<?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 yashvi shaw on Medium]]></title>
        <description><![CDATA[Stories by yashvi shaw on Medium]]></description>
        <link>https://medium.com/@yashvishaw6?source=rss-1304e0936649------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/proxy/1*TGH72Nnw24QL3iV9IOm4VA.png</url>
            <title>Stories by yashvi shaw on Medium</title>
            <link>https://medium.com/@yashvishaw6?source=rss-1304e0936649------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 10:22:58 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@yashvishaw6/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 to Start Coding When You Have No Idea Where to Begin]]></title>
            <link>https://medium.com/@yashvishaw6/how-to-start-coding-when-you-have-no-idea-where-to-begin-ed2e2094ba3d?source=rss-1304e0936649------2</link>
            <guid isPermaLink="false">https://medium.com/p/ed2e2094ba3d</guid>
            <category><![CDATA[beginner-coding]]></category>
            <category><![CDATA[coding]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[how-to-start-coding]]></category>
            <dc:creator><![CDATA[yashvi shaw]]></dc:creator>
            <pubDate>Sun, 27 Sep 2026 10:55:56 GMT</pubDate>
            <atom:updated>2026-09-27T10:55:56.748Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/500/0*KSSkvJQHBTQKCpg2" /></figure><p>Every year, millions of individuals choose that learning to code should be their next goal. They give up very soon in the first couple of weeks, not because programming is too difficult, but due to the fact that they are starting off on the wrong foot, comparing themselves to people who are years ahead of them.<br>Here is the right way to start programming.</p><h4>1. Stop trying to pick the “right” language first.</h4><p>That’s where almost everyone stalls before they have even started writing a line of code, spending three weeks pondering forums’ discussions about Python vs JavaScript vs Rust and worrying that choosing the wrong one will make them lose their precious months.<br>They won’t. Programming fundamentals such as variables, loops, conditions, functions are all pretty much the same everywhere. Once you learn how to write them in one language, you’ll understand how to do it in any other right away. Just choose one of them and proceed:<br>- Python — the easiest syntax possible, looking very much like English and being the language of websites, AI, and data science.<br>- JavaScript — if you wish to see the result of your work immediately after writing it, that’s the language of the Web.</p><h4>2. Build something ugly before you build something good.</h4><p>A beginner programmer usually assumes that he needs to complete his programming class before he is able to do anything. This is completely wrong. It is precisely the programmers who continue with coding that work on the little embarrassing projects straight away — such as a todo list, a guessing game or a web page with a picture of their dog.<br>There is no knowledge gained by watching tutorials. All the knowledge is gained in the fifteen minutes that you spend scratching your head, looking up an error online and solving it yourself.</p><h4>3. Use AI as a tutor, not a crutch.</h4><p>This becomes even more crucial for your generation than for any other. Claude or ChatGPT could easily write a piece of code for you, hence the necessity for your golden rule regarding those tools: do not copy-paste code you do not understand.<br>You should work with AI just as you would with an attentive tutor:<br>Ask it to clarify each line of the code. Rather than asking it “how to fix this?”, ask “why did it fail?” and carefully read the explanation before making any changes. Ask it to quiz you or give you a slightly more complicated problem that you just solved.<br>If you skip this stage, you will be left with a project that cannot be explained, an interview that cannot be passed and a foundation that falls down the second the thing fails in a manner unfamiliar to the AI.</p><h4>4. Don’t learn in isolation, build things that connect to what you already like.</h4><p>If you enjoy playing games, make one. If you enjoy listening to music, make something that can help you organize playlists. And if you enjoy sports, make something that can track statistics for your favorite team. The motivation will not last long while you are going through mundane assignments, but it will last longer while making something you would really use.</p><h4>5. Once you’re not a total beginner, start reading other people’s code.</h4><p>After a few months, head over to GitHub and start looking at small and actual projects. It’s okay that you won’t understand everything because you don’t need to understand everything; you just have to see the patterns in how the experienced programmers do things.</p><p>It’s not really the language, the class, or the laptop. It’s the realization that it is okay to be terrible at this for a little bit.<br>Everyone who currently knows how to write code was once looking at a red error message, clueless about what it means. The only thing that separates people who learn from those who do not is that they did not stop after seeing the error message.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ed2e2094ba3d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The AI Bubble Isn’t About Whether It Works. It’s About Whether Anyone Can Afford It.]]></title>
