that work on every screen.</li><li>You want a tool that <strong>respects your readers’ privacy</strong>.</li><li>You’re on a <strong>tight budget</strong> and need a strong free option.</li></ul><h3>Summary</h3><p>Datawrapper is a no-code tool, started by journalist Lorenz in 2011 and launched in 2012 with developer Gregor Aisch, that turns spreadsheet data into clean charts, maps and tables in minutes. Its smart defaults, responsive embeds, privacy-first design and generous free plan make it the go-to choice for newsrooms, governments, NGOs, researchers and businesses that need to publish data clearly. It won’t replace Tableau or Power BI for deep analysis, but when you need an accurate, readable chart in front of an audience fast, few tools do it better.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c88364df65f5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How Many People Does It Take? From One to Eight Billion]]></title>
            <link>https://medium.com/@sisooriyakumar/how-many-people-does-it-take-from-one-to-eight-billion-d0292ee08218?source=rss-668662b96dfd------2</link>
            <guid isPermaLink="false">https://medium.com/p/d0292ee08218</guid>
            <category><![CDATA[society]]></category>
            <category><![CDATA[life-lessons]]></category>
            <category><![CDATA[people]]></category>
            <category><![CDATA[motivation]]></category>
            <category><![CDATA[save-planet]]></category>
            <dc:creator><![CDATA[Sooriyakumar]]></dc:creator>
            <pubDate>Wed, 04 Nov 2026 07:11:56 GMT</pubDate>
            <atom:updated>2026-11-04T07:11:56.767Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/508/1*80D0cW-OUpvgXPCoakfqsQ.jpeg" /></figure><p><em>Every person matters. Every partnership multiplies. But some things only become possible when all of us show up.</em></p><p>The other day I caught myself wondering about a simple question: how many people does it actually take to do something?</p><p>Some things need just one person. Some need two. Some need a small team, a village, a whole nation. And a few things, the biggest things, need all of us.</p><p>So I started counting, from one upward. Here is what I found along the way.</p><h3>1 person: the work of becoming yourself</h3><p>Some jobs can only be done alone. Nobody can learn for you. Nobody can heal for you, rest for you, or love yourself on your behalf.</p><p>Self-improvement, self-love and self-care are one-person tasks. They are not selfish. They are the foundation everything else stands on.</p><p>Think of it this way: a society is just thousands of individuals added together. When one person becomes calmer, kinder, more skilled and more honest, that improvement spreads. It reaches their family, their workplace, their street. Better people quietly build better societies.</p><h4>But can one person achieve greatness alone?</h4><p>At first glance, it seems so. Albert Einstein reshaped physics. Isaac Newton gave us the laws of motion and gravity. Tom Cruise became the face of action cinema. Cristiano Ronaldo became one of the greatest footballers the world has seen.</p><p>Look closer, though, and the picture changes. Einstein had teachers, colleagues and the scientists whose work he built on. Newton himself famously said he saw further by standing on the shoulders of giants. Tom Cruise works with stunt teams, directors, crews and studios. Ronaldo had a mother who sacrificed for him, coaches who shaped him, teammates who passed him the ball, and millions of fans who lifted him.</p><p>Behind every famous individual stands a family, a team, an industry and a crowd of supporters.</p><p>So what makes that one person special? Their dedication. Their effort. Their patience when nothing was working. Others gave them the stage, but they are the ones who refused to walk off it. That combination of personal grit and collective support is what turns an ordinary life into an exceptional one.</p><h3>2 people: the partnership</h3><p>The best example of two is a life partnership: a married couple who choose to walk through life side by side.</p><p>Two people come together mainly to help each other. And the beautiful thing is that they rarely have the same strengths. One is patient where the other is quick. One is good with money where the other is good with people. One stays calm in a crisis while the other brings the energy to start something new.</p><p>Put them together and something clever happens: <strong>one person’s weakness is covered by the other’s strength.</strong> The pair becomes stronger and less vulnerable than either would be alone.</p><p>The practical benefits add up too. Two incomes. Two pairs of hands. Two minds solving problems. Two hearts to carry the hard days. When one falls, the other helps them up.</p><p>That, to me, is what marriage really stands for. Not two halves becoming one, but two whole people making each other better.</p><p>The same idea applies beyond marriage: best friends, business co-founders, a mentor and a student. Two is the smallest number where you are no longer alone.</p><h3>3 people: the first family</h3><p>Add a child to a couple and you have a family. Three people means the couple is no longer living only for each other. They are now building something for the future.</p><p>Three is also the number of a tie-breaker. When two people disagree, a third voice can help find balance. Many small committees and decision panels use three members for exactly this reason.</p><h3>4 people: the circle of support</h3><p>Four feels like a full table. A family of four. A group of close friends. A rowing crew where four oars keep the boat steady and straight.</p><p>With four, roles start to appear naturally. One plans, one does, one checks, one encourages. Nobody has to do everything, and everybody matters.</p><h3>5 people: the small team</h3><p>Five is where real teamwork begins. A basketball team puts five players on the court. Many rock bands have had five members, each playing a different instrument that only sounds complete together.</p><p>In workplaces, small teams of around five are often the sweet spot: big enough to cover different skills, small enough that everyone still knows what the others are doing.</p><h3>11, 100, 1,000: teams, villages and towns</h3><p>Keep counting and the pattern keeps growing. Eleven players make a football team, where even Ronaldo needs ten others on the pitch. A hundred people can build a school or run a small business. A thousand people make a village or a town, with farmers, teachers, doctors and shopkeepers all depending on one another.</p><p>No single person in a town grows all the food, builds all the houses and heals all the sick. Each does a part, and the whole community lives better because of it.</p><h3>Millions: nations</h3><p>Millions of people working together build roads, hospitals, power grids and education systems. They recover from floods and wars. They vaccinate whole generations against disease.</p><p>At this scale, no one even knows most of the people they are cooperating with. Yet every taxpayer, nurse, engineer and volunteer is part of the same effort. A nation is proof that strangers can achieve together what no small group ever could.</p><h3>Billions: one planet</h3><p>And then we reach the biggest number of all: roughly eight billion of us.</p><p>As far as we know, humans are the only beings who can consciously decide to protect a whole planet. And there are billions of us. Just think about how much that could mean.</p><p>If one person can change their own life, and two people can build a family, imagine what billions can do:</p><ul><li>If every person planted just one tree, we would add billions of trees to the Earth.</li><li>If every household cut a little waste, mountains of plastic would never reach the ocean.</li><li>If every person saved a little water and energy each day, the total would be enormous.</li></ul><p>One person can reach a milestone. Two people can build a life. But billions of people, working together, can achieve almost anything, including healing the only home we have.</p><p>The problems facing our planet, like climate change, pollution and disappearing forests, are too big for any hero to solve alone. Not Einstein, not Newton, not any leader or celebrity. They can only be solved by everyone.</p><h3>So, how many people does it take?</h3><p>It depends on what you want to build.</p><ul><li><strong>1</strong> to grow yourself.</li><li><strong>2</strong> to share a life.</li><li><strong>3 to 5</strong> to build a family or a team.</li><li><strong>Thousands</strong> to build a community.</li><li><strong>Millions</strong> to build a nation.</li><li><strong>Billions</strong> to save a planet.</li></ul><p>Here is what I believe after all this counting.</p><p><strong>Each person is important</strong>, because every big number is made of ones. <strong>A supporting partner is even more important</strong>, because two people lift each other higher than either could climb alone. <strong>But saving the Earth is the most important of all</strong>, and that is the one job none of us can do alone.</p><p>So start with yourself. Hold on to the people beside you. And then join the billions doing good, one tree, one choice, one kind act at a time.</p><p>The count starts with you.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d0292ee08218" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Transformer: The Architecture That Started the AI Boom]]></title>
