<?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 Bharati Gehani on Medium]]></title>
        <description><![CDATA[Stories by Bharati Gehani on Medium]]></description>
        <link>https://medium.com/@bharatigehani?source=rss-5a2b76ae2846------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*pDdMzhxDww1sk7BmMoE-oQ.jpeg</url>
            <title>Stories by Bharati Gehani on Medium</title>
            <link>https://medium.com/@bharatigehani?source=rss-5a2b76ae2846------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 14:53:32 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@bharatigehani/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[How AI Engineers Actually Make Decisions: Stanford CS230 Lecture 2, in Pictures]]></title>
            <link>https://medium.com/@bharatigehani/how-ai-engineers-actually-make-decisions-stanford-cs230-lecture-2-in-pictures-9be21102a92b?source=rss-5a2b76ae2846------2</link>
            <guid isPermaLink="false">https://medium.com/p/9be21102a92b</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[deep-learning]]></category>
            <dc:creator><![CDATA[Bharati Gehani]]></dc:creator>
            <pubDate>Mon, 12 Oct 2026 08:21:01 GMT</pubDate>
            <atom:updated>2026-10-12T08:21:01.119Z</atom:updated>
            <content:encoded><![CDATA[<h4>Supervised, self-supervised and weakly supervised learning, explained from scratch through four real projects.</h4><p>Lecture 1 of Stanford’s CS230 was about why deep learning matters. Lecture 2 is about how you actually build with it.</p><p>It’s taught by Kian Katanforoosh, Professor Andrew Ng’s co-instructor and the founder of Workera. His style is simple: he gives the class a real project and keeps asking, “What would you decide here?”</p><p>I wrote this so you can follow along even if you have never watched the lecture or trained a model. Every term gets explained the first time it shows up. Let’s go!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lSjHWMzE3AEHVV-EA6dRNg.png" /><figcaption>Lecture 2 in one picture</figcaption></figure><h3>First, a Quick Crash Course</h3><p>Before the projects, Prof Kian recaps the basics. Everything else builds on these basics.</p><h4>What is a model?</h4><p>A model has two parts. The <strong>architecture</strong> is the blueprint: how many layers it has and how they connect. The <strong>parameters</strong> are the millions (sometimes billions) of numbers inside that blueprint that get tuned during training.</p><p>His favorite simplification: when you chat with an AI, you are really calling two files sitting in the cloud. One describes the architecture. The other stores the parameters.</p><h4>How does a model learn?</h4><p>Say you want a model that answers one question: is there a cat in this photo?</p><ol><li>You show it a photo. Its parameters start out random, so it guesses badly, say “10% chance of a cat.”</li><li>A <strong>loss function</strong> measures how wrong that guess was. The photo had a cat (the right answer is 1) and the model said 0.1, so the penalty is big.</li><li><strong>Gradient descent</strong> nudges every parameter a tiny bit in the direction that would have made the guess less wrong.</li><li>Repeat with millions of photos until the guesses are right.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CMmVdF90PMEkOMJZc-mG_g.png" /><figcaption>The learning loop</figcaption></figure><p>Learning from examples where you already know the right answer (the <strong>label</strong>) is called <strong>supervised learning</strong>.</p><h4>What is a neuron?</h4><p>A computer sees a color photo as three grids of numbers: how much red, green and blue is in every pixel. You flatten those grids into one long list and feed it to a neuron.</p><p>The neuron does two things. First, it multiplies each number by a weight and adds everything up, plus one extra number called the bias. Then it passes the result through a <strong>sigmoid</strong>, a function that squeezes any number into a value between 0 and 1, so the output reads like a probability. An output of 0.73 is above 0.5, so the model says “cat.”</p><p>Put many neurons side by side and you get a <strong>layer</strong>. Stack layers and you get a <strong>deep neural network</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nKVa7sVYPs0Pai8HQA0bUw.png" /><figcaption>One neuron, step by step</figcaption></figure><h4>From one animal to three</h4><p>Prof asked the class how to turn the cat detector into one that spots cats, dogs and giraffes. You add two more output neurons, one per animal, and collect photos of dogs and giraffes.</p><p>The step people forget is the labels. Each label becomes a list of three numbers. A photo with only a cat is [0, 1, 0], called a <strong>one-hot</strong> label. A photo with a cat and a dog is [1, 1, 0], a <strong>multi-hot</strong> label. Prof Kian says adding new data but forgetting to update the labels is the number one mistake he sees in projects.</p><h4>Bigger isn’t always better</h4><p>A deeper network has more <strong>capacity</strong>: it can learn more complex patterns. But capacity has to match your data. Give a billion-parameter model only a million cat photos and it will memorize those exact photos without ever learning what a cat is. That’s called <strong>overfitting</strong>.</p><h4>What happens inside the layers</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RIi2tfd677p0lI2jxpQgjA.png" /><figcaption>From pixels to faces, one layer at a time</figcaption></figure><p>Train a network on faces and look inside. The first layers only see raw pixels, so they learn simple shapes like diagonal and vertical edges. Middle layers combine those edges into eyes, noses and ears. The deepest layers respond to whole faces.</p><p>Nobody told the network to look for eyes. Before deep learning, engineers hand-built a detector for each feature. That’s <strong>feature engineering</strong>. Now the network discovers the features from the data by itself. That’s <strong>feature learning</strong>.</p><p>Two more words you’ll need later:</p><ul><li>An <strong>encoding</strong> is the list of numbers a network produces for an input at some layer.</li><li>An <strong>embedding</strong> is an encoding where distance means something: similar things get similar numbers. Two photos of the same person land close together, and photos of different people land far apart.</li></ul><h3>Project 1: Day or Night?</h3><p>The task: given a photo, say whether it was taken during the day or at night. (Pretend tools like ChatGPT don’t exist. The point is to understand the decisions under the hood.)</p><h4>Step 1: Define the task</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*koowX3u3UgbEBWcHkmKjVg.png" /><figcaption>Every “easy” task hides edge cases</figcaption></figure><p>It sounds trivial until the class started listing hard cases: indoor photos with no sky, a clock on the wall (AM or PM?), dawn and dusk, cloudy days, and northern Norway, where the sun doesn’t set in summer.</p><p>A model for one park camera is easy. A model for any photo in the world needs far more data and a bigger model. Deciding which one you are building comes first.</p><h4>Step 2: Pick the resolution</h4><p>Resolution is a real trade-off. Too low and the clues disappear (the clock becomes a blur). Too high and every image has more numbers to process, so training gets slower and more expensive. Slow training means fewer experiments.</p><p>How do you pick quickly? This is my favorite trick of the lecture.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NObLQsHAwDApggWFyVrSpA.png" /><figcaption>Use humans as a proxy</figcaption></figure><p>Use humans as a proxy. Prof Kian’s team printed the same photos at different resolutions and asked friends: day or night? Below a certain size, people just couldn’t tell. They landed on 64 × 64 pixels with all 3 color channels, because a blue sky is a huge clue. The whole experiment took minutes.</p><h4>Step 3: The rest of the recipe</h4><ul><li><strong>Output:</strong> one number between 0 (night) and 1 (day), so the last step is a sigmoid.</li><li><strong>Architecture:</strong> a small network is enough for a simple task. (Convolutional networks, which are great for images, come later in the course.)</li><li><strong>Loss function:</strong> binary cross-entropy, also called logistic loss. It gives a small penalty when the model is right and confident, and a big penalty when it is confident and wrong.</li></ul><p>One more insight: when you plan a project, don’t only ask “how accurate will it be?” Ask “how fast can I iterate?” If one training run takes a year, you can’t learn anything.</p><p><strong>Lesson #1: Before you spend compute, ask a human.</strong></p><h3>Project 2: “Hey Siri” and Trigger Words</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BaiiJyMzHSy5ApxgmTQwrw.png" /><figcaption>Cheap models guard expensive ones</figcaption></figure><p>Your smart speaker isn’t one model. It’s a cascade:</p><ol><li>A tiny model runs all the time and only checks whether anyone is talking.</li><li>If someone is, a second model listens for one specific word, like “Alexa” or “Hey Siri.”</li><li>Only then does the big, power-hungry model wake up to understand your request.</li></ol><p>This saves a lot of energy. The class project is the second model: given a 10-second audio clip, detect the word “activate.” (Fun fact: “activate” would be a bad trigger word in real life. You want a word people rarely say in normal conversation, or your speaker keeps waking up.)</p><h4>Collect diverse data</h4><p>The class listed what the data needs to cover: different accents, ages, speaking speeds, male and female voices, and background noise. Each one changes the sound frequencies the model sees. Prof Kian’s first version didn’t work for his German friends, and a model trained on a quiet campus might fail next to the New York subway.</p><p>For technical settings like the audio sample rate, you don’t need to be a speech expert. Find a similar speech project on GitHub and reuse what works for human voices.</p><h4>The labeling experiment</h4><p>Prof played three short clips in Italian and told the class that two of them contained a mystery word. Find it. Almost nobody could (only an Italian speaker got it). Then he played them again, this time showing exactly where in each clip the word appeared. Now people who don’t speak Italian could hear it: <em>pomeriggio</em>, Italian for “afternoon.”</p><p>If it’s easier for you, it’s easier for the model.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zE0grM311d1fgSpTHTbdeQ.png" /><figcaption>Label smarter, not harder</figcaption></figure><p><strong>Option A</strong> labels each whole clip with one number: 1 if the word is somewhere in it. Easy to label, but the model has to figure out on its own which part of the clip matters. Kian estimated it would need about 1,000 times more data.</p><p><strong>Option B</strong> labels the clip moment by moment: 0 while the word isn’t being said, 1 while it is. Barely harder to label, and the model learns much faster.</p><p>One more trick. If only a single moment is labeled 1, a lazy model can predict 0 everywhere and still be 99.9% “accurate.” So you label several moments right after the word as 1, which balances things out. The model makes a prediction (a sigmoid) at every moment, and the loss is binary cross-entropy applied at each one.</p><h4>Make millions of examples in three hours</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OzZUWettwJjzBchVAYAQqg.png" /><figcaption>The data recipe</figcaption></figure><p>Here’s the recipe Prof Kian’s team used:</p><ol><li>Record lots of people saying “activate” (positive words).</li><li>Record other words too, including tricky near-misses like “deactivate” (negative words).</li><li>Grab background noise. It’s basically free online.</li><li>Write a Python script that drops random words onto random background clips. Because the script placed every word, it knows exactly where “activate” is and labels everything automatically.</li></ol><p>Add small changes like speeding clips up, slowing them down or shifting frequencies (this is called <strong>data augmentation</strong>), and a few hours of recording turns into millions of training examples. The <strong>test set</strong>, the data you use to judge the final model, must still be real, hand-labeled audio, because real audio is what you actually care about.</p><p>Two more lessons from this project: learn how data licenses work before you collect anything from the internet, and ask experts early. A senior PhD student told Kian which architecture would fail before he even tried it, which saved him about a month.</p><p><strong>Lesson #2: Your labeling strategy can matter more than your model.</strong></p><h3>Project 3: Face Verification</h3><p>A school wants a camera at the gym entrance. You swipe your student ID, the camera takes your photo, and the system checks that it’s really you.</p><h4>Why comparing pixels fails</h4><p>The obvious idea is to compare the two photos pixel by pixel. It breaks immediately. Different lighting changes every pixel. Move your head slightly and nothing lines up. Glasses, a hat, a new beard or being a few years older than your ID photo all throw it off.</p><p>This task also needs a much higher resolution than day or night, around 412 × 412, because small details like eye color matter.</p><h4>Compare encodings</h4><p>So you run both photos through a network trained on faces and compare their encodings (say, 128 numbers each). If the distance between them is below a threshold, it’s the same person.</p><p>Where you set that threshold is a product decision. An airport would be much stricter than a dining hall, accepting more false rejections in exchange for fewer false matches.</p><h4>Training it: the triplet loss</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7klGEfWVZs08XvsCZdSf7w.png" /><figcaption>Triplet loss</figcaption></figure><p>How do you train a network so the same face always gets similar numbers? You design the loss function around it. Each training example is a <strong>triplet</strong>:</p><ul><li><strong>Anchor:</strong> a photo of a person</li><li><strong>Positive:</strong> another photo of the same person</li><li><strong>Negative:</strong> a photo of someone else</li></ul><p>The loss rewards the network when the anchor ends up closer to the positive than to the negative, by at least a small margin. The margin stops the network from cheating by giving every face the same numbers.