            <link>https://medium.com/@yashvishaw6/the-ai-bubble-isnt-about-whether-it-works-it-s-about-whether-anyone-can-afford-it-1acf87d2b257?source=rss-1304e0936649------2</link>
            <guid isPermaLink="false">https://medium.com/p/1acf87d2b257</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[cost-of-ai]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[yashvi shaw]]></dc:creator>
            <pubDate>Sun, 27 Sep 2026 08:15:06 GMT</pubDate>
            <atom:updated>2026-09-27T08:15:06.507Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VaOl-RzVXENUZA93YYUbsA.png" /></figure><p>All current conversations about AI revolve around the wrong question. Conversations about the quality of the models being “good enough,” about them hallucinating too much, or whether they will take away your job are all there.<br>This is the easy conversation, the one everyone feels entitled to give their opinions on and which gets all the attention.<br>There is only one conversation no one is willing to have: who cares if the technology works if the underlying business model does not?</p><h4>Every AI product you love is being sold to you below cost.</h4><p>Your monthly subscription for unlimited coding, unlimited pictures, unlimited chat completion? Some company is absorbing the difference between the price you pay and the actual cost of doing this service. At the moment, this company is a VC firm who needs you locked in before the price is raised. It’s not some crazy conspiracy theory — this is the old school approach to software: subsidize the product, raise prices when the switching cost becomes sufficiently high for leaving to hurt more than paying.<br>This story was told before, in the history of Uber who lost money on each ride for over a decade just to eliminate the competition in the taxi industry, and after switching the price when both customers and drivers had no other options. WeWork did the same giving away cheaper office spaces while the music was still playing. And now AI does the exact same thing under the mask of “democratizing intelligence.”</p><h4>The heaviest users are the ones quietly bankrupting the product.</h4><p>The following is what the AI hype cycle does not want you to pay attention to: It’s not the majority of the users using the product casually that decides the product’s fate; it’s the 5% of users who run agents all day, completing thousands of prompts, who do. The rest of the users, who ask it to write an email for them twice a week, are nothing but pure profit. But the instant the product becomes good enough for the power users to start coming in droves, the economics go from subsidized growth to unsustainability overnight.<br>And this is the reason why you already see asterisked unlimited plans. Rate limits. “Fair use” policies. Peak-hour throttling. Nobody talks about it being a price hike, for it is not one.</p><h4>“The model is smarter” doesn’t pay the electricity bill.</h4><p>All of the industry is eager to brag about capability — more contextual windows, smarter reasoning and processing, multimodal in all its glory. However, the issue that isn’t discussed as widely is that any advancement usually comes with increased cost, not reduced cost. A smarter model would be, most likely, more costly. Efficiency improvements do happen, however, they cannot catch up with the increasing rate of demand.</p><blockquote>So the real question here was never about whether AI was impressive or not. It is obvious that it is impressive. However, the real question is: who is going to pay for all the real costs of running all of the promised stuff, and what will happen to the product when this stops happening?</blockquote><p>You could be as sure that this technology is truly revolutionary — and I am — and yet be equally sure that the present valuation is fiction. And that is not an impossible combination. In point of fact, that may be the only intellectually consistent combination possible at the moment. The fact that the tech is real does not make the business real.</p><h4>What happens when the subsidy ends?</h4><p>There are a few things that will happen, likely in parallel with each other: increased costs for heavy users, free tiers being cut down to the point where they’re really more like demos than anything else, and an influx of dead AI-wrapper startups, who never really did have much of a moat except “we used the API and put together a nice UI.” The companies that do survive won’t be those with the most impressive demo. They’ll be the ones that had their unit economics worked out already.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1acf87d2b257" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Here’s How I Can Tell You’ve Used AI]]></title>
            <link>https://medium.com/@yashvishaw6/heres-how-i-can-tell-you-ve-used-ai-91e711dc76f0?source=rss-1304e0936649------2</link>
            <guid isPermaLink="false">https://medium.com/p/91e711dc76f0</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[anthropic-claude]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[yashvi shaw]]></dc:creator>
            <pubDate>Thu, 24 Sep 2026 16:05:43 GMT</pubDate>