            <link>https://medium.com/@sisooriyakumar/the-transformer-the-architecture-that-started-the-ai-boom-9daa88cd6fa4?source=rss-668662b96dfd------2</link>
            <guid isPermaLink="false">https://medium.com/p/9daa88cd6fa4</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[generative-ai]]></category>
            <category><![CDATA[transformers]]></category>
            <category><![CDATA[deep-learning]]></category>
            <dc:creator><![CDATA[Sooriyakumar]]></dc:creator>
            <pubDate>Sun, 01 Nov 2026 06:16:11 GMT</pubDate>
            <atom:updated>2026-11-01T06:16:11.144Z</atom:updated>
            <content:encoded><![CDATA[<h4>From RNNs to LLMs</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jIz9Q4LCuvGSEOtaEiL6qQ.png" /></figure><p>We have all heard the word <strong>Transformer</strong>. It shows up in every AI headline, every LLM paper, and every “how ChatGPT works” blogs. GPT, Claude, Gemini, Llama and BERT all use it.</p><p>But here’s an honest question: <strong>do we actually know how a Transformer works, and why it was invented in the first place?</strong></p><p>If ChatGPT is the child, the Transformer is the parent. Almost every large language model in use today is built on one architecture, introduced in a 2017 paper with a confident title: <strong><em>“Attention Is All You Need.”</em></strong></p><p>The Transformer did not appear from nowhere. It was the answer to years of researchers hitting the same walls with older models. To understand why it matters, we need to follow that path from the start.</p><p>In this post we will:</p><ol><li>Understand <strong>what a Transformer is</strong>, in one paragraph.</li><li>Walk through the <strong>history</strong>: deep learning → RNN → LSTM → GRU → Encoder-Decoder → Attention → Transformer. At each step we’ll see what problem was solved and what new problem appeared.</li><li>Open up the <strong>Transformer architecture</strong> and explain every component in simple English.</li><li>Follow a <strong>real translation example</strong> from the original paper, word by word.</li><li>Meet the <strong>Transformer family</strong>: encoder-only, decoder-only, and encoder-decoder models.</li><li>See <strong>how the Transformer caused the AI boom</strong> and how today’s LLMs use it.</li></ol><p>Let’s begin.</p><h3>Part 1: What Is a Transformer?</h3><p>A <strong>Transformer</strong> is a type of neural network that takes a sequence as input, such as a sentence, and produces another sequence or an understanding of it. Its main idea is a mechanism called <strong>attention</strong>, which lets every word in a sentence look directly at every other word and decide which ones matter most.</p><p>Older models read a sentence one word at a time, from left to right. A Transformer looks at <strong>the whole sentence at once</strong>. That one change made it faster to train, better at understanding context, and able to scale to enormous sizes. Scale is what eventually gave us large language models.</p><p>That’s the short version. Now let’s see why we needed it.</p><h3>Part 2: The Road to Transformers</h3><p>Think of this section as a chain. Each idea solves a problem with the one before it, then shows a new weakness that the next idea has to fix.</p><h4>2.1 Deep Learning: The Foundation</h4><p><strong>Deep learning</strong> is a branch of machine learning that uses <strong>neural networks</strong> with many layers. A neural network is a stack of simple math units, loosely inspired by neurons. Each unit takes numbers in, combines them using learned <strong>weights</strong>, and passes the result on.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*D-CTYpJVHgFrEY7CA2F_aQ.jpeg" /></figure><p>During training, the network makes predictions, compares them with the correct answers, and adjusts its weights slightly to reduce the error. This adjustment process is called <strong>backpropagation</strong>. After millions of examples, the network becomes very good at the task.</p><p><strong>The problem:</strong> a basic neural network expects a <strong>fixed-size input</strong> and has <strong>no memory</strong>. It treats every input independently. Language is a <strong>sequence</strong>, and order matters:</p><blockquote><em>“The dog bit the man” ≠ “The man bit the dog”</em></blockquote><p>Sentences also have different lengths. We needed a network that could read sequences and remember what came before.</p><h4>2.2 Embeddings: Turning Words into Numbers</h4><p>First, a quick but important stop. Neural networks only understand numbers, so how do we feed them words?</p><p>We turn each word into a list of numbers called a <strong>vector</strong>, or <strong>embedding</strong>. The model learns these numbers so that <strong>words with similar meanings get similar vectors</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1u-8pAMu_iGly_dXp8T7SA.jpeg" /></figure><p>You can picture embeddings as points on a map of meaning. “Cat” and “dog” sit near each other, and “Paris” sits near “France”. Popular early methods included <strong>Word2Vec</strong> (2013) and <strong>GloVe</strong> (2014).</p><p><strong>The limitation:</strong> these classic embeddings give each word <strong>one fixed meaning</strong>. The word “bank” gets the same vector in “river bank” and “bank account”. Context was missing. Keep this in mind, because the Transformer fixes it.</p><h4>2.3 RNN: Giving Networks a Memory</h4><p>A <strong>Recurrent Neural Network (RNN)</strong> reads a sequence <strong>one word at a time</strong> and carries a <strong>hidden state</strong>, a kind of running memory, from one step to the next.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LIEuZv7UNoERm-0SfrrDxg.jpeg" /></figure><p>At each step, the RNN combines <strong>the current word</strong> with <strong>its memory of everything so far</strong> and produces an updated memory. It’s like reading a book and keeping a mental summary as you go.</p><p><strong>What it solved:</strong> it handles sequences of any length, and order matters.</p><p><strong>The problems:</strong></p><ol><li><strong>Vanishing gradients and short memory.</strong> When an RNN learns, the error signal has to travel backwards through every time step. Over long sentences, this signal shrinks until it’s almost zero, so the network <strong>forgets early words</strong>. In “The <strong>cat</strong>, which my neighbour adopted from the shelter last spring, <strong>was</strong> hungry,” a plain RNN struggles to connect “was” back to “cat.”</li><li><strong>Sequential and slow.</strong> Step 5 can’t be computed until step 4 is done. You can’t process words in parallel, so training on large data is slow.</li></ol><h4>2.4 LSTM: A Smarter Memory</h4><p><strong>Long Short-Term Memory (LSTM)</strong> networks, proposed by Hochreiter and Schmidhuber in 1997, were designed to fix the forgetting problem.</p><p>An LSTM adds a separate <strong>cell state</strong>, like a conveyor belt of memory running through the sequence, and three <strong>gates</strong> that control it:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qDT2XwzoMmF6dvJPxeXyfg.jpeg" /></figure><ul><li><strong>Forget gate:</strong> “What old information should I throw away?”</li><li><strong>Input gate:</strong> “What new information is worth storing?”</li><li><strong>Output gate:</strong> “What part of my memory should I show right now?”</li></ul><p>Because information can flow along the cell state with little change, LSTMs <strong>remember much longer</strong> than plain RNNs. For years, LSTMs powered speech recognition, translation and text prediction.</p><p><strong>Remaining problems:</strong></p><ul><li>Still <strong>sequential</strong>: one word at a time, so it’s still slow to train.</li><li>Memory is better but <strong>still limited</strong> for very long texts.</li><li>More gates mean <strong>more computation</strong> per step.</li></ul><h4>2.5 GRU: A Lighter LSTM</h4><p>The <strong>Gated Recurrent Unit (GRU)</strong>, introduced by Cho et al. in 2014, simplifies the LSTM. It merges the cell state and hidden state and uses only <strong>two gates</strong>:</p><ul><li><strong>Update gate:</strong> how much of the old memory to keep vs. replace.