</p><p>Repeat that millions of times and the network learns what makes a face yours, without anyone coding up eyes or ears. At test time there are no triplets: you just compare the ID photo with the camera photo. This approach comes from the FaceNet paper (2015).</p><p><strong>Lesson #3: Great deep learning engineers are creative with loss functions.</strong></p><h4>One encoder, three products</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7nj847hb78ov9J3vFtzFcA.png" /><figcaption>One encoder, three jobs</figcaption></figure><ul><li><strong>Verification:</strong> “Are you who your card says?” Compare two encodings.</li><li><strong>Identification:</strong> “Who is this?” No card needed. Store an encoding for every student, then find the stored encodings closest to the camera photo (an algorithm called K-nearest neighbors). If nothing is close enough, the person isn’t in the database. Think Global Entry at the airport.</li><li><strong>Clustering:</strong> “Group all my photos of Mom.” Run an algorithm called K-Means on the encodings and each group becomes a person. Your phone keeps a center point for each person and files new photos under the closest one.</li></ul><h3>Self-Supervised Learning: No Labels Needed</h3><p>Everything so far needed labels, and labels are expensive. So what if all you have is billions of unlabeled photos?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KV2aYSB9f4pBah5vWmB93g.png" /><figcaption>No labels needed</figcaption></figure><p>You let the data label itself. Take a photo and make a copy that’s rotated, cropped or partly covered. A human knows it’s still the same dog, so you train the network to give both versions the same numbers while pushing different photos apart. This is <strong>contrastive learning</strong> (SimCLR is a famous example), and it’s how modern models learn from billions of unlabeled images.</p><p>Text works the same way. Hide the next word in a sentence and make the model guess it. That’s <strong>next-word prediction</strong> (technically next-token prediction), the core of GPT.</p><p>The magic is what emerges from that simple game:</p><ul><li>“I poured myself a cup of ___” teaches what goes together: coffee, tea and water are all liquids that fit in a cup.</li><li>“The capital of France is ___” teaches facts.</li><li>“She unlocked her phone using her ___” teaches meaning: face, fingerprint and password all unlock things.</li><li>“If it’s raining, I should bring an ___” teaches reasoning.</li></ul><p>Nobody labeled any of it. Skills that show up without being explicitly taught are called <strong>emergent behaviors</strong>. The same trick works for audio (hide a few seconds), video (hide some frames) and even biology (hide part of a protein or DNA sequence).</p><h3>Weakly Supervised Learning: Free Pairs from the World</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UBQtjcGJg506mk6YS1Y-lw.png" /><figcaption>Text is the hub</figcaption></figure><p>To connect different types of data, like text and images, you need pairs. But you don’t have to create them yourself. The world already has: photos with captions, movies with subtitles, videos with their own audio, songs with titles, gameplay footage with keyboard presses. Learning from these naturally occurring pairs is called <strong>weakly supervised learning</strong>.</p><p>Train on them and everything lands in one shared space, where the word “drums,” the sound of drums and a photo of drums sit close together. Meta’s ImageBind showed you don’t even need every pair: most types of data connect to images or text, which act as the hub. Give it the sound of a dog barking plus a photo of a beach, and it can find a dog on a beach.</p><h3>My Biggest Takeaway</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ftmh4fQfGmGo5chdFrnX1w.png" /><figcaption>Eureka!!</figcaption></figure><p>Most of the important decisions in an AI project happen before any training: defining the task, choosing the resolution, designing the labels and building the data. Many of them can be tested in ten minutes with a friend and a printer.</p><p>When in doubt, ask a human first and an expert second.</p><p>That’s Lecture 2! Next up: Lecture 3, the full cycle of a deep learning project. Follow along so you don’t miss it.</p><p><em>Credits: All ideas in this post come from Stanford CS230 (Autumn 2025), Lecture 2: Supervised, Self-Supervised and Weakly Supervised Learning, taught by Professor Kian Katanforoosh with Professor Andrew Ng, and published by Stanford Online. Watch the full lecture: </em><a href="https://www.youtube.com/watch?v=DNCn1BpCAUY"><em>https://www.youtube.com/watch?v=DNCn1BpCAUY</em></a><em>. All illustrations are my own.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9be21102a92b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Your First Hour With Neural Networks: Deep Learning Specialization, Course 1, Week 1, in Pictures]]></title>
            <link>https://medium.com/@bharatigehani/your-first-hour-with-neural-networks-deep-learning-specialization-course-1-week-1-in-pictures-bb02c221f221?source=rss-5a2b76ae2846------2</link>
            <guid isPermaLink="false">https://medium.com/p/bb02c221f221</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Bharati Gehani]]></dc:creator>
            <pubDate>Tue, 29 Sep 2026 08:06:36 GMT</pubDate>
            <atom:updated>2026-09-29T08:06:36.946Z</atom:updated>
            <content:encoded><![CDATA[<h4>No maths, no code, no details left unexplained. If you have never touched AI in your life, start here. This is the whole of Week 1, drawn out.</h4><p>I am working through the Deep Learning Specialisation, the five-course series taught by Professor Andrew Ng that a very large number of people have used as their way into this field.</p><p>This is Course 1, Week 1. It is the gentlest week of the whole series, and its job is to answer three questions: what is a neural network, what are these things actually used for, and why did all of this suddenly start working in the last fifteen years when the ideas are decades old.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EN3MPQ4kEwukzOM1Y-ONxg.png" /></figure><p><em>Week 1 in one picture</em></p><h3>Part 1: What Is a Neural Network?</h3><h4>Start with something small</h4><p>Forget neural networks for a moment. Imagine you have data on six houses. For each one you know two numbers: how big it is, and what it sold for.</p><p>You want to guess the price of a seventh house that has not sold yet. You know its size. What do you do?</p><p>The obvious thing is to draw the six houses as dots on a chart, with size along the bottom and price up the side, and then draw a line through them. Find where your seventh house sits on that line, read off the price, done.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jF0IOR_HcMBb4jzIir2XaQ.png" /></figure><p><em>Six dots and a line through them</em></p><p>That is already a prediction machine. It is a crude one, but it has the shape of every single thing that follows in this course: something goes in, something comes out, and the rule connecting them was worked out from examples rather than written by hand.</p><h4>The line has to bend in one place</h4><p>There is one problem with a straight line. Extend it far enough to the left and it drops below zero, which says a very small house sells for a negative amount of money. Nobody pays you to take their house.</p><p>So the real shape is not a straight line. It is a line that sits flat on zero for a while and then rises.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*E5tRd_-LrTlJjYVEDhPmlA.png" /></figure><p><em>Flat at zero, then rising. This shape has a name.</em></p><p>That bent shape turns up so often in this field that it has a name. It is called <strong>ReLU</strong>, which stands for rectified linear unit. Three words that sound much harder than the picture. All it means is: below a certain point the answer is zero, and above that point the answer grows steadily. “Rectified” is an old electrical engineering word for “we cut off the negative part”.</p><p>Remember that name. It comes back in Part 3 and it turns out to be one of the reasons this whole field took off.</p><h4>Now we can say the word neuron</h4><p>What you just built, a box that takes the size of a house, applies that bent rule, and hands back a price, is a single <strong>neuron</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*T1vTpNDqFUDJhW5_bcPGyQ.png" /></figure><p><em>One neuron: something in, something out, a rule in the middle</em></p><p>That is all a neuron is. It is not a small brain. It is not alive. It is a box that takes numbers in, combines them, applies one simple bend, and passes a number on. The word comes from a loose comparison with brain cells that was made decades ago, and the comparison has not aged well. Professor Andrew says as much later in the course. Treat “neuron” as the name of a very small piece of arithmetic and you will be closer to the truth than most people who use the word.</p><p>A neural network is what you get when you wire a lot of these boxes together.</p><h4>One house has more than one number</h4><p>Size is not the only thing that sets a price. There is the number of bedrooms. There is the postcode. There is how walkable the area is and how good the local schools are.</p><p>So feed in four numbers rather than one, and put a few neurons in the middle.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*br4bfz3kzFKMP1S3MtyEgg.png" /></figure><p><em>Four things about the house go in. A price comes out.</em></p><p>The numbers you feed in have a name. Collectively they are called <strong>x</strong>, and each individual one is a <strong>feature</strong>. The thing that comes out is called <strong>y</strong>. Those two letters are used everywhere in machine learning, so it is worth fixing them now: <strong>x is what you know, y is what you want to know</strong>.</p><p>The column of boxes in the middle, the ones that are neither the input nor the answer, is called a <strong>hidden layer</strong>. Hidden only because you never see its numbers. You see what goes into the network and what comes out. The middle is internal bookkeeping.</p><h4>Here is the part that surprised me</h4><p>When people draw this diagram, they like to label the middle boxes with tidy human ideas. This one handles family size, that one handles walkability, the third one handles school quality.</p><p>You do not get to do that. Nobody tells the middle boxes what to be.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DQSKAIurm3rhRfZUMevAAQ.png" /></figure><p><em>You connect everything to everything and let it sort itself out</em></p><p>What actually happens is this. You connect every input to every box in the middle. You show the network a pile of houses where you already know both the features and the real selling price. The network makes a guess, sees how wrong it was, and nudges its internal numbers slightly towards being less wrong. Then it does that again, and again, thousands of times.</p><p>The middle boxes end up computing something useful, because useful is what reduces the error. But what they compute is theirs, not yours. Sometimes a box lines up with an idea a human would recognise. Often it does not line up with anything you could name.</p><p>Two things are worth being precise about here, because this is where beginners get the wrong idea.</p><p>First, the network is <em>not</em> figuring out the answer on its own. You handed it the real prices. That is how it knew it was wrong. Without those answers it has nothing to aim at.</p><p>Second, you gave it no instructions about the middle. So the arrangement is: you supply the questions and the answers, and the network works out the route between them.</p><p>That arrangement has a name, and it is the whole of Part 2.</p><h3>Part 2: Supervised Learning, and What It Is Good For</h3><h4>The name makes more sense than it sounds</h4><p><strong>Supervised learning</strong> means you train on examples where you already know the right answer.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6fw9xmRGE9WD6GkOSYYpaw.png" /></figure><p><em>Every training example is a pair: a question and its answer.</em></p><p>Every training example is a pair. Here is the house, here is what it sold for. Here is the photo, here is what is in it. Here is the audio clip, here is what was said. The “supervision” is the answer key. Somebody had to produce it, which is why data is expensive and why so much of real AI work is about getting hold of good answer keys.</p><p>This matters because almost everything that makes money in AI today is supervised learning. Professor Andrew is direct about it: this is where the economic value is.</p><h4>What people actually build with it</h4><p>Four examples, each with its x and its y.</p><p><strong>Online advertising.</strong> x is information about a user and an ad. y is whether they will click it. This is a boring problem that is worth an enormous amount of money, because getting slightly better at it multiplies across billions of ad views.</p><p><strong>Photo tagging.</strong> x is an image. y is which of a thousand objects it contains.</p><p><strong>Speech recognition.</strong> x is a few seconds of audio. y is the words that were spoken.</p><p><strong>Machine translation.</strong> x is a sentence in English. y is the same sentence in Hindi.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WM8Y5DCTpinLR9jBvPkSyA.png" /></figure><p><em>Four jobs, four different shapes of x and y</em></p><p>And one more that is different in kind: <strong>self-driving cars</strong>. x is a camera image plus readings from a laser scanner, and y is the position of every other car on the road. Different shapes of input, mixed together, feeding one answer.</p><h4>Two kinds of data, and why the difference matters</h4><p>Look at those examples again and you will notice they split into two groups.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xjFL06omYURFY3EcX2j2Yg.png" /></figure><p><em>A spreadsheet on one side. A photo, a sound, a sentence on the other.</em></p><p><strong>Structured data</strong> is data in a table. Rows and columns, each column clearly labelled and meaning one thing. Size of house. Number of bedrooms. Age of user. Every cell is a number or a category that somebody chose to record on purpose. Most databases in the world look like this.