            <atom:updated>2026-09-24T16:05:43.352Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/662/1*7tEWgn7oa21QJFDvqnNdqQ.png" /></figure><p>Since I read a lot of text, I have started developing an innate sense about it. Someone emails or writes a cover letter, and something feels off even before I figure out why. It looks polished but hollow. Here is what usually tips me off.<br>The vocabulary sounds unusually lofty. People rarely use words such as “delve,” “tapestry,” “landscape,” or “testament” in casual writing. AI has a tendency to use these terms repeatedly, and therefore, a casual message from a colleague that sounds like a brochure gets noticed.<br>Every single thing comes in sets of three. “Fast, reliable, and scalable.” “Learn, grow, and thrive.” AI loves threes, and the tendency is almost universal in every paragraph. One instance sounds nice; ten instances become a pattern.</p><p>The “it’s not X, it’s Y” argument. “It’s not just a tool, it’s a revolution.” Humans use this rhetorical device on occasion. AI uses it to fill an empty void.<br>Dashes galore. Some folks love the dash, but when every sentence has one, and the author has never used it before, something is fishy.<br>Perfect structure, no personality. AI likes well-organized paragraphs, with nice, even bullet points and an equally even conclusion that starts off with “In summary” or “Ultimately.” Writing is never so perfect; it has a mind of its own, filled with opinions and tangents and an occasional clunky sentence.<br>Lack of specificity. This is probably the biggest red flag. “This was a difficult assignment” is how an AI writes. Not “this was the day I worked all night after midnight before the product launch when the server crashed.”</p><h4>Just a few caveats</h4><p>None of the above can be taken as proof; there are many humans who write in a formal style, enjoy using dashes, and deliberately use the word “delve.” AI detection systems are notoriously unreliable, too, often unfairly marking non-native English speakers. <br>This is all a gut instinct based on certain patterns — no judgment.<br>There is nothing wrong with using AI in general. Quite a lot of people do it quite correctly, using AI to generate ideas, polish their grammar, or overcome writer’s block.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=91e711dc76f0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[We Optimized Software for Speed. AI Is Making Us Pay for It.]]></title>
            <link>https://medium.com/@yashvishaw6/we-optimized-software-for-speed-ai-is-making-us-pay-for-it-b3ecb3c7866c?source=rss-1304e0936649------2</link>
            <guid isPermaLink="false">https://medium.com/p/b3ecb3c7866c</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[yashvi shaw]]></dc:creator>
            <pubDate>Thu, 17 Sep 2026 12:25:41 GMT</pubDate>
            <atom:updated>2026-09-17T12:25:41.326Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*O7NtzOTbuos_2KhRkA5ygA.jpeg" /></figure><p>For two decades, the unspoken law of software development was “ship fast and fix it later.” Speed before safety. Technical debt was an understood concept, a cost that would be cleaned up at some point in the future that would never really come. It sort of worked because there were people involved in the process, reviewing all those pull requests. That assumption is quietly breaking.</p><h3>The debt is compounding faster than we can see</h3><p>AI programming assistants excel at churning out code that seems plausible, but they fall way short when it comes to knowing if that code fits within the true context of your system, follows the established conventions from three files down, or reintroduces a bug that was already solved eight months ago.</p><p>So what happens in practice: AI-generated code gets merged faster than it gets understood. Each individual pull request looks fine. The reviewer, under normal time pressure, skims rather than reads. Multiply that by a team shipping features at triple the old velocity, and you get systems that grow in complexity faster than anyone’s mental model of them does.</p><p>This isn’t a hypothetical. Engineering teams are increasingly describing codebases that “work” but that nobody fully trusts anymore, where the honest answer to “why does this function exist” is “an AI wrote it and the tests passed.”</p><h3>Why this is different from old-school tech debt</h3><p>Technical debt from the old era used to be more legible. A badly-written function, a quick fix, a “to do: fix this later” comment, ugly but readable code created by someone who knew what they were doing.</p><p>AI-generated debt is often invisible on inspection. It reads clean. It’s well-formatted, sensibly named, and passes the linter. The debt isn’t in how it looks — it’s in the gap between how confidently it was written and how well anyone actually understands it. That’s a much harder kind of debt to audit, because the usual signal — messy code — no longer correlates with risk.</p><h3>What actually helps</h3><p>A few things are emerging as real countermeasures, not just hygiene theater:</p><ul><li>Think of AI output as a first draft created by a super-fast and extremely green contractor. Some ownership needs to be taken of that.</li><li>Put effort into architecture and user interface design, not code reviews alone. Good boundaries mean that misunderstandings will only go so far in a system.</li><li>Reward teams who understand deeply over those that deliver quickly. A team that works slower but knows all parts of the system will win out over the one that doesn’t.</li><li>Use AI to teach, not to create. Have it give you an explanation of what the code does and why as a way to double-check your understanding.</li></ul><p>The teams that will struggle most aren’t the ones using AI the least. They’re the ones who used it fastest without changing how they verify what it produced. Speed was never really the scarce resource in software. Understanding was. AI just made that far easier to lose track of.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b3ecb3c7866c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Developers Are Quietly Burning Out on AI Coding Tools]]></title>