</li><li><strong>Reset gate:</strong> how much of the past to ignore when reading the new word.</li></ul><p>GRUs often perform about as well as LSTMs with fewer parameters, so they train a bit faster.</p><p><strong>But the core issue remains:</strong> GRUs are still <strong>recurrent</strong>. They read word by word, can’t be parallelised across a sentence, and still have trouble with very long-range dependencies. Faster gates couldn’t fix the problem. The design itself had to change.</p><h4>2.6 The Encoder-Decoder (Seq2Seq) Model</h4><p>Now consider <strong>translation</strong>: English in, French out. The input and output lengths differ, and the word order can change. In 2014, Sutskever et al. (and Cho et al.) proposed the <strong>Encoder-Decoder</strong> or <strong>Sequence-to-Sequence (Seq2Seq)</strong> architecture, built from two RNNs or LSTMs:</p><ul><li>The <strong>Encoder</strong> reads the full input sentence and compresses it into a single vector called the <strong>context vector</strong>.</li><li>The <strong>Decoder</strong> takes that vector and generates the output sentence word by word.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OL1izuQu7-rpWG_BYSeOIg.jpeg" /></figure><p>This was a big breakthrough for machine translation. Keep the words <strong>encoder</strong> and <strong>decoder</strong> in mind, because the Transformer keeps this two-part structure.</p><p><strong>The problem: the bottleneck.</strong> The entire input, whether 5 words or 50, must be squeezed into <strong>one fixed-size vector</strong>. Imagine summarising a full page into one sentence and then translating from only that sentence. Translation quality dropped sharply as sentences got longer.</p><h4>2.7 Attention: Look Back at What Matters</h4><p>In 2014, <strong>Bahdanau, Cho and Bengio</strong> introduced <strong>attention</strong>, which was refined by Luong et al. in 2015. The idea is simple:</p><blockquote><em>Instead of forcing the decoder to rely on one summary vector, let it </em><strong><em>look back at all the encoder’s hidden states</em></strong><em> and decide </em><strong><em>which input words matter most for the word it’s producing right now</em></strong><em>.</em></blockquote><p>When the decoder is about to write “étudiant”, it should focus mostly on “student”:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*j4EfOTnxjGHHw_ofI-PBhg.jpeg" /></figure><p>These <strong>attention weights</strong> are learned automatically and add up to 1. The decoder builds a fresh, focused context vector for <strong>every</strong> output word. The bottleneck is gone, and translation of long sentences improved a lot.</p><p><strong>Remaining problem:</strong> attention was still <strong>attached to RNNs</strong>. The encoder and decoder still read word by word. Training was still sequential, slow, and hard to scale to large datasets.</p><p>So researchers asked a bold question:</p><blockquote><strong><em>If attention is doing the heavy lifting, why do we need the RNN at all?</em></strong></blockquote><h4>2.8 Enter the Transformer (2017)</h4><p>In June 2017, eight researchers at Google (Vaswani, Shazeer, Parmar, Uszkoreit, Jones, Gomez, Kaiser and Polosukhin) published <strong>“Attention Is All You Need.”</strong> They <strong>removed recurrence completely</strong> and built a model using only attention plus simple feed-forward layers.</p><p>Here’s the whole history in one table:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VhFvob6tlfpw2VvzFldOUA.jpeg" /></figure><p>The Transformer reached state-of-the-art translation quality and trained in a <strong>fraction of the time</strong>. The base model trained in about 12 hours on 8 GPUs. Now let’s look inside.</p><h3>Part 3: The Transformer Architecture, Explained</h3><h4>3.1 The Big Picture</h4><p>The original Transformer keeps the <strong>Encoder-Decoder</strong> structure:</p><ul><li>The <strong>Encoder</strong> reads the input sentence (English) and builds a rich, context-aware understanding of every word.</li><li>The <strong>Decoder</strong> uses that understanding to generate the output sentence (French), one word at a time.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ErdJBSoLnS_flpIhQ5MmEg.jpeg" /></figure><p>In the original paper, both the encoder and decoder are stacks of <strong>6 identical layers</strong>. Let’s follow the data through, step by step.</p><h4>3.2 Step 1: Tokenization</h4><p>Before anything else, text is split into <strong>tokens</strong>. A token can be a full word, part of a word, or punctuation. The original paper used <strong>sub-word tokenization</strong> (byte-pair encoding and word-piece). Common words stay whole, and rare words get split into pieces:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2Sdt8Xj1CW-Qgu0vOn82vg.jpeg" /></figure><p>Each token is then mapped to an <strong>ID number</strong> from a fixed vocabulary of about 37,000 tokens for the English-German model.</p><p><strong>Why sub-words?</strong> The model can handle words it has never seen by building them from known pieces, and the vocabulary stays a manageable size.</p><h4>3.3 Step 2: Input Embedding</h4><p>Each token ID is turned into an <strong>embedding vector</strong>, the “list of numbers that captures meaning” from Part 2. In the original Transformer, each vector has <strong>512 numbers</strong>, a size called <em>d_model</em>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UzW5BQEuHc8LkocHEUFClg.jpeg" /></figure><p>These embeddings are <strong>learned during training</strong>. At this stage each vector only knows the word itself, not its neighbours. Context comes later.</p><h4>3.4 Step 3: Positional Encoding</h4><p>The Transformer reads all words <strong>at the same time</strong>, not one after another. That’s great for speed, but it creates a problem:</p><blockquote><em>Without help, the model sees “dog bites man” and “man bites dog” as the </em><strong><em>same bag of words</em></strong><em>.</em></blockquote><p>The fix is to <strong>add a position signal</strong> to each embedding. The original paper used a fixed pattern of sine and cosine waves of different frequencies. Each position (1st, 2nd, 3rd, …) gets its own unique pattern, a bit like a fingerprint.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PwtQw3MpGorBFz9I3hKtyg.jpeg" /></figure><p>Now every vector carries two things: <strong>what the word is</strong> and <strong>where it is</strong> in the sentence. (Many later models learn the position signal instead of using fixed waves. Newer LLMs often use variants such as <em>rotary</em> position embeddings, but the purpose is the same.)</p><h4>3.5 Step 4: Self-Attention, the Heart of the Transformer</h4><p>This is the main idea, so we’ll take it slowly.</p><p><strong>The goal:</strong> turn each word’s vector into a <strong>context-aware</strong> vector by letting it gather information from the other words in the sentence.</p><p>Look at this sentence:</p><blockquote><em>“The </em><strong><em>animal</em></strong><em> didn’t cross the street because </em><strong><em>it</em></strong><em> was too tired.”</em></blockquote><p>What does “it” refer to? A human knows it’s “the animal.” Self-attention lets the model work this out: when processing “it,” the model <strong>pays strong attention to “animal.”</strong> Change “tired” to “wide” and “it” should now attend to “street.”</p><p>The Q, K, V idea: a library analogy</p><p>For each word, the model creates three new vectors from its embedding, using three learned weight matrices:</p><ul><li><strong>Query (Q):</strong> <em>“What am I looking for?”</em></li><li><strong>Key (K):</strong> <em>“What do I contain? What can I be matched on?”</em></li><li><strong>Value (V):</strong> <em>“What information do I actually hand over if someone picks me?”</em></li></ul><p>Think of a library:</p><ul><li>You walk in with a <strong>question</strong> (Query): “books about space travel.”</li><li>Every book has a <strong>label on its spine</strong> (Key).</li><li>You compare your question with every label and see which ones match best.</li><li>Then you read the <strong>contents</strong> (Value) of the best-matching books, paying more attention to the better matches.</li></ul><p>How it works, step by step</p><p>For a word like “it”:</p><ol><li><strong>Score:</strong> compare the Query of “it” with the Key of <strong>every</strong> word, including itself. A higher score means more relevant.</li><li><strong>Scale:</strong> divide the scores by a constant (the square root of the key size) so the numbers don’t get too large and training stays stable.</li><li><strong>Softmax:</strong> turn the scores into <strong>attention weights</strong> between 0 and 1 that add up to 1.</li><li><strong>Blend:</strong> take a weighted mix of all the Value vectors using those weights.</li></ol><p>The result is a new vector for “it” that’s <strong>mostly made of “animal”’s information</strong>, mixed with a little from the others.