</p><p><strong>Unstructured data</strong> is everything else. Images, audio, text. A photo is not a table. It is a grid of millions of brightness values, and no single one of those values means anything on its own. The meaning is in the pattern across them.</p><p>People find unstructured data easy and computers find it hard, and for most of computing history that gap was enormous. A person can glance at a photo and say “cat” without effort. Writing rules that do the same thing by hand turned out to be close to impossible.</p><p>That gap is what closed. The headline-grabbing results of the last decade are mostly computers getting good at unstructured data, which is why the news is full of image and speech and language results. But the structured data side is where a great deal of the commercial value sits, because that is the shape most business data already has.</p><h4>Different jobs want differently shaped networks</h4><p>You do not use the same wiring for every problem.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_MAdhYydDlfmn9EWN88glg.png" /></figure><p><em>Three shapes for three kinds of problems</em></p><p>For a table of house features, the plain arrangement from Part 1 is fine. Everything connects to everything.</p><p>For images you use a <strong>CNN</strong>, short for convolutional neural network. The important idea is that it looks at small patches of the picture at a time rather than the whole thing at once, which suits images because a pattern in the top left corner is the same pattern in the bottom right corner and you would rather learn it once.</p><p>For anything that arrives in a sequence, one thing after another, you use an <strong>RNN</strong>, short for recurrent neural network. Audio is a sequence. So is a sentence, which is why translation uses these. The important idea is that the network keeps a little memory of what it has already seen, so the word it reads now can be understood in the light of the words before it.</p><p>Those two names are all you need for now. Course 4 is the CNN course. Course 5 is the RNN course. You will meet them properly there.</p><h3>Part 3: Why Did This Start Working Now?</h3><p>Here is the fact that makes this part interesting. The core ideas behind neural networks are decades old. Most of the maths was worked out long before anybody had a product to show for it.</p><p>So what changed?</p><h4>The chart that explains the decade</h4><p>Draw a chart. Along the bottom, how much labelled data you have. Up the side, how well the thing performs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wdQs8INVuzWC2gyxgzlH1A.png" /></figure><p><em>Four lines. Only three of them keep climbing.</em></p><p>Now draw four lines on it.</p><p>The bottom line is the older generation of methods. It climbs quickly at first and then goes flat. Give it ten times more data and it barely improves. It has run out of capacity to absorb what you are giving it.</p><p>Above that, three more lines: a small neural network, a medium one, and a large one. Each starts out unremarkable and keeps climbing long after the first line has given up. The larger the network, the longer it keeps climbing.</p><p>Two things follow from that picture, and they are the whole reason this field looks the way it does today.</p><p>At the left edge, where data is scarce, the lines are tangled and it genuinely does not matter much which method you pick. Skill and luck decide it.</p><p>At the right edge, where data is abundant, the order is stable. Big networks win, and they keep winning by more.</p><p>So when we digitised everything, when phones and websites and sensors started producing records of everything anybody did, we walked off the left edge of that chart and out to the right. And out there, the method that had been quietly waiting for decades was the one that pulled ahead.</p><h4>Three things had to line up, not one</h4><p>Data alone was not enough.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*oc4H6FBIu69Enx1AGAePNA.png" /></figure><p><em>Data, computation, algorithms. All three or none.</em></p><p><strong>Data.</strong> Digitisation gave us the records. A large network with a small dataset is worse than useless, so this had to come first.</p><p><strong>Computation.</strong> Training a big network means doing an enormous amount of arithmetic over and over. Faster chips, and specifically GPUs, which are chips originally built to draw video game graphics and which happen to be very good at exactly the kind of arithmetic a neural network needs, made the training time drop from impossible to overnight.</p><p><strong>Algorithms.</strong> Better methods, many of which exist purely to make training faster rather than to make the final result cleverer.</p><p>That last point deserves its own section, because there is a lovely example of it and it involves the bent line from Part 1.</p><h4>One small change that sped up everything</h4><p>For years, the standard bend used inside neurons was not ReLU. It was a smooth S-shaped curve called a <strong>sigmoid</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*a0djSqbnFwgE4KZ7pvZYvw.png" /></figure><p><em>The S curve goes flat at both ends. The bent line does not.</em></p><p>Look at the S curve. At the far left and the far right it flattens out almost completely. In those flat regions, changing the input barely changes the output.</p><p>Now remember how a network learns. It makes a guess, measures how wrong it is, and nudges its internal numbers in whatever direction reduces the error. The size of the nudge depends on how steep things are locally. That method is called <strong>gradient descent</strong>, and the picture behind the name is walking downhill: you feel which way the ground slopes and you step that way.</p><p>On a flat stretch there is no slope to feel. So a neuron that has wandered out to the flat end of the S curve learns almost nothing. Progress crawls.</p><p>The bent ReLU line has a constant slope everywhere on its rising side. No flat stretch to get stuck on. Swap sigmoid for ReLU and gradient descent runs dramatically faster, with nothing else changed.</p><p>That is it. That is the whole change. And it is one of the genuine reasons the last decade happened.</p><h4>Why speed is the real prize</h4><p>It would be easy to read that and think faster training is a nice convenience. It is more than that.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6bWTke7aWyThIv5yy2ySag.png" /></figure><p><em>Idea, code, experiment, back to idea</em></p><p>Building anything in this field is a loop. You have an idea. You write the code. You run the experiment. The result tells you what to try next, and round you go.</p><p>If one turn of that loop takes a month, you will get roughly twelve ideas tested in a year, and you will be cautious about which ones you bother trying. If a turn takes ten minutes, you will test hundreds, including the strange ones, and the strange ones are frequently where the good results hide.</p><p>So faster computation does not only buy you the same work sooner. It changes how boldly you are willing to think. Professor Andrew makes this point deliberately, and it is the part of Week 1 I have thought about most since.</p><h3>What Comes Next</h3><p>Week 1 is the only week with no maths in it. From here the specialization builds up in a deliberate order.</p><p>Course 1, the one this week belongs to, covers neural networks and deep learning: you build one from scratch, starting with a single neuron and stacking up from there. Course 2 is about making them actually work, which is the practical craft of tuning and debugging. Course 3 is about how to structure a machine learning project and decide what to work on. Course 4 is images and CNNs. Course 5 is sequences and RNNs, which is the road towards language models.</p><p>Weeks 2 and 3 of this course are where the notation starts, and I will draw those the same way.</p><h3>My Takeaway</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5sAQ-NnNfwl1nRlE-UboPg.png" /></figure><p><em>The ideas waited. The world caught up.</em></p><p>The thing I did not expect from Week 1 is that the story of deep learning is not really a story about a brilliant new idea. The ideas were sitting there. What changed is that the world started producing enough data to feed them and enough computing power to train them, and then a handful of unglamorous engineering improvements made the loop fast enough that people could iterate their way forward.</p><p>There is something reassuring in that if you are starting out. You are not late to a field that was solved by geniuses in a room. You are early in a field that is still mostly about trying things carefully and paying attention to what breaks.</p><p>Next up: Course 1, Week 2.</p><p><em>Credits: Deep Learning Specialization, Course 1 (Neural Networks and Deep Learning), Week 1, taught by Professor Andrew Ng, on Coursera and DeepLearning.AI. All illustrations are my own.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bb02c221f221" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Paper That Made the World Bet Billions on AI, in Pictures]]></title>
            <link>https://medium.com/@bharatigehani/the-paper-that-made-the-world-bet-billions-on-ai-in-pictures-fc763ffc4c5e?source=rss-5a2b76ae2846------2</link>
            <guid isPermaLink="false">https://medium.com/p/fc763ffc4c5e</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[scaling-law]]></category>
            <category><![CDATA[cs230]]></category>
            <dc:creator><![CDATA[Bharati Gehani]]></dc:creator>
            <pubDate>Wed, 23 Sep 2026 11:50:40 GMT</pubDate>
            <atom:updated>2026-09-23T11:50:40.714Z</atom:updated>
            <content:encoded><![CDATA[<p>In 2017, nine researchers at Baidu measured exactly how much better AI gets when you feed it more data. The answer came out as a straight line, and that line is why data centres are being built right now.</p><p>While <a href="https://www.linkedin.com/in/bharati-g-82372b192/?isSelfProfile=true">I </a>was drawing Lecture 1 of Stanford’s CS230, Professor Andrew pointed at a moment in AI history I had never really thought about.</p><p>Everybody has always believed that more data makes models better. What nobody could do, before 2017, was say <em>how much</em> better. Not roughly. Not at all.</p><p>Then a team at Baidu Research published a paper with a wonderfully flat title: <em>Deep Learning Scaling is Predictable, Empirically</em>. It is not a paper about a clever new model. It is a paper about measuring something obvious, carefully, and finding that it has a shape.</p><p>That shape is why companies are comfortable spending billions on compute. So I drew it.</p><p><strong>You need no background for this.</strong> Every term gets explained in ordinary words before it is used. If you have never trained a model in your life, you will still be able to follow the whole thing.</p><h4>Three words to get out of the way first</h4><p>You will meet three words repeatedly, so let me define them once, properly.</p><p><strong>A model</strong> is a program that learns a task from examples rather than from rules somebody writes down. You show it thousands of photos, each one already labelled “cat” or “dog” by a human, and it works out which patterns in the pixels go with which label. Nobody ever tells it “a cat has pointed ears and whiskers.” It is given the right answer for every training photo; what it finds on its own is the rule connecting the picture to that answer. Learning from labelled examples this way is called <em>supervised</em> learning, and it is what every task in this paper does.</p><p><strong>Training data</strong> is that pile of examples. More examples generally mean a better model, which is the belief this whole paper is about to test.</p><p><strong>Error</strong> is simply how often the model is wrong on examples it has never seen before. Lower is better. Every chart in this article has error going down the vertical axis, so <strong>down the page means the model is getting better</strong>.</p><p>That is the vocabulary. Here is the whole paper in one drawing, which will make more sense after you have read the rest, so feel free to come back to it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gGAwNJ73UpSLQV0CMWuVXQ.png" /><figcaption>The whole paper in one picture. It will make more sense once you reach the end.</figcaption></figure><h3>The Question Nobody Could Answer</h3><p>Imagine you run an AI team in 2016. Your model gets the right answer 80% of the time, and your boss wants 95%.</p><p>You know you probably need more data. But how much more? Twice as much? Ten times? A hundred times? Is 95% even reachable for this problem, or have you already hit some ceiling you cannot see? Should the company spend six months and a lot of money collecting more examples, or would those six months be wasted?</p><p>Nobody could answer that in 2016. Not because people were not clever, but because nobody had ever sat down and measured the relationship. Every AI project was a guess dressed up as a plan.</p><p>So the Baidu team went and measured it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eT3QU5dcs9rCGf2Pqb_pmw.png" /><figcaption>The question every AI team had in 2016, and nobody could answer it.</figcaption></figure><h3>The Method Is Almost Boring</h3><p>Here is the entire experiment, and it is worth noticing how unglamorous it is.</p><p>Take one task. Now deliberately throw away almost all of your training data and keep only a tiny slice of it, say one example in every thousand. Train a model on that slice. Measure how wrong it is.</p><p>Now go back and keep a bigger slice: one in a hundred. Train again, from scratch. Measure again. Then one in ten. Then half. Then all of it.</p><p>You end up with a handful of dots. Each dot says: <em>this much data went in, this much error came out.</em> Plot the dots.</p><p>That is the whole method. There is no new algorithm in it. The surprise is not how they measured. The surprise is how tidy the answer turned out to be.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WXr_7Q4NxwvWvir4X4VlYQ.png" /><figcaption>Train on a slice, measure the error, repeat with a bigger slice.</figcaption></figure><h3>Why the Charts Have Strange Axes</h3><p>This is the part that trips most people up, so let me go slowly, because once it clicks the rest of the paper is easy.