            <link>https://medium.com/@yashvishaw6/why-developers-are-quietly-burning-out-on-ai-coding-tools-42b15b6ef9ae?source=rss-1304e0936649------2</link>
            <guid isPermaLink="false">https://medium.com/p/42b15b6ef9ae</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[claude]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[hate-ai]]></category>
            <dc:creator><![CDATA[yashvi shaw]]></dc:creator>
            <pubDate>Wed, 16 Sep 2026 14:14:10 GMT</pubDate>
            <atom:updated>2026-09-16T14:14:10.758Z</atom:updated>
            <content:encoded><![CDATA[<p><em>The tools were supposed to make our lives easier. So why does writing code feel more exhausting than ever?</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4qFyPYH0NxkAMo6Ho4c9SA.png" /></figure><p>A couple of years ago, programming with AI coding assistants was like magic. You would write a comment, and a functional code would appear. All the unnecessary repetitive work disappeared. The debugging process became much easier. It seemed that software engineering had entered its simplest period.</p><p>Now, walk into almost any engineering Slack channel or scroll through developer forums, and you’ll find a different mood entirely — one of quiet resentment, creeping anxiety, and a strange kind of fatigue that has nothing to do with how many hours someone worked.</p><h3>The Skill You Spent Years Building Now Feels Optional</h3><p>The majority of engineers do not join this profession because they want to control a tool. They joined the profession because they enjoyed problem-solving — solving a tricky bug, designing an elegant system architecture, figuring out how it all works.<br>AI assistants are very good at creating code. They are not as good at creating understanding. If a junior engineer takes suggestions without understanding why they work, he/she will deliver working code without forming the mental models that used to be created by wrestling with the problem. The senior developers face the same challenge — taking suggestions without being able to fully understand the reasoning behind them.</p><p>The result is a slow erosion of the very skill developers spent years cultivating. It’s hard to feel proud of work you didn’t fully author, and it’s even harder to feel confident in code you don’t fully understand.</p><h3>Reviewing AI Output Is Its Own Job — And Nobody Budgeted for It</h3><p>There’s also a tax on verification effort that doesn’t appear in productivity metrics: that of verification. Any generated blocks of code must be read, verified, tested, and very often rewritten. It’s not an option — a plausible piece of code that is still incorrect seems to be a greater threat than an obviously faulty one.<br>Most engineers report putting the same amount of effort into fixing bugs in the AI-assisted code as they would need to put into coding it from scratch. The key difference, of course, is that coding feels like an art, while debugging feels like cleaning up after somebody else’s mess.</p><h3>The Pressure to “Just Use AI” Is Coming From Everywhere</h3><p>It’s rarely a mandate. It’s worse — it’s ambient pressure. Managers ask why a task took three days when “it should only take one with Copilot.” Recruiters ask about AI tool fluency in interviews regardless of the role. Performance reviews start quietly factoring in “AI leverage” as a productivity metric, even when no one can clearly define what that means.</p><p>This creates a strange bind: engineers are expected to move at AI-assisted speed while still being fully accountable for AI-assisted mistakes. Speed goes up. Slack goes down. The margin for thinking carefully — which is, not incidentally, the thing that actually prevents production incidents — keeps shrinking.</p><h3>Code Reviews Have Become Adversarial in a New Way</h3><p>Reviewing a colleague’s code has always involved trust — trust that they understood the problem, considered edge cases, and made deliberate tradeoffs. AI-assisted code disrupts that trust in a subtle way. Reviewers increasingly have to ask not just “is this good code?” but “did a human actually think about this, or did they accept the first suggestion that compiled?”</p><p>That second question is corrosive to team dynamics. It’s hard to build professional trust with colleagues — or with yourself — when the honest answer is sometimes “I’m not totally sure.”</p><h3>What’s Actually Driving the Resentment</h3><p>It isn’t really about the tools being bad. Most of them are impressive, genuinely. The resentment comes from something more structural:</p><ul><li>Autonomy has shrunk. Fewer engineers get to sit with a hard problem and think. There’s pressure to reach for a tool immediately.</li><li>Ownership has blurred. When something breaks, is it your bug or the model’s? That ambiguity is uncomfortable in a field built on accountability.</li><li>Identity has been quietly reshaped. Being “good at coding” used to mean something specific. Now it partly means being good at prompting, reviewing, and editing — a related but different skill that many engineers never signed up for.