</p><p>Here’s the one formula of this post, which is all four steps above in one line:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YTnn62BTlJX8bxzw0RAzLw.jpeg" /></figure><ul><li>Q · Kᵀ → compare every Query with every Key (step 1)</li><li>/ √d_k → scale (step 2)</li><li>softmax → turn into weights (step 3)</li><li>· V → blend the Values (step 4)</li></ul><p>Diagram: self-attention for one word</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YUnSFVB0Lic2y7CpxQpBbw.jpeg" /></figure><p><strong>The key point:</strong> this happens for <strong>every word at the same time</strong>, as one big matrix operation. No word waits for another. That’s the parallelism RNNs never had.</p><h4>3.6 Step 5: Multi-Head Attention</h4><p>One attention calculation can only focus on one kind of relationship at a time. Language has many: grammar (subject ↔ verb), meaning (it ↔ animal), position (neighbouring words), and more.</p><p>So the Transformer runs <strong>8 attention “heads” in parallel</strong>, each with its own Q, K, V weights. Each head can learn to focus on something different:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MJHBZgFOMzdmhIDX9Kak9Q.jpeg" /></figure><p>It’s like asking eight experts to read the same sentence, each looking for something different, then combining their notes. Each head works with smaller vectors (512 ÷ 8 = 64 numbers), so the total cost is about the same as one big head.</p><h4>3.7 Step 6: Add &amp; Norm (Residual Connection + Layer Normalization)</h4><p>After attention, two small but important things happen:</p><ul><li><strong>Add (residual connection):</strong> the original input to the layer is <strong>added back</strong> to the attention output. The model only has to learn <strong>what to change</strong>, not rebuild everything from scratch. It also gives the learning signal a “highway” through deep stacks, which helps avoid the vanishing gradient problem we saw with RNNs.</li><li><strong>Norm (layer normalization):</strong> the numbers are rescaled to a stable range so training stays smooth and doesn’t blow up or die out.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6DsWrEYBJhiTdJOeFQvfnQ.jpeg" /></figure><p>This “Add &amp; Norm” wraps <strong>every</strong> sub-layer in both the encoder and the decoder.</p><h4>3.8 Step 7: Feed-Forward Network</h4><p>Next, each word’s vector goes through a small <strong>two-layer neural network</strong>, applied to every position separately and identically. It expands the vector from 512 to 2,048 numbers, applies a non-linearity (ReLU), and shrinks it back to 512.</p><p>A simple way to think about it:</p><ul><li><strong>Attention</strong> = words <strong>talking to each other</strong> (gathering context).</li><li><strong>Feed-forward</strong> = each word <strong>thinking on its own</strong> about what it just learned.</li></ul><p>Research suggests these layers also store a lot of the model’s learned “knowledge.”</p><h4>3.9 The Encoder Layer, Put Together</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LBoJhprThk_IOplngf6LUA.jpeg" /></figure><p>The output of one encoder layer becomes the input to the next. After 6 layers, every word’s vector is <strong>deeply context-aware</strong>. The vector for “bank” now clearly means “river bank” or “money bank” depending on its neighbours. This fixes the “one fixed meaning” problem of the old embeddings.</p><p>The final encoder output is passed to <strong>every decoder layer</strong>.</p><h4>3.10 The Decoder: Generating the Output</h4><p>The decoder generates the translation <strong>one token at a time</strong>. Each new token is predicted using (a) the tokens generated so far and (b) the encoder’s understanding of the input. Each decoder layer has <strong>three</strong> sub-layers.</p><p>(a) Masked Multi-Head Self-Attention</p><p>The decoder’s words attend to each other, with one rule: <strong>a word may only look at words before it, never after.</strong></p><p>Why? When generating word 3, words 4 and 5 don’t exist yet. During training we feed in the whole correct French sentence at once for speed, so we have to <strong>mask</strong> (hide) the future words. Otherwise the model would just copy the answer.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eR36tclyZm2PHH5KWablDQ.jpeg" /></figure><p>(b) Cross-Attention (Encoder-Decoder Attention)</p><p>This is where the decoder <strong>looks at the English sentence</strong>. It works exactly like attention, except:</p><ul><li><strong>Queries</strong> come from the <strong>decoder</strong> (“what French word am I trying to make?”),</li><li><strong>Keys and Values</strong> come from the <strong>encoder output</strong> (“here’s what the English sentence contains”).</li></ul><p>This is the 2014 attention idea in a new form, and it’s the bridge between the two languages.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qRWV86Tmc7crHJEa8_jSxg.jpeg" /></figure><p>The same as in the encoder. Each position processes its own updated information. Each sub-layer is again wrapped in Add &amp; Norm, and the decoder also has <strong>6 layers stacked</strong>.</p><h4>3.11 Step 8: Linear + Softmax, Picking the Next Word</h4><p>The decoder’s final vector for the current position goes through:</p><ol><li>A <strong>Linear layer</strong> that converts it into one score for <strong>every token in the vocabulary</strong>, tens of thousands of scores.</li><li>A <strong>Softmax</strong> that turns those scores into <strong>probabilities</strong>.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hL6RjN2EfS_F05PieZvuIw.jpeg" /></figure><p>The chosen token is added to the output, and the process repeats until the model produces an <strong>end-of-sentence</strong> token.</p><h4>3.12 Training vs. Inference</h4><ul><li><strong>Training:</strong> the model sees the English sentence and the <strong>correct</strong> French sentence at the same time. Thanks to masking, it predicts <strong>every</strong> next word in parallel and learns from its mistakes. Feeding the correct previous words is called <strong>teacher forcing</strong>. This parallel training is a big reason Transformers train so fast.</li><li><strong>Inference (actual use):</strong> there’s no correct answer to feed, so the decoder generates <strong>one token at a time</strong>, feeding each new token back in as input for the next step.</li></ul><h3>Part 4: A Worked Example from the Original Paper</h3><p>The original paper evaluated the Transformer on two machine translation tasks from the <strong>WMT 2014</strong> benchmark: <strong>English → German</strong> and <strong>English → French</strong>. It set new records on both, with BLEU scores of <strong>28.4</strong> (EN→DE) and <strong>41.8</strong> (EN→FR) using the larger model, at a much lower training cost than previous top systems.</p><p>Let’s trace a simple English → French example through the whole architecture:</p><blockquote><strong><em>Input:</em></strong><em> “I am a student”<br> </em><strong><em>Output:</em></strong><em> “Je suis étudiant”</em></blockquote><p>(Notice that French drops the “a.” The model has to learn that the article disappears here, which simple word-by-word translation would get wrong.)</p><h4>Stage 1: Preparing the input (English side)</h4><p><strong>Tokenize:</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*M6wpe3ap21hjByu6QWgG9A.png" /></figure><p>(IDs are made up for illustration.)</p><p><strong>Embed:</strong> each ID becomes a 512-number vector. At this point, “student” only means the general idea of “student.”</p><p><strong>Add position:</strong> “I” gets the position-1 signal, “am” position 2, “a” position 3, “student” position 4. Now the model knows the order.</p><h4>Stage 2: The encoder builds understanding</h4><p>All four tokens enter the encoder <strong>together</strong>.</p><p>In <strong>self-attention</strong>, each word gathers context from the others:</p><ul><li><strong>“I”</strong> attends strongly to <strong>“am”</strong>, since they form subject + verb.</li><li><strong>“am”</strong> attends to <strong>“I”</strong> (who is it?) and <strong>“student”</strong> (being what?).</li><li><strong>“a”</strong> attends to <strong>“student”</strong>, since it’s the article for that noun.</li><li><strong>“student”</strong> attends to <strong>“a”</strong> and <strong>“am”</strong> and learns it’s the thing being described, the complement of “I am.”</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gzPMewaFAS7BIEkqwACf9w.png" /></figure><p>Different <strong>heads</strong> capture different patterns: one may track grammar, another meaning, another neighbouring words.</p><p>The <strong>feed-forward</strong> layers then refine each word’s vector. After <strong>6 encoder layers</strong>, we have four rich vectors. Each one knows its own meaning and its role in the sentence: “I” as the subject, “student” as what the subject is, and so on.