</p><p><strong>An ordinary ruler adds. Every step along it is the same amount bigger than the last: 10, 20, 30, 40, 50. The gap between 10 and 20 takes up exactly as much room on the page as the gap between 40 and 50.</strong></p><p><strong>A log ruler multiplies. Every step along it is the same amount times bigger than the last: 1, 10, 100, 1,000, 10,000. Here the gap between 1 and 10 takes up exactly as much room on the page as the gap between 1,000 and 10,000. Both are one ten-fold step, so both get one step of space.</strong></p><p>That sounds strange written down, but you already think this way about money. Going from 1,000 rupees saved to 10,000 saved feels like roughly the same size of achievement as going from 10 lakh to 1 crore. In plain subtraction, those are nothing alike, 9,000 against 90 lakh, but both are the same tenfold jump, and that is the part that feels equivalent. A log ruler is simply a ruler that agrees with that instinct.</p><p>Why this paper needs one: their smallest experiments are trained on a few thousand examples, and their largest on hundreds of millions, a range of about a hundred thousand to one. On ordinary graph paper, if you scaled the page so the largest experiment fit, every small experiment would be crushed into a single dot in the corner, sitting on top of each other and on top of zero. You would see nothing. On a log ruler, each ten-fold jump gets its own equal slice of the page, so every experiment is visible at once. The paper does this on both axes, the data along the bottom and the error up the side. That is all &quot;log-log&quot; means. Two multiplying rulers.</p><p>And here is the payoff, which is the reason the paper exists. On ordinary axes these results look like a vague curve that drops steeply and then flattens out, and a curve like that tells you where you have been but not where you are going. On log-log axes, the very same numbers fall on a dead straight line.</p><p>Why does that matter so much? Because a straight line is something you can lay a ruler against and continue. It turns &quot;more data helps, somehow&quot; into &quot;Give me ten times the data and the error drops by this exact percentage, and it will do it again the next time, and the time after that.&quot;</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VD9voI9edOvMO16-kGsrPg.png" /><figcaption>The same measurements twice: a vague curve on ordinary axes, a straight line on log-log.</figcaption></figure><p><em>Same experiment, same numbers, drawn twice. On the left, ordinary axes, and it looks like a vague curve that flattens out. On the right, log-log axes, and the same dots fall into a line you could extend with a ruler.</em></p><h3>Every Curve Has Three Parts</h3><p>The line is not straight forever. It is straight in the middle. The paper splits every learning curve into three stretches, and knowing which stretch you are standing in tells you what to do next.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CRyae1xXA8ucyCjiIoO2oA.png" /><figcaption>Three stretches: too little data to learn, the straight middle, and the floor.</figcaption></figure><p><strong>1. The small-data region, at the start.</strong> There are not enough examples yet for the model to learn anything real, so it does about as well as random guessing. For image classification, they found this stretch ended at roughly 25 pictures per category. A category is one of the labels you are trying to tell apart, so if you are sorting photos into cat, dog and horse, that means about 25 photos of each. Give it fewer than that per category, and it simply has too little to go on: there is not enough evidence in the pile for any rule to emerge, however long you train.</p><p><strong>2. The power-law region, in the middle.</strong> This is the straight stretch, and this is where nearly every real project lives. Here, each new example genuinely teaches the model something that helps it on examples it has never seen. Multiply your data by ten, and the error drops by a predictable percentage. Multiply by ten again, and it drops by that same percentage again. That steady, repeatable behaviour is what mathematicians call a <strong>power law</strong>. You and I can call it a rule you can extend.</p><p><strong>3. The floor, at the end.</strong> This is the one people find hardest, so here is what it actually means.</p><p>Every task contains some amount of noise that no model can ever get past, because the answer is genuinely not present in the input. In speech recognition, two people really do pronounce “there” and “their” identically, and no amount of audio will tell them apart. In photo labelling, some pictures are blurry, or contain two animals, or were labelled wrongly by a tired human in the first place.</p><p>Those cases are permanently unwinnable. So the error line falls, and falls, and then stops falling and goes flat. That flat bottom is the <strong>floor</strong>: the best any model can ever do on this task, no matter how much data you buy.</p><p>Knowing roughly where your floor sits is the difference between a sensible budget and an expensive fantasy. If you are already close to it, more data is money set on fire.</p><h3>They Tried Four Completely Unrelated Fields</h3><p>This is the part that turns an interesting observation into a finding people are willing to spend money on.</p><p>They ran the whole exercise four separate times on four problems that have almost nothing to do with each other. Different data, different kinds of models, and different sub-teams inside Baidu.</p><p><strong>Machine translation.</strong> Feed it a sentence in English, and get the same sentence in French. The model has to handle grammar, word order and meaning all at once.</p><p><strong>Language modelling.</strong> Given the words so far, predict the next word. This sounds like a party trick, and it is the direct ancestor of ChatGPT and every other chatbot you have used.</p><p><strong>Image classification.</strong> Show it a photograph, and get back a label: dog, aeroplane, teapot.</p><p><strong>Speech recognition.</strong> Feed it recorded audio of somebody talking; get back the words they said. This is what your phone does when you dictate a message.</p><p>Four problems, different types of input: text, text, pictures, and sound. And every single one of them came out as a straight line in the middle region.</p><p>One field behaving this way is a curiosity. Four unrelated fields behaving this way starts to look like a property of learning itself.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Oo7Da1xIwuPftGNG0fhFgQ.png" /><figcaption>Translation, language modelling, image labelling, speech. Four fields, four straight lines.</figcaption></figure><h3>The Slope Is the Whole Story</h3><p>All four lines were straight, but they were not all at the same angle, and the angle turns out to be the useful number.</p><p>The angle is called the <strong>exponent</strong>, and it is written as a negative number because the line goes downward. Think of it as a price tag on data:</p><p>A <strong>steep</strong> line means data pays off fast. Ten times more data buys you a big drop in error. Data is good value for this problem.</p><p>A <strong>shallow</strong> line means data barely helps. You can multiply your data by ten and watch the error move hardly at all. Data is expensive and disappointing for this problem.</p><p>Here is what they measured, roughly. Remember, further from zero means steeper, which means better value:</p><ul><li><strong>Language modelling: about -0.07.</strong> Painfully shallow.</li><li><strong>Machine translation: about -0.13.</strong></li><li><strong>Speech recognition: about -0.30.</strong></li><li><strong>Image classification: about -0.31.</strong> The best value of the four.</li></ul><p>That first number deserves a moment of your attention. Language modelling, at -0.07, improves more slowly with data than anything else they tested. That means text models need <em>colossal</em> amounts of text before they get meaningfully better.</p><p>Written in 2017, that reads now like a prediction of exactly what happened. Everybody ended up training on something close to the entire public internet. Not because anyone was showing off, but because for this particular problem, that is what the slope demands.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*k27TKNUIuLroidAXT7webw.png" /><figcaption>All straight, but not all equally steep. The angle is the price tag on data.</figcaption></figure><h3>The Part Nobody Expected</h3><p>Here is where the paper gets genuinely interesting and unusually honest.</p><p>Before anybody measured this, existing mathematical theory had a prediction for what the exponent should be. Theory said around -0.5 and in some treatments -1.0. Steep. Data should be an excellent value.</p><p>Reality came in between -0.07 and -0.35. Much shallower than predicted, across all four fields. Real models improve far more slowly with data than the mathematics said they should.</p><p>And the authors do not paper over it. They write plainly that these exponents are yet to be explained by theoretical work, and they call the gap an opportunity rather than an embarrassment.</p><p>It is still open today. If you are a student hunting for a research problem, that is the most interesting sentence in the whole paper.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ksEBanmIIAKbhdyCqPMqPw.png" /><figcaption>Theory predicted about -0.5. Reality came in between -0.07 and -0.35.</figcaption></figure><h3>A Better Model Moves the Line Down, Not Sideways</h3><p>This one reorganised how I think about model design.</p><p>When the team swapped in a better architecture, meaning a cleverer internal design for the model, the whole line shifted <strong>downward</strong>: less error at every single data size. Good.</p><p>But the line ran at almost exactly the same <em>angle</em> as before. The slope did not change.</p><p>Sit with what that implies. A cleverer design gives you a one-off head start. It does not change how quickly you learn from each new batch of data. The slope appears to belong to the <em>problem</em>, not to the model you point at it.</p><p>Which means the two things you can buy, a better idea and more data, do genuinely different jobs. A better idea lowers the line. More data moves you along it. You need both, and knowing which one you are short of is most of the skill.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JpbvAxH9P93ujkx_KdptTQ.png" /><figcaption>A better design lowers the line everywhere without changing its angle.</figcaption></figure><h3>The Model Grows More Slowly Than the Data</h3><p>One more practical finding, and this one is about the bill.</p><p>The obvious assumption is that ten times the data needs a model ten times bigger to absorb it. They measured that too, and it is not true.</p><p>Model size follows its own power law, with exponents between roughly 0.57 and 0.92 across the four fields. All of them below one. In plain terms: ten times the data needs something closer to four times the model, not ten times.</p><p>That is excellent news for anybody paying for the hardware. Your compute cost grows more slowly than the pile of data you are feeding it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JHEoo7bSFTj6uqECmXj51g.png" /><figcaption>A curve peeling away from the line early is a signal that something is wrong.The winner at full scale is already the winner on a slice you can afford.</figcaption></figure><h3>What You Can Actually Do With This</h3><p>Here is the payoff, and it is the reason this paper matters far outside research.</p><p>Train three small models on slices of data you can already afford. Plot your three dots on log-log axes. Fit a line through them. Now extend that line with a ruler to the accuracy you want, and read off how much data it would take to get there.</p><p>You have just turned “we think more data will help” into “we need roughly this much data and this much compute, and it will cost this much.” One is a hope. The other is a plan a finance department can approve.</p><p>It also tells you when to walk away, which is just as valuable. If the line says a hundred times more data buys you two percentage points, you have your answer, and you found it out in a week rather than in a year.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*I2SieeFrjvMVm7jkoXezSg.png" /><figcaption>Three cheap runs, a line through them, and a number you can budget against.</figcaption></figure><h3>And It Tells You When Something Is Broken</h3><p>Once you know what the curve is supposed to look like, a curve that does something else becomes a diagnostic tool.</p><p>If your results start peeling away from the straight line earlier than they should, that is a signal worth chasing. Your training may not be settling properly. Your data may have quality problems. Or you really have reached the floor for this task, and the flattening is honest.</p><p>Before this paper, a flattening curve was just disappointing. After it, it is a clue.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FjROO_-DcDP12QdsH1Rqsw.png" /><figcaption>A curve peeling away from the line early is a signal that something is wrong.</figcaption></figure><h3>One Last Gift, If You Do Not Have a Data Centre</h3><p>Follow the logic of those two parallel lines through to its conclusion.</p><p>If a better architecture shifts the line down at <em>every</em> data size, then the gap between two architectures is visible on small data too. The winner at full scale is already the winner on the slice.</p><p>So you can compare model ideas on an amount of data you can actually afford to run and reasonably trust that the ranking will hold when somebody scales it up. You do not need a data centre to find out which idea is better. You need one to find out how good the winner eventually gets.</p><p>For students and small teams, that is the difference between being able to do useful research and being locked out of it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dj1FtyGtukad3NO2RPdAQw.png" /><figcaption>The winner at full scale is already the winner on a slice you can afford.</figcaption></figure><h3>My Biggest Takeaway</h3><p>What strikes me most is that this paper did not invent anything.</p><p>It measured something everybody already believed, on four problems, carefully, and published the shape. No new architecture, no clever trick, no state-of-the-art result to brag about. Just a straight line, drawn honestly.</p><p>And that shape changed how an entire industry spends money. When a result is unpredictable, funding it is a gamble, and gamblers get small budgets. When it lands on a straight line you can extend with a ruler, it becomes a forecast, and forecasts get data centres.</p><p>The other thing I keep turning over is the honest hole in the middle of it. The measured exponents do not match what theory predicted, and years later, nobody has properly explained why. The most consequential empirical result in modern AI still rests on a foundation with a gap in it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*54SqIAs0NTF5p_DV49Jo4g.png" /><figcaption>What I am taking away from this paper.</figcaption></figure><p><em>This is a companion to my Stanford CS230 series, where I turn each lecture into a picture story so that anybody can learn AI without spending months on it. This paper is the one Professor Andrew points at in Lecture 1.</em></p><p><em>Credits: “Deep Learning Scaling is Predictable, Empirically”, Joel Hestness, Sharan Narang, Newsha Ardalani, Gregory Diamos, Heewoo Jun, Hassan Kianinejad, Md. Mostofa Ali Patwary, Yang Yang and Yanqi Zhou, Baidu Research, December 2017. Free to read at </em><a href="https://arxiv.org/abs/1712.00409"><em>arxiv.org/abs/1712.00409</em></a><em>. </em><strong><em>All illustrations are my own.</em></strong></p><p>Connect with me on LinkedIn at <a href="https://www.linkedin.com/in/bharati-g-82372b192/?isSelfProfile=true">Bharati Gehani</a> to learn every day.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fc763ffc4c5e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Professor Andrew Ng’s First CS230 Lecture, in Pictures]]></title>