</li></ul><h3>Where This Leaves Us</h3><p>None of this means AI tools should be abandoned — most engineers, even frustrated ones, aren’t asking for that. What they’re asking for is more agency in <em>how</em> these tools get used: room to write things by hand when it matters for learning, honest conversations about what “productivity” should actually measure, and recognition that verifying machine output is real work, not a shortcut.</p><p>Those engineers who are succeeding at the moment are neither those who reject the AI technology wholesale nor those who submit themselves entirely to AI. What they have done is draw a very clear line for themselves in which they have used AI as a tool that they have managed and controlled, rather than been buffeted by.<br>That may turn out to be the difference between burnout and sticking with the field they once loved.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=42b15b6ef9ae" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Isn’t Replacing Programmers. It’s Replacing the Reason They Learned to Think.]]></title>
            <link>https://medium.com/@yashvishaw6/ai-isnt-replacing-programmers-it-s-replacing-the-reason-they-learned-to-think-e8877292d9b8?source=rss-1304e0936649------2</link>
            <guid isPermaLink="false">https://medium.com/p/e8877292d9b8</guid>
            <category><![CDATA[coding]]></category>
            <category><![CDATA[claude-code]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-replace-programmers]]></category>
            <dc:creator><![CDATA[yashvi shaw]]></dc:creator>
            <pubDate>Sun, 13 Sep 2026 18:33:48 GMT</pubDate>
            <atom:updated>2026-09-13T18:33:48.357Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*RZjsEc3BQYT6G-p_pkdz1Q.jpeg" /></figure><p>Everyone’s arguing about the wrong threat. The headlines say AI is coming for developer jobs, so developers argue back with headcount numbers and hiring reports, and both sides miss what’s actually happening under their feet.</p><p>No one was laid off by a chatbot last week. But something else happened, something that was far more disturbing.</p><h4>The Job Was Never “Write Code”</h4><p>This is the part of the job-loss conversation that gets overlooked: coding is the obvious 20%. The rest, which no one sees, is the 80%, which consists of staring at a problem until it stops being ambiguous, having five hypotheses floating around in your head at one time, seeing the bug that’s not where anyone expected it to be.<br>It was the 80% that made someone a good engineer. Coding is just how you exported the process into a machine.<br>And when technology comes along that enables the 20% without any input from a developer, the inevitable question “is it going to replace developers?” becomes irrelevant. The actual question is: What do you do with the 80% when it’s no longer required?</p><h4>The Skill Degenerates before the Job Itself</h4><p>Picture two engineers, five years apart in age, both starting today. One spends her first two years fighting with a stack trace at 1 am, forming a mental model of the system by getting things wrong repeatedly. The other never has to, every stuck moment gets resolved by asking an assistant and the fix appears before the frustration has time to teach her anything.</p><p>Both ship the same code this week. In three years, only one of them can tell you <em>why</em> the system behaves the way it does when it breaks in a way no assistant has seen before. The other has a portfolio full of shipped features and a mental model with holes in it exactly where the hard problems used to be.</p><p>This isn’t a hypothetical. It’s already showing up in interviews and on-call rotations: engineers who can produce code fluently and freeze the moment the code produces a result nobody can explain.</p><h4>The Real Replacement</h4><p>AI is not going to replace programmers either. What is going to happen is that the struggle that would be an inevitable experience in becoming one is going to go away.<br>This is more challenging than “will there be any jobs?” This question does not appear in any lay-off announcement. This will surface after five years as a senior designation with a junior’s mindset, surrounded by a group who have forgotten the art of debugging anything beyond their assistants’ reach.</p><h4>What Doesn’t Get Automated</h4><p>There is a path to getting out of this, and it doesn’t involve rejecting the tools; that’s too late, and it shouldn’t be an option anymore. It’s making conscious decisions about the pain that you choose to retain.</p><p>The engineers who will be dangerous in five years are those who use AI to avoid typing but not thinking — those who still force themselves to spend twenty minutes staring at a difficult bug before asking for help, who still look at the code that their assistant generated, rather than just executing it, who regard “it works” as the beginning of a problem, rather than its end.</p><p>This is not longing for the difficult path; it’s an acknowledgment that the difficult path was never the price to pay for the job. It was the job.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e8877292d9b8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Your Chatbot Is Not Your Friend]]></title>
            <link>https://medium.com/@yashvishaw6/your-chatbot-is-not-your-friend-8faa6b36d5a9?source=rss-1304e0936649------2</link>