</p><p>This set of vectors is the <strong>encoder output</strong>, the model’s “understanding” of the English sentence. It’s computed <strong>once</strong> and reused for every French word.</p><h4>Stage 3: The decoder generates French, one token at a time</h4><p><strong>Step 1: generate the first word</strong></p><ul><li>Decoder input: [&lt;start&gt;]</li><li><strong>Masked self-attention:</strong> only &lt;start&gt; exists, so it attends to itself.</li><li><strong>Cross-attention:</strong> the decoder’s Query is roughly “how should a French sentence start here?” It looks at the English vectors and focuses mostly on <strong>“I”</strong> (and “am”).</li><li><strong>Feed-forward → Linear → Softmax:</strong> highest probability is <strong>“Je”</strong>.</li></ul><pre>Output so far: Je</pre><p><strong>Step 2: generate the second word</strong></p><ul><li>Decoder input: [&lt;start&gt;, Je]</li><li><strong>Masked self-attention:</strong> “Je” looks at &lt;start&gt; and itself. It knows a subject pronoun has just been written, so a verb probably comes next.</li><li><strong>Cross-attention:</strong> focuses on <strong>“am”</strong>.</li><li><strong>Prediction:</strong> <strong>“suis”</strong>, the French form of “am” that goes with “Je.”</li></ul><pre>Output so far: Je suis</pre><p><strong>Step 3: generate the third word</strong></p><ul><li>Decoder input: [&lt;start&gt;, Je, suis]</li><li><strong>Masked self-attention:</strong> the decoder knows it has “Je suis” (I am) and now needs what comes after.</li><li><strong>Cross-attention:</strong> puts <strong>little weight on “a”</strong> and <strong>strong weight on “student.”</strong> Across its training data, the model learned that French says “Je suis étudiant” with no article for professions after “être.”</li><li><strong>Prediction:</strong> <strong>“étudiant”</strong></li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nrJwYA3RH8HVuudXF_n6Dg.png" /></figure><p><strong>Step 4: finish</strong></p><ul><li>Decoder input: [&lt;start&gt;, Je, suis, étudiant]</li><li>Cross-attention finds that every part of the English input has been covered.</li><li><strong>Prediction:</strong> &lt;end&gt;. Translation complete.</li></ul><pre>Final output: &quot;Je suis étudiant&quot;</pre><h4>What happened during training</h4><p>During training, the model saw millions of sentence pairs like this. For our pair, it received the full input [&lt;start&gt;, Je, suis, étudiant] at once. With masking, it predicted “Je,” “suis,” “étudiant,” and &lt;end&gt; <strong>all in parallel</strong>, compared its predictions with the right answers, and adjusted its weights. Repeat this millions of times and the attention patterns above emerge on their own. Nobody programmed “focus on student when writing étudiant.” The model learned it.</p><h3>Part 5: The Transformer Family</h3><p>After 2017, researchers realised you don’t always need both halves of the Transformer. Three main families grew from the original design.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fsaAMfRfV8n-3y6UUZacXA.png" /></figure><p><em>*Exact architectures of some commercial models aren’t fully public, but they’re widely understood to be decoder-style Transformers.</em></p><h4>5.1 Encoder-Only Models: Built to Understand</h4><p><strong>How they differ:</strong> they keep <strong>only the encoder stack</strong> and drop the decoder. With no masking, every word sees <strong>every other word in both directions</strong>, so these models are <strong>bidirectional</strong>.</p><p><strong>How they’re trained:</strong> mainly with <strong>Masked Language Modeling</strong>. Hide random words and ask the model to fill in the blanks:</p><blockquote><em>“The cat sat on the </em><strong><em>[MASK]</em></strong><em>.” → “mat”</em></blockquote><p><strong>Famous examples:</strong></p><ul><li><strong>BERT</strong> (Google, 2018): the model that made “pre-train, then fine-tune” standard in NLP. Also trained with “next sentence prediction.”</li><li><strong>RoBERTa</strong> (Meta, 2019): BERT trained more carefully, with more data, longer training, and without next sentence prediction. Same architecture, better results.</li><li><strong>ALBERT:</strong> a lighter BERT that shares parameters across layers.</li><li><strong>DistilBERT:</strong> a smaller, faster BERT trained to copy a larger one (“distillation”).</li><li><strong>DeBERTa:</strong> improves how attention handles word content and position.</li></ul><p><strong>Best for:</strong> understanding tasks such as sentiment analysis, text classification, named entity recognition, search ranking, and question answering by highlighting an answer span. They <strong>don’t generate</strong> long text well.</p><h4>5.2 Decoder-Only Models: Built to Generate</h4><p><strong>How they differ:</strong> they keep <strong>only the decoder stack</strong>. With no encoder, the <strong>cross-attention layer is removed</strong>, leaving masked self-attention + feed-forward. Each token can only look <strong>left</strong>, at earlier tokens.</p><p><strong>How they’re trained:</strong> <strong>next-token prediction</strong>. Given the text so far, predict the next token:</p><blockquote><em>“The capital of France is” → “Paris”</em></blockquote><p>This simple goal, at huge scale, turns out to teach grammar, facts, reasoning patterns and much more.</p><p><strong>Famous examples:</strong></p><ul><li><strong>GPT, GPT-2, GPT-3, GPT-4 and later</strong> (OpenAI): the line that showed scale works.</li><li><strong>Llama</strong> (Meta): popular open-weight models.</li><li><strong>Claude</strong> (Anthropic), <strong>Gemini</strong> (Google), <strong>Mistral</strong>, and most of today’s chatbots.</li></ul><p><strong>Best for:</strong> text generation, chat, writing, coding, summarisation, and reasoning. Today they also handle many “understanding” tasks just by being prompted.</p><h4>5.3 Encoder-Decoder Models: Built to Transform Input into Output</h4><p><strong>How they differ:</strong> they keep the <strong>full original structure</strong>, both stacks plus cross-attention. The improvements are mostly in training data, objectives and scale.</p><p><strong>Famous examples:</strong></p><ul><li><strong>T5</strong> (Google, 2019): “Text-to-Text Transfer Transformer.” Turns <strong>every</strong> task into text in, text out. For example, “translate English to German: …” or “summarize: …”.</li><li><strong>BART</strong> (Meta, 2019): trained to repair corrupted text, which makes it strong at summarisation.</li><li><strong>mT5:</strong> a multilingual T5.</li><li><strong>MarianMT:</strong> efficient translation models.</li></ul><p><strong>Best for:</strong> tasks with a clear input → output mapping, such as translation, summarisation, and rewriting.</p><h4>Quick Comparison</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SZisUbYWUe-DYdfHX_qzcw.png" /></figure><h3>Part 6: How the Transformer Started the AI Boom</h3><p>Why did this one architecture change the field so much? Here are the benefits, and how each one led to today’s LLMs.</p><h4>Benefit 1: Parallel Processing</h4><p>RNNs read word by word. Transformers process <strong>all tokens at once</strong> using matrix operations, which is exactly what <strong>GPUs and TPUs</strong> are built for.</p><p><strong>→ Led to:</strong> training on <strong>much larger datasets</strong> in reasonable time. Instead of millions of sentences, models could learn from a large share of the public internet.</p><h4>Benefit 2: Long-Range Understanding</h4><p>In attention, any word connects <strong>directly</strong> to any other word in one step. In an RNN, information had to pass through every word in between and faded along the way.</p><p><strong>→ Led to:</strong> models that can follow long documents, long conversations, and large codebases. Context windows grew from hundreds of tokens to hundreds of thousands, even millions, in recent models.</p><h4>Benefit 3: Scalability</h4><p>The Transformer is a simple stack of identical blocks. Making it bigger is mostly a matter of adding layers, widening vectors and adding heads. Research on <strong>scaling laws</strong> around 2020 found that performance improves <strong>smoothly and predictably</strong> as you increase model size, data and compute.</p><p><strong>→ Led to:</strong> the race to scale, from about 65 million parameters in the original base model, to 1.5 billion in GPT-2, 175 billion in GPT-3, and beyond. At larger scales, models showed new abilities such as few-shot learning, following instructions, and step-by-step reasoning.</p><h4>Benefit 4: Pre-training and Transfer Learning</h4><p>You can pre-train <strong>one</strong> large Transformer on general text, then fine-tune or prompt it for many tasks.</p><p><strong>→ Led to:</strong> <strong>foundation models</strong>. Instead of building a new model for every task, companies build on one very capable base. That made AI far cheaper and faster to adopt.