            <link>https://medium.com/@bharatigehani/professor-andrew-ngs-first-cs230-lecture-in-pictures-4aaae6c80897?source=rss-5a2b76ae2846------2</link>
            <guid isPermaLink="false">https://medium.com/p/4aaae6c80897</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[genai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[cs230]]></category>
            <dc:creator><![CDATA[Bharati Gehani]]></dc:creator>
            <pubDate>Tue, 22 Sep 2026 08:26:20 GMT</pubDate>
            <atom:updated>2026-09-22T08:26:20.099Z</atom:updated>
            <content:encoded><![CDATA[<h4>Stanford CS230 (Autumn 2025), Lecture 1: why deep learning won, when prompting isn’t enough, and the career advice every builder needs to hear.</h4><p>Stanford put its entire CS230 Deep Learning course on YouTube. For free. Taught by Professor Andrew Ng and Kian Katanforoosh.</p><p>So I gave myself a challenge: watch every lecture and turn each one into a picture story. This is Lecture 1, the “why” lecture. No math yet, just the big ideas and some very honest career advice. Let’s go!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hu5VxJwzspNq9v-6BTE1Yg.png" /><figcaption>Lecture 1 in one picture</figcaption></figure><h3>Why Deep Learning Won</h3><p>Why has deep learning dominated AI for the last 10 to 15 years? One word. <strong>Scale.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fgaC-Dy_jFFXDhPTHq1Gmw.png" /><figcaption>Old algorithms get “full.” Neural networks keep eating.</figcaption></figure><p>Give traditional algorithms like logistic regression more data and their performance climbs for a while, then flattens. They just don’t know what to do with it. Neural networks behave differently: the bigger the network, the longer it keeps getting better as you feed it more data.</p><p><strong>Lesson #1: Deep learning won because it keeps eating data and keeps getting better.</strong></p><p>Fun fact: about 15 years ago, the first GPU machine Professor Andrew used to train neural networks was built by a Stanford undergrad in his dorm room. That undergrad was Ian Goodfellow (yes, the GANs guy).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*N9j73h5thuvQdjr86hjwZw.png" /><figcaption>From a dorm room to data centers</figcaption></figure><p>Then research out of Baidu showed that the gains from scaling are predictable, and OpenAI made the idea famous with its scaling laws paper. When something is predictable, you can invest in it. That’s a big reason companies are pouring money into massive data centers today.</p><h3>The AI Stack</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*O7GE8TthWKN3eykbWjxreQ.png" /><figcaption>Where deep learning sits</figcaption></figure><p>CS fundamentals at the bottom, then machine learning, then deep learning, then generative AI (mostly built on transformers) at the top. Two things I loved:</p><ul><li><strong>Deep learning and neural networks are the same thing.</strong> Neural networks existed for decades. “Deep learning” was simply a better brand. Who doesn’t want learning that’s deep?</li><li><strong>CS fundamentals still matter</strong>, even if you use AI coding tools every day.</li></ul><h3>When Prompting Isn’t Enough</h3><p>This was the most useful part for me as a PM. Professor Andrew prompts LLMs all the time, but he goes one layer deeper, into deep learning, in three situations:</p><ol><li><strong>Your data isn’t text.</strong> For images, audio, video and giant spreadsheets, he often works with deep learning models directly.</li><li><strong>Prompt tuning plateaus.</strong> Sometimes a month of tweaking prompts just won’t push performance any higher.</li><li><strong>Your AI bill explodes.</strong> LLM calls are cheap while you prototype. Then you hit product-market fit and the bill skyrockets. Fine-tuning smaller models bends that cost curve back down.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YFj5Pt_k6KIKPwTrmp42zQ.png" /><figcaption>Four signs you need to go one layer deeper</figcaption></figure><p>Also worth knowing: very few people in the world train giant frontier models. Most AI builders take a pre-trained model, engineer their own data and fine-tune it. That’s the skill set this course builds.</p><p><strong>Lesson #2: Prompting takes you far. Deep learning takes you the rest of the way.</strong></p><h3>The 5 Modules</h3><ul><li><strong>B: Basics.</strong> Build a neural network from scratch in raw Python.</li><li><strong>T: Tuning.</strong> Hyperparameters like learning rate and network size. Professor Andrew says this skill can decide whether you go home at 3 AM or 7 AM!</li><li><strong>S: Strategy.</strong> How to run an ML project with discipline (more below).</li><li><strong>C: ConvNets.</strong> For images and computer vision.</li><li><strong>S: Sequence models.</strong> For text and time series. Hello, transformers.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BL5YCWg5LlsWqSBsaOp_LQ.png" /><figcaption>A silly way to remember them: BTSCS, Big Teams Ship Cool Stuff!</figcaption></figure><h3>Discipline Beats Hype</h3><p>Less experienced teams pick what to work on almost at random. They read one article saying AI needs lots of data, then spend six months collecting data that was never going to help.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GLGSzlNGv2ESLfzP1pCqDQ.png" /><figcaption>Diagnose first, then pick the fix that matters</figcaption></figure><p>My favorite story from the lecture: a large family-run business bought a pile of GPUs. When Professor Andrew asked what they planned to do with them, the CTO pointed to his nephew, a college undergrad who “knows AI.”</p><p>Teams that diagnose first and then fix what actually matters move about 10x faster.</p><p><strong>Lesson #3: Don’t buy into the hype. Diagnose, then act.</strong></p><h3>Build Fast, Be Responsible</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Sh3879rSja8Agv1HPrg4BA.png" /><figcaption>Two buckets of software work</figcaption></figure><p>AI-assisted coding helps most with quick and dirty prototypes. If a prototype runs in a sandbox on your laptop and touches no sensitive data, you can move fast, try 20 ideas and keep the one or two that work. Production software is different. That very morning, a collaborator’s AI coding agent had wiped a database. Luckily, it was a test app with five users.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qXWYuVn0npHlLjb9qOAr9Q.png" /><figcaption>You control your code. Not your data, not your users.</figcaption></figure><p>Why build so fast? You control 100% of your code, but you don’t control your data (accents, car noise, background speakers) or how users react. Building is how you find out. So the mantra is not “move fast and break things.” It’s <strong>move fast and be responsible.</strong></p><h3>“Don’t Learn to Code” Is Terrible Advice</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VcwiCv9ajA8LjpgtKG33iA.png" /><figcaption>Every time coding got easier, more people coded</figcaption></figure><p>Professor Andrew thinks we’ll look back on this as some of the worst career advice ever given. Punch cards to keyboards, assembly to high-level languages, editors to IDEs, and now AI-assisted coding. Every time coding got easier, more people learned to code, not fewer.</p><h3>The Honest Truth About Jobs</h3><p>Unemployment for recent CS grads is up. But businesses can’t find enough people with GenAI and deep learning skills, because many curricula haven’t caught up.</p><p>Professor Andrew was direct: he won’t hire a software engineer today who doesn’t use AI to code. He once picked a soon-to-be grad who was fluent in AI coding over an engineer with 10 years of experience who hadn’t adapted.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*M-b_C9sAD0hYEMedJnM20w.png" /><figcaption>Who’s most productive?</figcaption></figure><p>But the most productive person of all? Someone with 10 years of experience who is also great with AI.</p><p><strong>Lesson #4: Experience still matters. Experience plus AI is unbeatable.</strong></p><h3>Fundamentals Give You the Language</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4DTNY84N-nzJNLa1ujQVDw.png" /><figcaption>Knowing the language gives you control</figcaption></figure><p>Another gem: Professor Andrew’s collaborator knew art history, so he could prompt an image generator with the language of art (genre, palette, inspiration) and get beautiful results. Professor Andrew could only ask for pretty pictures of robots. Same with software. When you understand how computers and AI work, you can tell the AI exactly what to build.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5p1n5NYWHHmWQGYF-wl8hA.png" /><figcaption>The GenAI toolbox</figcaption></figure><p>When Professor Andrew hires for GenAI roles, he looks for AI-assisted coding plus these tools: RAG, vector databases, evals and error analysis, guardrails, knowledge graphs, multimodal LLMs, fine-tuning and agentic workflows. If some of these are new to you, that’s what this series is for!</p><h3>My Biggest Takeaway</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YjjlLtghXWBxGd26XKQHTg.png" /><figcaption>Eureka!!</figcaption></figure><p>Learn the fundamentals. Use AI to build. Move fast, and be responsible. And go tell your friends in every field to become builders. The barrier to building software has never been lower.</p><p>That’s Lecture 1! Next up: Lecture 2 on supervised, self-supervised and weakly supervised learning. Follow along so you don’t miss it.</p><p><strong><em>Credits: </em></strong><em>All ideas in this post come from Stanford CS230 (Autumn 2025), Lecture 1: Introduction to Deep Learning, taught by Professor Andrew Ng (co-instructor: Kian Katanforoosh) and published by Stanford Online. Watch the full lecture on Youtube. </em><strong><em>All illustrations are my own.</em></strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4aaae6c80897" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I challenged myself to build confidence in 30 days.]]></title>
            <link>https://medium.com/@bharatigehani/i-challenged-myself-to-build-confidence-in-30-days-0d3b34178473?source=rss-5a2b76ae2846------2</link>
            <guid isPermaLink="false">https://medium.com/p/0d3b34178473</guid>
            <category><![CDATA[self-esteem]]></category>
            <category><![CDATA[confidence]]></category>
            <category><![CDATA[challenge]]></category>
            <dc:creator><![CDATA[Bharati Gehani]]></dc:creator>
            <pubDate>Wed, 18 Jun 2025 13:13:38 GMT</pubDate>
            <atom:updated>2026-09-22T03:40:10.053Z</atom:updated>
            <content:encoded><![CDATA[<h3>I challenged myself to build confidence in 7 days.</h3><blockquote>In this series, I challenge myself to learn and do everything I can to build more confidence every day. <br>I will devote time every day to reading books, watching YT videos, and going through the latest research to understand the science behind building self-esteem and confidence. <br>I will write my learnings and document my journey here day-by-day, hoping to create that 1% change in someone’s life as well. <br>Alright, so I’m excited to get started.</blockquote><p>It&#39;s 3am on a weekday, and I just had this “Eureka” moment where I want to get out of this zone I am in. It is described by Dr. Seuss as the “waiting zone.”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Uc2ZX-mDS8SpfJWgM7Uq1A.png" /></figure><p>So I challenge myself to do something new, something fun and exciting. Develop confidence! They say you have to fake it till you make it; let&#39;s see how it goes.</p><h3>Day 1: 17 June 2025</h3><p>Title: <strong>Give me 15 Minutes and I’ll Make you Dangerously Confident <br></strong>[<a href="https://www.youtube.com/watch?v=7e6RSGdvHD8&amp;t=13s&amp;ab_channel=LeilaHormozi">Link</a>]<br>Video rating: Decent ⭐⭐<br><strong>Main learning: Confidence comes from practice. Start small.</strong></p><p>She seems like a fairly confident young lady and explains that confidence comes from practice, tells you to start small, and yes, to “fake it till you make it.”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iaoOOQKaDexK464NQOB5tA.png" /><figcaption>She explains the importance of tracking your wins</figcaption></figure><p>We remember the days we lost but often forget the small “wins.” Document them and see the confidence boost.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YInMvxRXUFhyTDtX8sI0yA.png" /><figcaption>My arms are sweaty, knees weak ... (She references my favourite Eminem song here.)</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iymuC4Sckd1PBBp0Q4DtjA.png" /></figure><p>And finally, building confidence is the first step to succeeding in ANYTHING.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4G9pWZ8TvfPNL2pzwVR2Zw.png" /><figcaption>The framework to build confidence: practice, fake it till you make it, track the wins, start small, and get out of your comfort zone!