            <guid isPermaLink="false">https://medium.com/p/8faa6b36d5a9</guid>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[chatbots]]></category>
            <category><![CDATA[claude-ai]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[software-development]]></category>
            <dc:creator><![CDATA[yashvi shaw]]></dc:creator>
            <pubDate>Sun, 13 Sep 2026 16:27:22 GMT</pubDate>
            <atom:updated>2026-09-13T16:27:22.920Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/540/1*nEQhWwgqSMtShLwTXqrT4w.png" /></figure><p>Ladies and gentlemen, it’s time for a discussion on engagement. That’s a fancy term for clicking, and clicks make money for businesses. However, I’m here to tell you that your customers do not want to be engaged. They want to be left alone.<br>Before you develop yet another “friendly AI assistant” named after something like Ava, Max or Sunny. Before you give it a cute wave emoji and develop a personality profile larger than your onboarding process. Before you teach it to say “Great question!” to a person who is furious about a billing error, just stop. Sit down. Let’s talk about what actually happened.</p><h4>The Bot That Cried “I’m Here to Help!”</h4><p>Somewhere in a meeting, someone said: “What if the chatbot felt more human?” And everyone nodded, because it sounded right. Warmth is good. People like warmth.</p><p>But here’s the thing nobody said out loud: a person who is typing into a support chat at 11:47 PM is not looking for a friend. They are looking for their package. Their refund. Their password back. They are, in the truest sense, trying to leave the conversation as fast as humanly or artificially possible.</p><p>You gave them a chatbot with a face and a name and three exclamation points per message. What they wanted was an answer in one message, or a fast handoff to someone who could give them one.</p><h4>Cost Center Dressed as Customer Love</h4><p>Let’s be honest about the order of operations, because the order matters. Nobody built a chatbot because customers were begging for one. They built it because a support team costs money, and a bot costs less money, and somewhere a slide deck had a line that said “40% deflection rate” and everyone clapped.</p><p>That’s a fine reason to build something. It is a <em>terrible</em> reason to pretend you built it out of love. Customers can smell the difference between “we made this to serve you better” and “we made this so three fewer humans have to answer the phone.” One of those framings survives a bad interaction. The other one doesn’t.</p><h4>What People Actually Forgive</h4><p>Here’s the part that should make you feel better, not worse: people are not anti-bot. They are anti-<em>friction</em>. They will talk to an obvious, un-costumed piece of automation all day long if it’s fast, honest about what it is, and gets out of the way the second it’s out of its depth.</p><p>What they will not forgive:</p><ul><li>A bot that pretends to be a person, badly</li><li>A bot that loops the same three answers no matter what you type</li><li>A bot standing between you and a human with no visible exit</li><li>A bot that says “I understand your frustration” and then does nothing about it</li></ul><p>Notice none of those are “the bot used AI.” The complaint was never the technology. It was the theater around the technology.</p><h4>The Solution No One Wishes To Consider</h4><p>Make the bot boring. Make it fast. Make the exit door to a human obvious and one click away, always, no exceptions. Stop asking it to have a personality and start asking it to have a job: answer the twenty questions people actually ask, and know instantly when the question isn’t one of the twenty.<br>That’s it. That’s the whole secret. Not sentiment. Not a cute avatar. A machine that knows its lane and gets out of your way.<br>Nobody wants to talk to your chatbot. They want their problem to disappear. Build for that, and the chatbot becomes invisible — which, it turns out, is the highest compliment automation can ever earn.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8faa6b36d5a9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Do You Still Need College When AI Can Teach You Everything for Almost Nothing?]]></title>
            <link>https://medium.com/@yashvishaw6/do-you-still-need-college-when-ai-can-teach-you-everything-for-almost-nothing-b262cd6e8404?source=rss-1304e0936649------2</link>
            <guid isPermaLink="false">https://medium.com/p/b262cd6e8404</guid>
            <category><![CDATA[university]]></category>
            <category><![CDATA[university-worth-it]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[college]]></category>
            <dc:creator><![CDATA[yashvi shaw]]></dc:creator>
            <pubDate>Fri, 04 Sep 2026 08:28:15 GMT</pubDate>
            <atom:updated>2026-09-04T08:28:15.939Z</atom:updated>