</p><h4>Benefit 5: Contextual Understanding</h4><p>The same word gets different representations in different contexts (“bank” by a river vs. “bank” with money).</p><p><strong>→ Led to:</strong> a big jump in language understanding quality, starting with BERT’s results in 2018.</p><h4>Benefit 6: Works Beyond Text</h4><p>Attention works on <strong>any sequence</strong>: image patches, audio frames, protein structures.</p><p><strong>→ Led to:</strong> <strong>Vision Transformers (ViT)</strong> for images, <strong>Whisper</strong> for speech, <strong>AlphaFold 2</strong> (which uses attention-based modules) for protein structure, and today’s <strong>multimodal</strong> models that read text, images and audio together.</p><h4>How Today’s LLMs Use the Transformer</h4><p>Today’s chatbots are essentially <strong>very large decoder-only Transformers</strong>, built in roughly three stages:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uPMz4otQD1GUFfyF3ocKcQ.png" /></figure><p>When you ask a chatbot a question, it runs the same loop we traced in Part 4: tokenize, embed, add positions, pass through dozens of Transformer layers full of multi-head attention, apply softmax over the vocabulary, pick the next token, and <strong>repeat</strong>. The answer you see is generated one token at a time.</p><p>The release of ChatGPT in November 2022 brought this to the public. Behind it was a direct line back to one 2017 translation paper.</p><h4>Fair Warning: The Transformer Has Limits Too</h4><p>In keeping with the theme of this post, every solution brings a new problem:</p><ul><li><strong>Quadratic cost:</strong> every token attends to every other token, so doubling the input length roughly <strong>quadruples</strong> the attention computation. Very long contexts are expensive.</li><li><strong>Huge resource needs:</strong> large models need large data, a lot of energy, and costly hardware.</li><li><strong>Hallucinations:</strong> predicting the most likely next token doesn’t guarantee a true answer.</li></ul><p>Researchers are actively working on these issues with efficient attention variants (such as FlashAttention and sparse attention), mixture-of-experts models, and alternatives like state-space models (for example Mamba).</p><h3>Summary</h3><p>Here’s the whole story in a few lines:</p><ul><li><strong>Deep learning</strong> gave us networks that learn from data, but they had no memory for sequences.</li><li><strong>RNNs</strong> added memory but forgot long context and were slow.</li><li><strong>LSTMs</strong> and <strong>GRUs</strong> used gates to remember longer, but were still sequential.</li><li><strong>Encoder-decoder</strong> models handled translation, but squeezed everything into one vector.</li><li><strong>Attention</strong> let the decoder look back at the relevant words, but was still built on RNNs.</li><li>The <strong>Transformer</strong> (2017) removed recurrence and kept attention. The result was parallel, fast, scalable, and better at context.</li></ul><p>Inside, the Transformer follows a clear flow: <strong>tokens → embeddings → positional encoding → encoder (self-attention + feed-forward) → decoder (masked self-attention + cross-attention + feed-forward) → linear + softmax → next token.</strong> The core is <strong>Query, Key, Value</strong> attention: every word asks a question, matches it against every other word’s label, and collects the information it needs.</p><p>From that one design came <strong>BERT</strong> (understanding), <strong>GPT</strong> (generation), <strong>T5</strong> (input-to-output tasks), and eventually the LLMs we use every day. Parallelism, scalability, and the ability to learn from huge data are what made the AI boom possible.</p><p>So the next time you chat with an AI, remember that each word it writes comes out of attention layers built on a 2017 paper about translating English into French and German.</p><h3>Over to You 💬</h3><p>Here’s a question for you:</p><blockquote><strong><em>The Transformer replaced RNNs because they couldn’t scale. Today, the Transformer’s own weakness is the cost of very long contexts. Do you think the next big AI breakthrough will be an improved Transformer, or a completely new architecture that replaces it, just as the Transformer replaced the RNN ?????</em></strong></blockquote><p>Share your thoughts in the comments. I’d love to read them. If this post helped you finally understand Transformers, give it a clap 👏 and share it with someone who’s still wondering what “attention” means.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9daa88cd6fa4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Is Coding Still Relevant in 2027?]]></title>
            <link>https://medium.com/@sisooriyakumar/is-coding-still-relevant-in-2027-f4d8fd0b4acf?source=rss-668662b96dfd------2</link>
            <guid isPermaLink="false">https://medium.com/p/f4d8fd0b4acf</guid>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[2027]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[coding]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Sooriyakumar]]></dc:creator>
            <pubDate>Sat, 31 Oct 2026 14:51:45 GMT</pubDate>
            <atom:updated>2026-10-31T14:51:45.646Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zHd5VGqi-_r7oq7HdNlNgg.jpeg" /></figure><p><em>AI writes most of our code now. So why should anyone still learn to write it by hand?</em></p><p>Short answer: yes, coding is still relevant in 2027. And it will stay relevant for as long as software jobs exist.</p><p>But to see why, we first need to look at how we got here. The way developers work has changed more in the last five years than in the twenty before it. Let’s walk through that journey in four phases, and then come back to the big question.</p><h3>Phase 1: Every line, by hand (before 2022)</h3><p>Not so long ago, there was no AI in a developer’s day. We opened the editor and wrote everything ourselves. Every if condition, every for loop, every while loop, every variable name came from our own fingers.</p><p>When we got stuck, the routine was always the same:</p><ul><li>Copy the error message into Google.</li><li>Open the first Stack Overflow thread.</li><li>Scroll past the accepted answer to find the one that actually works.</li><li>Read the docs, a blog post, or a GitHub issue from three years ago.</li></ul><p>Then we went back to the editor and typed the fix ourselves. It was slow, but it built something valuable. You understood every line because you wrote every line.</p><h3>Phase 2: Ask, copy, paste (late 2022 onwards)</h3><p>OpenAI’s GPT models had been around for a few years (GPT-3 arrived in 2020), but they lived mostly in research labs and APIs. That changed on 30 November 2022, when ChatGPT launched as a simple chat app anyone could use.</p><p>Almost overnight, a new habit was born. Instead of searching Stack Overflow, we opened a browser tab and asked the AI directly:</p><blockquote><em>“Write a function that validates an email address and returns true or false.”</em></blockquote><p>We described what the function should do, got a block of code back, and copied it into our project.</p><h4>The catch: everything depended on the prompt</h4><p>The quality of the answer depended almost entirely on the quality of the question. A vague prompt gave you vague code: missing edge cases, outdated libraries, functions that looked right but quietly failed. A detailed prompt gave you something usable.</p><p>So “prompt engineering” became a skill of its own. People who could describe the problem precisely got good results. Everyone else spent a lot of time fixing AI-generated bugs.</p><h4>How the models closed the gap</h4><p>Over time, the LLM providers filled that gap themselves. Starting around late 2024, models gained the ability to “think” before answering: to plan, reason through the problem and check their own work.</p><p>Today, even a short, simple prompt usually produces solid code, because the model fills in the missing details on its own. That reasoning ability is now standard in every frontier model, and even in many smaller ones.</p><h3>Phase 3: AI moves into the IDE</h3><p>In Phase 2, AI was a separate place. We had to leave our editor, go to the AI, ask, copy and come back. The next step was obvious: bring the AI to where the code lives.</p><p>It started with <strong>code completion</strong>. Tools like GitHub Copilot (first previewed in 2021) sat inside the editor and suggested the next line, or the rest of the function, as you typed. Press Tab, and it was done.</p><p>This felt like autocomplete on steroids. You were still driving, but the AI finished your sentences. Over time these tools added inline chat too, so you could highlight a block of code and ask the AI to explain or fix it without leaving the file.