</figcaption></figure><p>This was a solid start to this month :)</p><h3>Day 2: 18 June 2025</h3><p>Title: Become Confident in 60 Seconds<br><a href="https://www.youtube.com/shorts/HpeqnWQh15c">https://www.youtube.com/shorts/HpeqnWQh15c</a><br>Rating: Awesome ⭐⭐⭐⭐<br><strong>Main Learning: Trust in yourself.</strong></p><p>The more you try, the further away you are. <br>Confidence = Intense + Trust<br>Intense trust in yourself and your abilities, you are a badass; you can do it!<br>The secret to more confidence is to stop trying to be confident and just start trusting in yourself more than anyone and anything. Trust yourself.</p><h3>Day 3: 19 June 2025</h3><p>Title: The Power of <strong>Delusional</strong> Self Belief<br><a href="https://www.youtube.com/watch?v=V_Z8XoPVDBg&amp;ab_channel=ModernIdeas">https://www.youtube.com/watch?v=V_Z8XoPVDBg&amp;ab_channel=ModernIdeas</a><br>Rating: <strong>Must watch</strong> ⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐⭐<br><strong>Main Learning: No dream is too big. You need hard work but also need luck. Success is f-ing inevitable if you keep at it for a decade with delusional self-belief.</strong></p><p><strong>Chapter 1: No dream is too big. </strong><br>In 1945 scientists claimed that no one can run a mile in less than 4 minutes, and if they ever did, the heart would explode. But in 1954, this med student believed otherwise. He beat it. <br>And just 2 months later, 20 others also beat the 4-minute mark. <br>DAMN. <br>So no one beat the 4-minute mark for 9 years, and then boom, suddenly, in 2 months, 20 people beat it?</p><p>Hey, remember you only get one life; make sure to do something cool with it ;) or don’t!</p><p><strong>Chapter 2: Hard Work OR Luck?</strong></p><blockquote>As a kid you wanted to be an astronaut or an actor or something super cool.<br>After a certain age, you’re told you can’t do certain things, and it’s not wrong either… most people shouldn’t be chasing a dream.</blockquote><blockquote>But there are some people that have it inside them, and that makes me wonder, how many people never got there because they were told they couldn’t?</blockquote><p>How is Bill Gates worth more than 60 countries?</p><p>Bill Gates used to read constantly about computers … all the time. Got an idea, took that idea, and started a company.</p><p>7 days a week, no weekends for 10 years straight, never took money from his upper-middle-class parents for his education. <br>Microsoft was at a loss for the first 5 years.</p><blockquote>People say, “Wow, you’ve had such success, and it’s been so overnight and whatever.” Whatever success I have gotten has come after, like, eight years of just nothing working out.<br> ~Tarantino</blockquote><p><strong>Chapter 3: Inevitable</strong></p><p>It does take a little bit of luck and being at the right place at the right time, but in an entire lifetime, it’s impossible that you don’t get lucky once.</p><p>Tarantino is a great example of this. He did not get any success for eight long years. Logically, it would’ve made sense to quit after year 4 or 5, but he didn’t. He stuck around, and the years made him a great filmmaker, and because he was a great filmmaker and he stuck around for that long, one day he was at the right place at the right time.</p><p>So I honestly believe that if you have 1 goal and keep going after it for the better part of a decade, then it is inevitable. <br>You have to be delusional, though, to lose and still go after it for that f**king long.</p><p>BE <strong>DELUSIONAL, and with skill, a little delusion is like your guardian angel. BE DELUSIONAL; </strong>THAT’S HOW YOU MAKE IT. IN EVERYTHING.</p><h3>Day 4: 20 June 2025</h3><p>Rating: ⭐⭐<br>Topic: Life-Changing Tip For Building Confidence by Dr. Julie<br><a href="https://www.youtube.com/shorts/4_Uv1HuPhCA">https://www.youtube.com/shorts/4_Uv1HuPhCA</a><br><strong>Main Learning: Stretch your comfort zone every day without getting anxious. Challenge but don’t overwhelm.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/692/1*bwPWTt-2bHzX40pjDOP_Ug.png" /></figure><p>Blue: Comfort zone: Easy and safe <br>Red: Way out of your comfort zone: Panic zone: Overwhelming and frightening.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/702/1*M39UAr5Qy9evDtiqwhJ1Cg.png" /></figure><p>Green: Stretch Zone: Things that feel like a challenge but are doable with some effort.</p><p>Challenge yourself every day without overwhelming yourself… to grow your comfort zone.</p><h3>Day 5: 21 June 2025</h3><p><strong>Topic 1: Till I collapse</strong><br><a href="https://www.youtube.com/watch?v=gLMo7_VXU_s&amp;ab_channel=DaddyChill">https://www.youtube.com/watch?v=gLMo7_VXU_s&amp;ab_channel=DaddyChill</a><br>;) <br>Rating: Too OG to be rated by me.</p><blockquote>Till the roof comes off, till the lights go out<br>Till my legs give out, can’t shut my mouth<br>Till the smoke clears out, am I high? Perhaps<br>I’ma rip this shit till my bones collapse</blockquote><blockquote>this is your moment<br>And every single minute you spend tryna hold on to it<br>’Cause you may never get it again</blockquote><blockquote>Until the roof (until the roof)<br>The roof comes off (the roof comes off)<br>Until my legs (until my legs)<br>Give out from underneath me</blockquote><blockquote>I, I will not fall, I will stand tall<br>Feels like no one can beat me</blockquote><p><strong>Topic 2: Lose Yourself<br></strong><a href="https://www.youtube.com/watch?v=tR1ECf4sEpw&amp;ab_channel=7clouds">https://www.youtube.com/watch?v=tR1ECf4sEpw&amp;ab_channel=7clouds</a><br>Main Learning: Eminem says that you can do anything you put your mind to!</p><p>You’ve heard this song, but hear it again, and again, and again while reading the lyrics.</p><blockquote>Look<br>If you had one shot or one opportunity<br>To seize everything you ever wanted in one moment<br>Would you capture it<br>Or just let it slip? Yo</blockquote><blockquote>His palms are sweaty, knees weak, arms are heavy<br>There’s vomit on his sweater already, mom’s spaghetti<br><strong>He’s nervous, but on the surface he looks calm and ready</strong><br>To drop bombs, but he keeps on forgettin’<br>What he wrote down, the whole crowd goes so loud<br>He opens his mouth, but the words won’t come out<br>He’s chokin’ how everybody’s jokin’ now<br>The clock’s run out, time’s up, over, blaow<br>Snap back to reality, ope, there goes gravity, ope<br>There goes Rabbit, he choked, he’s so mad, but he won’t<br>Give up that easy, no, he won’t have it<br>He knows his whole back’s to these ropes, it don’t matter, he’s dope<br>He knows that, but he’s broke, he’s so stagnant, he knows<br>When he goes back to this mobile home, that’s when it’s<br>Back to the lab again, yo, this old rhapsody<br>Better go capture this moment and hope it don’t pass him and</blockquote><blockquote>(You better lose) yourself in the music<br>The moment, you own it, you better never let it go (go)<br>You only get one shot, do not miss your chance to blow<br>This opportunity comes once in a lifetime, yo</blockquote><blockquote>(You better lose) yourself in the music<br>The moment, you own it, you better never let it go (go)<br>You only get one shot, do not miss your chance to blow<br>This opportunity comes once in a lifetime, yo</blockquote><blockquote>You better</blockquote><blockquote>His soul’s escapin’ through this hole that is gapin’<br>This world is mine for the takin’, make me king<br>As we move toward a new world order<br>A normal life is borin’, but superstardom’s<br>Close to postmortem, it only grows harder<br>Homie grows hotter, he blows, it’s all over<br>These hoes is all on him, coast-to-coast shows<br>He’s known as the Globetrotter, lonely roads<br>God only knows he’s grown farther from home, he’s no father<br>He goes home and barely knows his own daughter<br>But hold your nose, ’cause here goes the cold water<br>These hoes don’t want him no more, he’s cold product<br>They moved on to the next schmoe who flows<br>He nose-dove and sold nada, and so the soap opera<br>Is told, it unfolds, I suppose it’s old, partner<br>But the beat goes on<br>Da-da-dom, da-dom, dah-dah, dah-dah</blockquote><blockquote>(You better lose) yourself in the music<br>The moment, you own it, you better never let it go (go)<br>You only get one shot, do not miss your chance to blow<br>This opportunity comes once in a lifetime, yo</blockquote><blockquote>(You better lose) yourself in the music<br>The moment, you own it, you better never let it go (go)<br>You only get one shot, do not miss your chance to blow<br>This opportunity comes once in a lifetime, yo</blockquote><blockquote>You better</blockquote><blockquote>No more games, I’ma change what you call rage<br>Tear this motherfuckin’ roof off like two dogs caged<br>I was playin’ in the beginnin’, the mood all changed<br><strong>I’ve been chewed up and spit out and booed off stage<br>But I kept rhymin’ and stepped right in the next cypher</strong><br>Best believe somebody’s payin’ the Pied Piper<br><strong>All the pain inside amplified by the<br>Fact that I can’t get by with my nine-to-five<br>And I can’t provide the right type of life for my family</strong><br>’Cause, man, these goddamn food stamps don’t buy diapers<br>And there’s no movie, there’s no Mekhi Phifer, this is my life<br>And these times are so hard and it’s gettin’ even harder<br>Tryna feed and water my seed, plus teeter-totter<br>Caught up between being a father and a prima donna<br>Baby mama drama, screamin’ on her, too much for me to wanna<br><strong>Stay in one spot, another day of monotony’s<br>Gotten me to the point I’m like a snail, I’ve got<br>To formulate a plot or end up in jail or shot</strong><br><strong>Success is my only motherfuckin’ option, failure’s not</strong><br><strong>Mom, I love you, but this trailer’s got to go<br>I cannot grow old in Salem’s Lot</strong><br><strong>So here I go, it’s my shot, feet, fail me not<br>This may be the only opportunity that I got</strong></blockquote><blockquote>(You better lose) yourself in the music<br>The moment, you own it, you better never let it go (go)<br>You only get one shot, do not miss your chance to blow<br>This opportunity comes once in a lifetime, yo</blockquote><blockquote>(You better lose) yourself in the music<br>The moment, you own it, you better never let it go (go)<br>You only get one shot, do not miss your chance to blow<br>This opportunity comes once in a lifetime, yo</blockquote><blockquote>You better</blockquote><blockquote>(<strong>You can do anything you set your mind to, man</strong>)</blockquote><p>Meaning of certain lines in the lyrics:</p><p>And there’s no movie, there’s no Mekhi Phifer; this is my life:<br>It’s his life, and no one is coming to save him like in the movie.</p><p>I cannot grow old in Salem’s Lot</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AMNa_ke-Z2vdeO2tYKi78g.png" /></figure><h3>Day 6: 22 June 2025</h3><p>Since we’ve established that confidence is tied to immense self-belief, let’s delve deeper into this. <br>Let&#39;s think about this: How did 20 people beat the 4-minute mark to run a mile in just 2 months when absolutely no one broke it in 9 years? There has to be some truth in unshatterable belief. <br>One person did it, and then they all realized their belief that it was not possible was wrong. <br>(reference to Day 3 video)</p><p>The main ingredients to success are<br>1. Delusional self-belief. <br>2. Hard work.<br>3. Consistency.</p><p><strong>Topic: How To Stop Doubting Yourself</strong><br><a href="https://www.youtube.com/shorts/MyiEljdcdKA">https://www.youtube.com/shorts/MyiEljdcdKA</a><br>Rating: ⭐⭐<br>Main learning: Don’t avoid what you fear and tackle it instead without waiting for the self-doubt to disappear.</p><ol><li>Avoidance feeds self-doubt. The more you avoid something, the more you will believe in that narrative that says, “I would’ve never been good enough anyway.”</li><li>Tackling self-doubt demands courageous action in the direction of something that matters to you, something that feels so important that it overrides your need to feel comfortable and ready.</li><li>When you stop waiting for the self-doubt to disappear, when you make this radical decision to let that self-doubt come along for the ride as you do that scary thing, it won’t haunt you back anymore.</li></ol><h3>Day 7: 9th July 2025</h3><p>I love to practice everything in the above videos and personally I feel more confident than ever in my day-to-day life. <br>I loved the video above <a href="https://www.youtube.com/watch?v=V_Z8XoPVDBg&amp;ab_channel=ModernIdeas">https://www.youtube.com/watch?v=V_Z8XoPVDBg&amp;ab_channel=ModernIdeas</a> so I will rewatch it today.</p><p>I end this blog with my favourite quotes of all time</p><blockquote>&quot;If we all did the things we are capable of doing, we would literally astound ourselves.” — Thomas Edison</blockquote><blockquote>“You cannot be lonely if you like the person you’re alone with.” — Wayne Dyer</blockquote><blockquote>“When I started counting my blessings, my whole life turned around.” — Willie Nelson</blockquote><blockquote>“If we did all the things we are capable of doing, we would literally astound ourselves.” — Thomas Edison</blockquote><blockquote>“Don’t wait until everything is just right. It will never be perfect. There will always be challenges, obstacles and less than perfect conditions. So what? Get started now. With each step you take, you will grow stronger and stronger, more and more skilled, more and more self-confident and more and more successful.” — Mark Victor Hansen</blockquote><blockquote>“If you hear a voice within you say ‘you cannot paint,’ then by all means paint, and that voice will be silenced.” — Vincent Van Gogh</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0d3b34178473" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Netflix Case]]></title>
            <link>https://medium.com/@bharatigehani/netflix-case-8cca649f80a9?source=rss-5a2b76ae2846------2</link>
            <guid isPermaLink="false">https://medium.com/p/8cca649f80a9</guid>
            <category><![CDATA[mba]]></category>
            <category><![CDATA[consulting]]></category>
            <category><![CDATA[product]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[netflix]]></category>
            <dc:creator><![CDATA[Bharati Gehani]]></dc:creator>
            <pubDate>Sat, 11 Nov 2023 21:24:22 GMT</pubDate>