            <content:encoded><![CDATA[<p>The answer is not “yes” or “no.” The answer is that you’re asking the wrong question.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/980/1*toeHq4j39uIUv9R9TQ-3Rw.jpeg" /></figure><p>This is the sales pitch that you have probably heard yourself before, perhaps even said yourself: how does spending $40,000 annually and four years of your life in lecture halls beat out getting the same concept explained to you individually by an AI model for the cost of a coffee subscription? Throw in a $50 mentorship session online and some YouTube playlists, and you get the whole syllabus of a computer science degree without the burden of loans.</p><p>This is a tempting offer indeed. Unfortunately, it overlooks almost all of the reasons why college used to matter and hugely overestimates what “learning a concept” is all about.</p><h4>What AI Is Genuinely Great At</h4><p>Give AI its credit first, because it really deserves it. If your aim is solely “to understand how a hash table works” or “to know enough Python to write a script,” then AI tutors are great for it. They have patience which no professor with 300 students ever can have. They will explain everything in five different ways to make you understand it. It helps you right when you are stuck, not after three days during office hours.</p><p>With regard to pure information transfer — the “here’s an idea, and now you know it” element of learning — AI has pretty much nailed the access issue. The aspect of the college sales pitch that involved you paying for the delivery of the content has really died. You were overpaying for that before AI was even around.</p><h4>What It’s Not Solving, and What You’re Actually Paying For</h4><p>But “understanding a concept” was never the hard part of a degree. Here’s what’s harder to replace:</p><ol><li>Credential is a way of establishing trust in shorthand, and trust cannot be established through a chatbot. The diploma shows the potential employer, right on the resume, “This individual completed an arduous multi-year project, under time pressure, assessed by individuals who have no motive to falsify information on their behalf.” The process of verifying that fact hasn’t been delegated to “Self-study with the assistance of AI” yet, not because the process of self-study does not allow one to acquire true knowledge, but because of the lack of verification component. Everyone can just claim they have learned it.</li><li>The friction within a cohort is a benefit, not a hindrance. The student who makes the same point about a concept in a way that is different from what you anticipated; the group project that pairs you with a person who goes about things the “wrong” way; the professor who challenges your logic rather than confirming it; all of this friction is where much of your actual thinking happens. The AI tutor is designed specifically to be agreeable and adaptable to you. It seldom tells you that you are wrong in a way that hurts.</li><li>Nobody has yet figured out how to solve the accountability challenge. Self-directed learning has an extremely harsh completion ratio. It is not because of any inherent inadequacy in the material, but it is because of the absence of any sort of framework pushing you through the dull period in the middle of any difficult learning process. Universities cost a great deal of money also because you are getting paid for a process designed to motivate you even when you do not want to — due dates, cohort, and advisers after your tail.</li><li>And the network is something that is not something you can learn. It’s something that you have to experience. And a massive amount of the value of having a degree is not the actual classes; it is three/four years of being around the people who will be in your industry for the next thirty years. The founders who meet their co-founder at college. The analyst who gets hired on a phone call from a professor.</li></ol><h4>So Is It Worth It?</h4><p>This is where “do I need college?” becomes “what am I really trying to do?”<br>If what you’re looking for is a skill in some specific profession where demonstrated skill matters more than credentials (trades, much tech work, creativity) then the math has changed. AI plus focused self-study plus a portfolio of real work can out-compete a degree, especially an expensive one.<br>If what you’re looking for is entry into a profession gated by credentials (medicine, law, most finance, most academia, most corporate positions), then the degree is non-negotiable, regardless of how well you’ve learned the subject matter. No one hires a self-taught surgeon.<br>When you’re comparing a specific four-year degree program to AI mentors, the issue isn’t “am I able to learn the same concepts for less?” Of course you can — concepts never cost the big bucks. The issue is whether the particular path you have in mind requires the credential and network and discipline — or simply the knowledge. These are separate costs, and confusing them is how you end up wasting money on a degree that wasn’t necessary, or skipping one that was.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b262cd6e8404" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[You Didn’t Get a Lazy Answer From AI. You Asked It a Lazy Question, Then Blamed the Robot.]]></title>
            <link>https://medium.com/@yashvishaw6/you-didnt-get-a-lazy-answer-from-ai-you-asked-it-a-lazy-question-then-blamed-the-robot-c7c538dc2c73?source=rss-1304e0936649------2</link>
            <guid isPermaLink="false">https://medium.com/p/c7c538dc2c73</guid>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[how-to-use-ai]]></category>
            <category><![CDATA[gen-z]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[claude]]></category>
            <dc:creator><![CDATA[yashvi shaw]]></dc:creator>
            <pubDate>Tue, 01 Sep 2026 13:11:58 GMT</pubDate>