</p><h3>Phase 4: The AI agent becomes the workspace (today)</h3><p>Now we are in the fourth phase. AI coding agents such as Claude Code (launched in 2025) plug into the IDE as an extension. They don’t just complete a line. They read the whole codebase, plan the change, edit multiple files, run the tests and fix what breaks.</p><p>This has quietly flipped how developers spend their day. When many of us open the IDE now:</p><ul><li><strong>70 to 80% of the time</strong> goes into the AI chat session: defining the task, writing the requirement, reviewing the plan and asking the agent to build it.</li><li><strong>Only 10 to 20% of the time</strong> goes into the actual source files, whether that’s JavaScript, Java, Python or anything else.</li></ul><p>The developer’s job has shifted from <em>typing the code</em> to <em>defining the work and checking the result</em>.</p><p>And since we now talk to the AI in natural language (English, or even our own mother tongue), a tempting idea has spread: <em>maybe we don’t need to know coding anymore.</em></p><h3>The myth: “Just know English and write a good prompt”</h3><p>The idea goes like this: if the AI understands plain language, all you need is something like:</p><blockquote><em>“You are an expert full-stack developer. Build a login page.”</em></blockquote><p>That will produce <em>a</em> login page. But is it the login page your product needs? A real login page raises questions the prompt never answered:</p><ul><li>Do users sign in with a username, an email, or both?</li><li>What are the password rules? At least 8 characters, with a special character?</li><li>Are passwords hashed before they are stored, and with which algorithm?</li><li>What happens after five failed attempts? Is there rate limiting or a lockout?</li><li>Is there “forgot password”, social login or two-factor authentication?</li></ul><p>The AI can implement every one of these. But <em>you</em> have to know they exist, and know which ones matter for your system. That knowledge doesn’t come from English. It comes from understanding software: security, data, architecture and how things fail.</p><p>So the prompt is only as good as the technical depth of the person writing it. Without that depth, you get a system that looks fine in a demo but isn’t reliable, stable or safe for real customers.</p><h3>Four reasons coding still matters</h3><h4>1. Someone has to review the code</h4><p>AI generates code incredibly fast. But at the end of the day, it’s a machine, and we can’t trust it 100%. It can misread the requirement, invent an API that doesn’t exist, or introduce a subtle security hole.</p><p>So every change still needs a human review and approval, at least at a high level. And you can’t review what you can’t read. To judge whether code is correct, safe and maintainable, you need to know how good code looks. That means knowing how to code.</p><h4>2. Small tasks don’t need an AI (and it costs tokens)</h4><p>Say you want to center a div. You could ask the agent to do it. It will search the codebase, read several files, find the right component and change one CSS property. That can burn thousands of tokens for a ten-second job.</p><p>If you know the code, you search for the element’s ID, open the component and add the property yourself. Done.</p><p>When tokens are plentiful, this may not matter. When they’re limited, knowing how to code means you save the AI for the work that actually needs it.</p><h4>3. Usage limits run out, deadlines don’t</h4><p>AI coding tools come with usage limits. With Claude, for example, usage works in session windows that reset every few hours. If you hit your limit with two hours of work still left, you can’t tell your manager, “Sorry, my AI session ended, so I can’t code.”</p><p>You still have to deliver. And to deliver, you need to finish the task yourself. That’s only possible if you can actually write the code.</p><h4>4. Coding is the foundation</h4><p>Everything else (good prompts, good reviews, good architecture decisions) sits on top of understanding how code works. Without that foundation, you are not directing the AI. You are just hoping it gets things right.</p><h3>Restaurants exist. Should you still learn to cook?</h3><p>“AI writes code, so do I need to know coding?” is really the same question as “Restaurants exist, so do I need to know cooking?”</p><p>The honest answer to both is: <strong>it’s a hybrid.</strong></p><p>Most working people can’t cook every single day. Some days you’re busy, some days you’re tired, some days you just want to relax, so you eat out. That’s perfectly fine. AI is the restaurant: fast, convenient and often very good.</p><p>But people who know how to cook understand food. They know what goes into a dish, how it should look and taste, and when something is off. They can still feed themselves when the restaurant is closed.</p><p>Coding works the same way. Let the AI handle the heavy lifting on most days. But know the fundamentals well enough to review its work, handle the small fixes yourself and keep shipping when the AI isn’t available.</p><h3>The bottom line</h3><p>Coding is not going away in 2027. Not in 2050 either, as long as software jobs exist. What’s changing is <em>how</em> we code: less typing, more thinking, planning and reviewing.</p><p>The developers who thrive won’t be the ones who stopped learning to code. They’ll be the ones who know code deeply enough to make the AI genuinely useful.</p><p><em>What about you? How much of your day is spent in the AI chat versus the actual code? Share in the comments.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f4d8fd0b4acf" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The BAT Model for Software Development: Build, Audit, Test]]></title>
            <link>https://medium.com/@sisooriyakumar/the-bat-model-for-software-development-build-audit-test-17b8f022ffc1?source=rss-668662b96dfd------2</link>
            <guid isPermaLink="false">https://medium.com/p/17b8f022ffc1</guid>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[product-development]]></category>
            <category><![CDATA[software-testing]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Sooriyakumar]]></dc:creator>
            <pubDate>Sat, 31 Oct 2026 10:31:15 GMT</pubDate>
            <atom:updated>2026-10-31T10:31:15.934Z</atom:updated>
            <content:encoded><![CDATA[<p><em>How AI changed the way we ship software, and the three-phase model we now use to keep quality high.</em></p><p>For the past few months, our team has been building software a little differently. We stopped splitting work the old way, frontend here and backend there, and started following a simple three-step model we call <strong>BAT: Build, Audit, Test</strong>.</p><p>In this post I want to share what BAT is, why we landed on it, and what I have learned from using it day to day. This is my personal take, not a rulebook, but I hope it gives you a useful glimpse into how AI is reshaping the development process.</p><h3>Traditional software development</h3><p>Before we talk about BAT, let’s look at how most teams have built software until very recently.</p><p>A typical feature needed at least two developers:</p><ul><li><strong>The frontend developer</strong> built the UI: screens, forms, layouts and user interactions.</li><li><strong>The backend developer</strong> built the APIs: business logic, database access and integrations.</li></ul><p>The two worked in parallel. As soon as the backend developer finished an API, the frontend developer picked it up and wired it into the UI. This integration step was often where the friction lived: mismatched field names, missing endpoints, unclear error formats, and a lot of back-and-forth messages.</p><p>Once the frontend and backend came together, the system was deployed to a test environment. Then the QA team took over. They tested everything, logged bugs, and sent them back to the developers.</p><p>So the flow looked like this:</p><ol><li>Frontend and backend are developed separately.</li><li>The two are integrated, piece by piece, as APIs become ready.</li><li>The system is deployed to a test environment.</li><li>Testers test it and report bugs.</li></ol><p>This model worked for years. But it depended heavily on handoffs between people, and every handoff added waiting time.</p><h3>What changed: AI made building accessible to everyone</h3><p>In the last year or two, things have shifted fast. With today’s AI coding tools, almost anyone can build software. A 12 or 15 year old who knows how to describe what they want clearly can get a working app out of an AI assistant. The hard part is no longer typing the code. The hard part is knowing what to ask for.