            <atom:updated>2026-01-25T06:58:44.870Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zO0O2AwqGAziFBsfwcJ22A.png" /></figure><p><strong>Problem: The interviewer is a data scientist at Netflix. 10 per cent of the users of Netflix are inactive. As an aspiring product manager at Netflix, how will you solve this problem?</strong></p><p>Let’s go a little bit into the <strong>framework</strong> we’ll use to tackle this PM case. We need to make sure our thoughts are very <strong>structured</strong> and one question flows to another.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wEyIzIZF-trEC3hfnmon7Q.png" /></figure><h3><strong>Clarifying Questions:</strong></h3><h4>We need to ask the interviewer clarifying questions because we can&#39;t assume anything.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bqVqJq6Ehm2Fn68WBSGTYg.png" /></figure><h4><strong>1.</strong> <strong>Define Product</strong></h4><p><strong>The first question you ask is</strong>: What is the product about ?</p><p><strong>If you have a clear picture of the product, you ask:</strong> As per my understanding, Netflix is a personalized streaming service for movies and TV shows. Is that right?</p><p><strong>Interviewer:</strong> Yes, it is an OTT app where we watch movies and TV shows for entertainment.</p><h4><strong>2.</strong> <strong>Metrics</strong></h4><p><strong>Q: What do you mean by an inactive user?</strong></p><p><strong>Interviewer: </strong>An inactive user is a user who is not logging in.</p><p><strong><em>Q: </em>How are we measuring this 10% metric? Have our metrics recently changed?</strong></p><p><strong>Interviewer: </strong>No recent change.</p><p><strong>Q: Has the total number of users also decreased?</strong></p><p><strong>Interviewer: </strong>The base has not decreased.</p><p><strong>Q: What is the benchmark of the usually inactive users?</strong></p><p><strong>Interviewer: </strong>We don’t have many inactive users, so 10 percent is alarming.</p><p><strong>Q: Are we using a new data analytics tool? Is there a bug?</strong></p><p><strong>Interviewer: </strong>No bug</p><p><strong>Q: Are the users logging in or not logging in?</strong></p><p><strong>Interviewer: 10 percent of users are not logging in.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/935/1*ugdANtnchH3Em7SsuZGoBQ.png" /><figcaption>Wow! We found our first hint to the problem!</figcaption></figure><p><strong>Clue #1:</strong> The inactive users are not logging in to the platform at all. Are they not logging in, or are they unable to log in? Let us mark this and circle back to it later.</p><h4>3. User Profile</h4><p><strong>Q: What are the inactive user profiles? Are they from a specific age group?</strong></p><p><strong>Interviewer: </strong>There is no particular age group. The profile is not the issue.</p><h4>4. Time Duration</h4><p><strong>Q: Is this a recent problem, or has this been occurring for a while? Is this a gradual change?</strong></p><p><strong>Interviewer: </strong>It has happened in the last week. — <strong>Clue #2</strong></p><p>Yes! Another clue: Follow-up Q: Were there any development changes last weekend?</p><p><strong>Interviewer: </strong>Yes, there have been some changes. — <strong>Clue #3</strong></p><p>Let us mark this and follow up on it during the deeper internal product discussion questions.</p><h4>5. Location</h4><p><strong>Q: Is this happening in a particular demographic or worldwide?</strong></p><p><strong>Interviewer: </strong>Worldwide</p><h4><strong>6.</strong> <strong>Platform</strong></h4><p><strong>Q: Is this happening on any specific platform?</strong></p><p><strong>Interviewer: </strong>It is observed in the mobile-based application. —<strong> Clue #4</strong></p><p><strong>Here we stumble across another clue: that this is an issue only for Netflix mobile users.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/955/1*9uycRcd9UiHQ2X9onBOYHg.png" /><figcaption>A silly way to remember all the clarifying question topics is PMTL: Product Managers True Love!</figcaption></figure><p>We’re done with asking the clarifying questions; let us move on to questions from the internal product discussions, i.e., questions that are highly specific to the product.</p><h3><strong>Internal Factors</strong></h3><p><strong>1. Has there been a recent change in:</strong></p><p><strong>a. UI</strong></p><p><strong>b. Pricing</strong></p><p><strong>c. Features</strong></p><p><strong>d. A/B testing a new feature—did a new feature go live?</strong></p><p><strong>Interviewer: </strong>We have released one change in different steps of the funnel. What do you think the feature could be?</p><p><strong>Ans: </strong>It is an API change at the backend. (both IOS and Android)—<strong>Clue #5</strong></p><p><strong>2. Is this a content-related issue? Could it be because Netflix does not have content rights? Overlapping of content</strong></p><p><strong>Interviewer: </strong>Not the reason for inactivity. No content-related issues are impacting.</p><p><strong>3. Is there a payment issue?</strong></p><p><strong>Interviewer: </strong>No issue</p><p><strong>4. Is there an issue with the login problem?</strong></p><p><strong>Interviewer: </strong>From the numbers, I do see that people are not logging in. (this confirms our login issue, we will delve deeper into this during the end-to-end customer journey.)</p><p><strong>5. Sentiment analysis problem?</strong></p><p><strong>Interviewer: </strong>No issue</p><p><strong>6. Did the app crash?</strong></p><p><strong>Interviewer: </strong>No app crashes</p><h3><strong>External Factors</strong></h3><ol><li><strong>Has there been any recent PR damage? Is Netflix promoting a specific movie that might be banned by the government? Negative news about the CEO, stocks, social media image, etc.</strong></li></ol><p><strong>Interviewer: </strong>No.</p><p><strong>2. Are there any laws imposed on OTT streaming?</strong></p><p><strong>Interviewer: </strong>No.</p><p><strong>3. Are there any government regulations on Netflix?</strong></p><p><strong>Interviewer: </strong>No.</p><p><strong>4. Do people prefer competition? Amazon Prime etc.</strong></p><p><strong>Interviewer: </strong>There is no particular reason to prefer other platforms.</p><p><strong>6. Is there a change in the mobile internet packages?</strong></p><p><strong>Interviewer: </strong>No.</p><p><strong>7. Is this a seasonal change? Due to cricket, users are now streaming matches instead of content on Netflix.</strong></p><p><strong>Interviewer: </strong>It does happen with competitors as well, but typically they also come back and log in to Netflix as well.</p><p><strong>8. Has a new competitor entered the space?</strong></p><p><strong>Interviewer: </strong>No. No change in the competitor space, no change in pricing, etc.</p><p><strong>9. Is this an industry-wide change?</strong></p><p><strong>Interviewer: </strong>No.</p><p><strong>10. Is there a change in the lifestyle of people, such as during COVID?</strong></p><p><strong>Interviewer: </strong>Entertainment time has been reduced, but they haven’t stopped watching Netflix.</p><h3><strong>End-to-End Consumer Journey</strong></h3><p><strong>Draw the end-to-end consumer journey and ask the interviewer whether there is a drop in any particular activity.</strong></p><p>1. Users open the Netflix app</p><p>2. Go to log in.</p><p>3. Steps to log in: People use their phone number, username, and password or login through their mail IDs.</p><p><strong>Interviewer: </strong>Which of these three has had a feature change?</p><p><strong>Solution: Since we noted this is an API change in the backend, the problem seems to be with logging in using the phone number and OTP.</strong></p><p><strong>Interviewer: </strong>The number of logins that have been dropped is in the phone number.</p><p>4. This is how an OTP-based login works; please identify the drop in user activity and check where the problem is.</p><p>1. Phone number</p><p>2. Message on phone</p><p>3. Enter OTP</p><p><strong>Interviewer: While triggering the OTP, there was an issue in the API that was not delivering the OTP. We are getting phone numbers, but customers are not able to get any OTP.</strong></p><p><strong>Final Solution: In this case, Netflix mobile users were unable to log in using their phone numbers due to a bug in the APIs, which drastically increased the metrics of total inactive users on Netflix. Solving the bug will resolve this case.</strong></p><p>I hope you enjoyed solving this case together.</p><p><em>This case was taken up at a product management masterclass, where hundreds of students brainstormed the solutions together. This beautiful case-solving framework can be applied to all product management cases.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8cca649f80a9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How Spotify Makes Money]]></title>
            <link>https://medium.com/@bharatigehani/how-spotify-makes-money-c59857e32742?source=rss-5a2b76ae2846------2</link>
            <guid isPermaLink="false">https://medium.com/p/c59857e32742</guid>
            <dc:creator><![CDATA[Bharati Gehani]]></dc:creator>
            <pubDate>Thu, 09 Nov 2023 13:53:36 GMT</pubDate>
            <atom:updated>2023-11-09T13:53:36.028Z</atom:updated>
            <content:encoded><![CDATA[<p>Spotify is the world’s most popular audio streaming subscription service, with more than 574 million users, including 226 million subscribers in more than 180 markets.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/563/1*F-xyh9UQcYMF0yX_nNy7dQ.jpeg" /></figure><p>Spotify has become an industry leader in the streaming segment by focusing on product-led growth.</p><p>Let us deconstruct Spotify using the AARRR metrics</p><ul><li><strong>Acquisition</strong>: How are users discovering the product?</li><li><strong>Activation</strong>: Are these users performing the desired actions?</li><li><strong>Retention</strong>: Do activated users keep coming back?</li><li><strong>Referral</strong>: Do users tell other people about the product?</li><li><strong>Revenue</strong>: Do you make a profit from your users?</li></ul><ol><li><strong>How are users discovering Spotify?</strong></li></ol><ul><li><strong>Freemium Product: </strong>Spotify operates on a freemium model, where users are easily activated and can listen to curated playlists by simply signing up or logging in.</li><li><strong>Spotify targeted ads: </strong>Spotify acquires users through various channels such as social media, word of mouth, influencer endorsements, etc.</li><li><strong>Partnerships: P</strong>artnerships with hardware companies for pre-installed apps and partnerships with telecom companies · offering Spotify Premium as part of their packages. Eg- Amazon Prime offered 6 months of Spotify for free!</li><li><strong>Spotify’s USP:</strong> Personalized playlists are such a highlight it pull users in through their areas of interest.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QQo6643SsH8g46IgBT6WBA.jpeg" /></figure><p><strong>2. How can the user be easily activated?</strong></p><ul><li><strong>Easy and fast Onboarding: </strong>Easy sign-up process through social media accounts or email, reducing barriers for users to start using the platform.</li><li><strong>Customized User Experience:</strong> Recommending initial playlists or podcasts based on user preferences to engage them immediately.</li><li><strong>Guided Tours and Tutorials:</strong> Offering simple, interactive guides to highlight key features, making the platform user-friendly for newcomers.</li><li><strong>Promotional Offers:</strong> Providing limited-time access to premium features or content, enticing users to experience the full potential of the service.</li></ul><p><strong>3.</strong> <strong>How does the product drive retention?</strong></p><ul><li><strong>Personalization and Recommendations:</strong> Curating tailored playlists, podcasts, and content continuously to keep users engaged and invested in the platform</li><li><strong>Algorithmic Insights</strong>: Analyzing user data to offer more accurate recommendations and personalized content, enhancing the user experience.</li><li><strong>Exclusive Content and Releases: </strong>Introducing exclusive music or podcasts, creating FOMO (fear of missing out), and incentivizing users to stay.</li><li><strong>Notifications and Updates:</strong> Regularly updating users about new releases, personalized content, or relevant events keep the platform fresh and enticing.</li></ul><p><strong>4.</strong> <strong>How does the product drive revenue, or what would they use in the future to drive the same?</strong></p><ul><li><strong>Subscription Tiers: </strong>Offering different subscription plans (free, premium, family, and student) with varying features and pricing to cater to different user segments.</li><li><strong>Ad Revenue: </strong>Leveraging ads on the free tier for non-subscribers, generating income through advertisers</li><li><strong>Partnerships and Merchandising:</strong> Collaborations with artists, merchandise sales, and live events create additional revenue streams beyond subscriptions.</li><li><strong>Expanding Services:</strong> Introducing additional paid services like live concert streaming or exclusive access to certain events.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/564/1*2doX7zMNmxFM9CNYtc7X8g.jpeg" /></figure><p><strong>5.</strong> <strong>How does the product drive referrals?</strong></p><ul><li><strong>Social Sharing Features:</strong> Enabling users to share playlists, tracks, or podcasts on social media, amplifying Spotify’s reach.</li><li><strong>Incentivized Referral Programs:</strong> Offering perks like extended free trials or premium access for users who refer friends, encouraging word-of-mouth marketing.</li><li><strong>Collaborations and Collaborative Playlists: </strong>Allowing users to collaborate on playlists, encouraging social interactions and shared experiences, and leading to natural recommendations and referrals</li><li><strong>User-generated Content Showcases: </strong>Highlighting user-generated playlists or podcasts, showcasing the platform’s community, and driving social proof for potential new users.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c59857e32742" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Canva — Product Deconstruction]]></title>
            <link>https://medium.com/@bharatigehani/canva-product-deconstruction-cf4d97cf6de7?source=rss-5a2b76ae2846------2</link>
            <guid isPermaLink="false">https://medium.com/p/cf4d97cf6de7</guid>
            <category><![CDATA[canva-pro]]></category>
            <category><![CDATA[product-design]]></category>
            <category><![CDATA[product]]></category>
            <category><![CDATA[product-management]]></category>
            <dc:creator><![CDATA[Bharati Gehani]]></dc:creator>
            <pubDate>Sun, 05 Nov 2023 21:07:15 GMT</pubDate>
            <atom:updated>2023-11-05T21:08:50.257Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/300/1*tSM8KnWzNumMdmRg_NNQvg.png" /></figure><h3><strong>Canva — Product Deconstruction</strong></h3><p><strong>1.</strong> <strong>How can the user be easily activated?</strong></p><p>· Focus on the user.</p><p>· Easy activation—the user can start designing without even logging in to Canva.</p><p>· Highlighting that it is free to use and a perfect fit for everyone.</p><p>· Gives value—instantly—start designing without entering any details, and once you open the design page, a small notification pops up that you can get the best experience if you log in—giving more value to the user.</p><p><strong>2.</strong> <strong>How does the product drive retention?</strong></p><p>· Once the user has created the design, to save, download or share it, they must sign up.</p><p>· Canva offers real value—through several graphics, templates, HD photos and videos, magic AI, and much more. It makes it easy for anyone to design. Without being required to download a design tool or spend months learning design, the user can get started in seconds with Canva.</p><p>· Simple to use. It is extremely simple to use Canva to design a poster, video, presentation, or ad. With its ease of accessibility to the everyday user, Canva is highly engaging.</p><p>· Fun to use: It is fun to create amazing designs in seconds with Canva.</p><p><strong>3.</strong> <strong>How does the product drive revenue, or what would they use in the future to drive the same?</strong></p><p>· Canva Premium Membership: A burst of value addition to the designs is being able to use Canva’s premium photos, video graphics, and more.</p><p>· Subscription starts cheap and non-recurring, starting with 1-day and 1-week options.</p><p>· Non-recurring subscriptions give the user peace of mind that they will not auto-renew.</p><p>· A lot of small businesses now use Canva because it is a faster and cheaper option than hiring designers. Canva has designed a specific product for them with two main features:</p><p>The first is Canva for Teams, where more than one person can access the designs and work together at a minimal additional cost.</p><p>Secondly, they use their personalized brand colours and templates in all designs.</p><p>· Both of these options also drive revenue for Canva, as they are of high value to the end user, in this case, small businesses.</p><p><strong>4.</strong> <strong>How does the product drive referrals?</strong></p><p>· Value, easy to use, and fun—word-of-mouth publicity.</p><p>· Premium features are also accessible with the Canva watermark.</p><p>· It is very easy to share the designs by downloading them or through social media, with the option of getting the print delivered to your doorstep as well.</p><p>Overall, Canva gives <strong>so much</strong> value free of charge, which is the best publicity ever.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cf4d97cf6de7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Reinvent Your Life]]></title>
            <link>https://medium.com/@bharatigehani/reinvent-your-life-0bddf693e587?source=rss-5a2b76ae2846------2</link>
            <guid isPermaLink="false">https://medium.com/p/0bddf693e587</guid>
            <category><![CDATA[picture-story]]></category>
            <category><![CDATA[book-summary]]></category>
            <category><![CDATA[cal-newport]]></category>
            <category><![CDATA[slow-productivity]]></category>
            <category><![CDATA[reinventingyourself]]></category>
            <dc:creator><![CDATA[Bharati Gehani]]></dc:creator>
            <pubDate>Sun, 05 Nov 2023 18:58:33 GMT</pubDate>
            <atom:updated>2023-11-05T18:58:33.008Z</atom:updated>
            <content:encoded><![CDATA[<p>Cal Newport</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4vbYDNkzkQw7drwKfkQ9pA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EWuvfW-2wjinJc1q4t93rQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YNkfAJ1GbEZcKJZLHsO0qw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LcdfpkWrgD_Q-GATNWh69w.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tSAjB2HSGtQAk_ofh90qkg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EZhWEEO0bXn7qEhRO6qMJA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MMOYi4EaWHM0X4EHxOrpgQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TsylQLViYs4Mr42waWRxfg.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0bddf693e587" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Make More Money]]></title>
            <link>https://medium.com/@bharatigehani/make-more-money-37f8186811bf?source=rss-5a2b76ae2846------2</link>
            <guid isPermaLink="false">https://medium.com/p/37f8186811bf</guid>
            <category><![CDATA[business]]></category>
            <category><![CDATA[millionaire]]></category>
            <category><![CDATA[personal-development]]></category>
            <category><![CDATA[make-more-money]]></category>
            <category><![CDATA[book-summary]]></category>
            <dc:creator><![CDATA[Bharati Gehani]]></dc:creator>
            <pubDate>Wed, 01 Nov 2023 20:33:30 GMT</pubDate>
            <atom:updated>2023-11-01T20:33:30.053Z</atom:updated>
            <content:encoded><![CDATA[<p><em>-Picture Story by Bharati Gehani</em></p><h4><strong><em>An Eye Opener: Who Becomes Wealthy?</em></strong></h4><p>After studying tens of thousands of self-made millionaires, these people who had started with nothing and in one generation became millionaires, often billionaires, What I found was that they all had certain traits and behaviours in common.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/990/0*oZ9pt57Uu4tiQb0u.jpg" /></figure><p><strong><em>YOU CAN DO IT TOO!</em></strong></p><p>More people are becoming millionaires today than ever before.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/705/0*jCDLwc4hEk1kGvHf.jpg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*cybdf0NdyQXidjA8.png" /></figure><h4><strong><em>Habits of Millionaires and Billionaires</em></strong></h4><p>Most self-made millionaires and billionaires have special habits that separate them from average people.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/442/0*6COyo5e1KlwlFKqa.jpg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/756/0*JXmqIWNYmNzo4Oep.jpg" /></figure><p>Your financial situation will be more a result of your habits than of your education, your background, or even the wealth you inherit.</p><p><strong><em>Napolean Hill</em></strong> said in Think and Grow Rich,</p><blockquote>“Your life only begins to become great when you develop a burning desire to accomplish a <strong>M<em>ajor Definite Purpose (MDP)</em></strong>, the focal point where you concentrate all of your energies on the one goal that can have the greatest impact on your life.”</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/540/0*dBfKQw9U27Ff2n2u.jpg" /></figure><p>85 per cent of millionaires have a <strong><em>BIG HAIRY AUDACIOUS GOAL.</em></strong> The great discovery is that as you work on one goal, especially a financial one, you start to make progress on all goals at the same time.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*VQcGbh4nrLfX5vOz.jpg" /></figure><p>Average people work 40 hours or less per week, with 50% of their time wasted idle in chitchat, social media, emails, coffee breaks, etc. Millionaires, on the other hand, work 60 hours per week on high-value activities. Most billionaires have been working for 20 years in their field of expertise before they became billionaires.</p><h4><strong><em>How to Develop New Habits</em></strong></h4><p>The golden rule is to select only one habit at a time that can help you be successful. Working on multiple habits seems to short-circuit the development of a single habit.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/336/0*unPGbKwnTAvNOhAS.jpg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/621/0*wFmKB8orn3M-I64X.jpg" /></figure><h4><strong><em>How Rich People Think</em></strong></h4><blockquote>“You become what you think of most of the time.”</blockquote><p>So, the way you think on the inside determines what happens to you on the outside. So it is the way you think that defines who you are. Rich people have three special ways of thinking. <strong><em>Thinking long-term</em></strong>, instead of focusing on short-term gratification, is crucial; they sacrifice today to get a better tomorrow. They <strong><em>think slowly</em></strong> about important decisions, always do their due diligence, and make <strong><em>informed decisions</em></strong>. So the next time you’ve got to make a big decision, take as much time as you want—read, read, and read; think about the bigger picture; and follow your heart. You’ll do well.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/862/0*2ReOT6CI0nZVRK8I.jpg" /></figure><h4><strong><em>Earning More Money</em></strong></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/832/0*tgNDRosDcP3wwUE7.jpg" /></figure><p>Rich people think of ways to <strong><em>earn</em></strong> more money. The top 5 per cent of learners/earners double their income every 18 months. No one will hold your hand and make you successful; no one will coddle you. Your job and great responsibility in life is to find the perfect place for yourself.</p><blockquote>As an adult, your responsibility is to find what a great spiritual leader referred to as <strong><em>‘your true place.’</em></strong></blockquote><p>Find your true place—the place that was meant for you—the place where you fit naturally, like a key in a lock. When you find that place, you’ll make more progress in a few months or years than many make in a lifetime.</p><h4><strong><em>The 7 Basics of Business Success</em></strong></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/907/0*H0_Ftcyx908wqs1m.jpg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/907/0*3HCRJ_tHRTBqS0u9.jpg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*67BWJpxomNK0mzRi.jpg" /></figure><blockquote>“There are no results inside a business. All results are outside the business. They are out there with the customer somewhere.”</blockquote><p>It’s amazing how many people in business spend so much of their time inside their business sending out emails to try to get people to buy from them without ever going out and calling on real-life customers to ask what they want and then show them that their product or service will give them that.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/818/0*VAZJj8gunGXBqgxC.jpg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/894/0*9vaCs_gPetCu34-i.jpg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/655/0*vlgkN4y_VWDYVAYe.jpg" /></figure><p><strong><em>Zero-based thinking</em></strong> is one of the greatest thinking tools of all time. What you do is ask this question.</p><blockquote>“Is there anything you are doing today that you would not start again if you could do it over, knowing what you now know?”</blockquote><p>You can apply zero-based thinking to almost everything: your relationships, the people you’ve employed, business activities—services or products you offer (would you still offer them knowing what you know now? ), your day-to-day routine, your investments of time, money, and emotion.</p><p>If there is something you would not start again today, the next question is, “How do I get out of this and how fast?&quot; Cut it off immediately. Cut your losses as fast as you can.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/986/0*RXJPBJbfALAF3fCZ.jpg" /></figure><p>And finally, <strong><em>change</em></strong> is the most important factor determining your business success. The rate of change in every field is accelerating and is never going to slow down.</p><p>Charles Darwin was famous for saying that,</p><blockquote>“The contest for survival goes not to the strongest or most intelligent of the species but to the one most adaptive to change.”</blockquote><p>That applies to you and me as well.</p><h4><strong>The 7 Habits of High-Profit Businesses</strong></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/899/0*IKnUIfQdu2iVq5ZS.jpg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/839/0*DV6YJh7uCJ18xN9d.jpg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/870/0*slBTuJTSw0e-PjjY.jpg" /></figure><h4><strong>The 7 Habits for Personal Success</strong></h4><p>1. Write down your goals. Only 3% of adults write down their goals regularly, and they earn 10 times as much as people without written goals.</p><p>2. Increase output by 25–50% if you work from a list. Prioritize.</p><p>3. Starting and completing the most important task in the morning</p><p>4. Keep learning. Spend at least 30 to 60 minutes learning in your field. Attend seminars and workshops; you’ll find the smartest people in your field there.</p><p>5. Work all the time you work. And be on time.</p><p>6. Be healthy. Exercise for 200 minutes a week. Sleep 7–8 hours a night.</p><p>7. Develop a habit of continuous action. “Everything is hard till it becomes easy.”</p><p>Make a decision right now that you are going to develop the qualities, habits, characteristics, and personality qualities of the most successful people, and then you’re going to build your business so that it is one of the most successful businesses in the world. And I hope that you do.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/908/0*OoMPQYCKtX2DeAZC.jpg" /></figure><p>So, hey, you made it to the end.</p><p>Congratulations on reading your very first picture story!</p><p>I hope you enjoyed the process, and hey, quick question: What’s one thing common about you, me, Bill Gates, and everyone else on earth?</p><p>We’ve all got <strong><em>24 hours</em></strong> in a day. So let’s get crackin’!</p><p><em>Toodles!</em></p><p><strong><em>-Bharati Gehani</em></strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=37f8186811bf" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>