</p><p>I like to think of AI as a very skilled assistant. It knows <strong>how</strong> to do things: how to write a UI component, how to design an API, how to connect a database. What it does not know is <strong>what</strong> to do for your product. That is where the prompt comes in. A clear, specific instruction tells the assistant what to build, and the assistant takes care of the how.</p><p>This changes the economics of a team. One person, with good prompts and good judgment, can now build an entire platform, portal or product: UI, APIs, database and integration together. That single shift is what makes the BAT model possible.</p><h3>The BAT model: Build, Audit, Test</h3><p>BAT splits delivery into three phases. Build and Audit happen in development. Only Test happens in the test environment.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/825/1*ZeBSEtf9jqJKkkXTZM-Xew.png" /></figure><p><em>Audit catches the obvious issues inside development, so only edge cases reach the testers.</em></p><h4>B: Build</h4><p>In the Build phase, one person uses AI to create the base version of the product, end to end. That means the UI, the APIs, the database layer and the integration between them, all generated from clear prompts and requirements.</p><p>The builder’s job is not to type every line. It is to:</p><ul><li>Break the product into clear, well-described features.</li><li>Write precise prompts with the right context: business rules, data shapes, design expectations.</li><li>Review what the AI produces and steer it when it drifts.</li></ul><p>The output is a working base version. But it is rarely perfect. AI does not have the full context of your users, your business or your design system. So the build can carry behavioural bugs, UI glitches, broken responsiveness on mobile, or flows that technically work but feel wrong.</p><h4>A: Audit</h4><p>The Audit phase exists to catch those obvious issues before anyone else sees them. Either the builder switches hats, or a second person steps in.</p><p>The auditor walks through the <strong>happy paths</strong>: the main user journeys that the product must support. Can a user sign up? Log in? Create, edit and delete the core records? Does the layout hold up on a phone? The goal is simple: make sure <strong>at least 80% of the main flows work properly</strong>.</p><p>Anything broken is fixed right away, directly in the build, while the context is still fresh. Think of it as structured dev testing, done with intent rather than as an afterthought.</p><h4>T: Test</h4><p>Only after the audit does the product move to the test environment. Here, the tester focuses on the <strong>remaining 20%</strong>: edge cases, unusual inputs, error states, permission boundaries, concurrency and the strange things real users do in production.</p><p>Because the audit already removed the obvious bugs, testers spend their time on the problems that are genuinely hard to find.</p><h3>Isn’t this just full stack development?</h3><p>This is the first question I usually get. A full stack developer also builds the frontend and the backend alone, so what is new?</p><p>The short answer: <strong>full stack is a role, BAT is a process.</strong> And the Build phase works very differently from how a full stack developer traditionally works.</p><p>A full stack developer writes the code themselves, one layer at a time. They might build the API in the morning, switch to the UI in the afternoon, and wire the two together the next day. They are one person, but they still move through the stack sequentially, carrying the integration in their head. Their speed is limited by how fast they can write code and switch context between layers.</p><p>In the Build phase, the AI does most of the writing across all layers in the same pass. The builder describes a feature once, and the AI generates the UI, the endpoint, the data model and the glue between them together. The person’s job shifts from writing code to specifying, directing and reviewing it.</p><p>The second difference is structure. A full stack developer’s quality checks depend on personal habit. BAT makes them explicit, with a named Audit phase and a named Test phase that always follow Build.</p><p><strong>Full stack development vs the BAT model at a glance:</strong></p><ul><li><strong>What it is:</strong> Full stack is a developer role. BAT is a delivery process with three phases.</li><li><strong>Who writes the code:</strong> In full stack, the developer writes it by hand. In BAT, mostly AI writes it, directed by the builder.</li><li><strong>How layers are built:</strong> Full stack goes one at a time: backend, then frontend, then integration. BAT builds them together, feature by feature, in one pass.</li><li><strong>Main skill needed:</strong> Full stack needs deep coding skill across the stack. BAT needs clear requirements, good prompts and strong review.</li><li><strong>Integration:</strong> In full stack, it is done manually after both sides exist. In BAT, it is generated with the feature, then verified.</li><li><strong>Quality checks:</strong> In full stack, they are up to the individual. In BAT, they are built in: Audit in development, Test in QA.</li><li><strong>Speed limit:</strong> Full stack is limited by how fast one person can code and switch context. BAT is limited by how fast one person can specify and review.</li></ul><p>To be fair, a full stack developer who uses AI tools heavily is already doing something close to Build. BAT simply names that shift and wraps it with the checks that AI-generated code needs.</p><h3>One, two or three people: BAT fits your team size</h3><p>One of the things I like most about BAT is how flexible it is. The three phases stay the same. Only the number of people changes, depending on the project budget and the team you have.</p><ul><li><strong>1 person:</strong> Person A builds, audits and tests. Good for small budgets, MVPs and internal tools.</li><li><strong>2 people (option 1):</strong> Person A builds; Person B audits and tests. The builder stays focused, and a fresh pair of eyes checks everything.</li><li><strong>2 people (option 2):</strong> Person A builds and audits; Person B tests. The builder fixes their own obvious issues, and a dedicated tester handles edge cases.</li><li><strong>3 people:</strong> Person A builds, Person B audits, Person C tests. Good for larger products, client projects and a higher quality bar.</li></ul><p>The more people you add, the more independent each check becomes. A person auditing their own build will always have some blind spots, so when the budget allows, a separate auditor or tester is worth it.</p><h3>Why not just Build and Test? The case for Audit</h3><p>A fair question: if we already have testers, why add an Audit phase in the middle?</p><p>Build is obviously essential; without it, there is nothing to test. But sending a raw AI-generated build straight to the test environment creates a mess:</p><ul><li>Testers get flooded with obvious bugs that the builder could have caught in minutes.</li><li>Every bug becomes a ticket, a handoff and a context switch for the builder.</li><li>Tracking and retesting all of that takes far more time than fixing it on the spot.</li></ul><p>The Audit phase stops this. Obvious issues are found and fixed during development, before the build is ever released to the test environment. That saves the builder’s time, and it saves the testers’ time for the work only they can do.</p><p>AI helps teams ship faster, but not automatically safer. The more code a machine generates, the more places there are for small mistakes to hide. A dedicated audit step is our way of closing that gap.</p><h3>And why testing still matters</h3><p>Imagine two products. The first has a plain UI but has been through rigorous testing; most real-world cases are covered. The second looks beautiful but has barely been tested. Which one would you trust with your business? The first one, every time.</p><p>A product becomes great through testing. And that is even more true when a machine wrote most of the code. AI is fast and capable, but it does not understand your users, and it can be confidently wrong. So no matter how good the build looks, the Test phase stays.</p><h3>Final thoughts</h3><p>AI has changed who can build software and how fast it can be built. But speed alone does not make a good product. The BAT model is our attempt to keep the speed AI gives us while protecting the quality our users expect:</p><ul><li><strong>Build</strong> fast with AI, end to end.</li><li><strong>Audit</strong> the happy paths and fix the obvious issues before release.</li><li><strong>Test</strong> the edge cases that real users will eventually hit.</li></ul><p>It is working well for us so far, and it scales from a one-person project to a full team. I would love to hear how your team is adapting its process to AI. Are you doing something similar? Let me know in the comments.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=17b8f022ffc1" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>