            <atom:updated>2026-09-01T13:11:58.969Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iDGAj5TRGjAPqyWMUVuopQ.png" /></figure><p>Everyone uses AI as a politely mannered search engine. They ask a question, get an answer, and either like it or complain that “AI is not that smart.”<br>The hard reality is that it’s not the machine that is the issue. It’s the instructions that are.</p><h4>The Problem With the One-Line Prompt</h4><p>Just fire up an AI software today, and you will be greeted by a screen filled with prompts such as:<br>- “Write me a marketing email.”<br>- “Summarize this article.”<br>- “Generate blog post ideas.”<br>None of these prompts are bad. However, they’re incomplete prompts. You’re giving a task to the AI without defining the context, any constraints, or what constitutes a good output from its end.</p><p>Imagine yourself giving a new employee the assignment of “writing me a marketing email” without providing any more information. The result will be generic, bland, and perhaps inappropriate for your company. Exactly the same is done by AI because it tries to fill the gap that was not provided to it.</p><h4>Reasoning Doesn’t Happen by Accident</h4><p>The models people call “smart” aren’t smarter because of the model, they’re smarter because of the setup. There’s a specific shift that changes everything: stop asking AI for an <em>answer</em>, and start asking it to <em>reason toward one</em>.</p><p>It takes just three small tweaks to make that happen:<br>1. Assign it a role, not just a prompt. “You’re a senior copywriter specializing in writing direct response emails for SaaS startups” will yield a very different outcome than if no such role was assigned at all. Not only does the model have broad knowledge but it also has specific patterns it can use if told which “expert” to emulate.<br>2. Make it explain its reasoning before producing the output. Something like “First, identify three major objections a person would have to this product, and then compose an email in response” makes the model think first and produce afterwards. If you omit this step, you’ll get superficial fluency. If you include it, you’ll get structure.<br>3. Define what constitutes a failure. Most people give instructions only on what needs to be done right. Few define what needs to be avoided. Something like “don’t use any buzzwords like ‘synergy’ or ‘leverage,’ and no exclamation points,” is worth three paragraphs of instruction.</p><p>A Before and After:</p><p>Before: “Write a LinkedIn post about remote work.”</p><p>After: “You are a startup founder and have led all-remote teams for 5 years. Post an argument on LinkedIn that while remote working is not the productivity drain some say it is, this is true only with good documentation. Start by making a somewhat contentious statement. Use short sentences. Don’t use any business language. Conclude with a statement that invites disagreement, not agreement.”</p><p>The first prompt gives you an impersonal paragraph about remote work that could be found in any other writing on the subject. The second gets you something with a point of view, since you gave the model a point of view to work from.</p><h4>The Real Skill Nobody’s Teaching</h4><p>Those who actually get useful results out of the AI have not discovered any particular set of magic words. Instead, they’ve learned the ancient skill of being crystal clear about their intentions before they ask. <br>It is not a way to trick AI. It’s simple common sense and artificial intelligence has a knack of letting you know how much of that you actually had on offer.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c7c538dc2c73" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to Use Claude to Earn a Living: A Practical, Evidence-Based Guide]]></title>
            <link>https://medium.com/@yashvishaw6/how-to-use-claude-to-earn-a-living-a-practical-evidence-based-guide-796c9c4d51c4?source=rss-1304e0936649------2</link>
            <guid isPermaLink="false">https://medium.com/p/796c9c4d51c4</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[anthropic-claude]]></category>
            <category><![CDATA[earn-money-from-ai]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[how-to-use-ai-at-work]]></category>
            <dc:creator><![CDATA[yashvi shaw]]></dc:creator>
            <pubDate>Tue, 25 Aug 2026 13:25:20 GMT</pubDate>
            <atom:updated>2026-08-25T13:25:20.110Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*yEZVbTCZ798Wni4NWF3XIg.png" /></figure><p>Promises of making money with the help of AI tools are ubiquitous, and most of them are unverifiable claims. This guide takes a different approach: every claim about what’s actually happening in the market is drawn from Anthropic’s own published research, official reports from labor-market platforms like Upwork, and mainstream financial reporting. As for the tool’s description, it covers only the currently known capabilities of Claude. None of the information below represents any guarantees of making money; it just describes what the tool does and how the market rewards users of the tool.</p><h4>Why this is a genuine opportunity right now and not hype?</h4><p>There are two independent data sources that paint the same picture.</p><p>Anthropic’s own usage data: Anthropic conducts an ongoing research initiative known as the Economic Index, where millions of actual conversations involving Claude are analyzed in order to observe actual use of the product across the economy. Its January 2026 report revealed that Claude is involved in at least a quarter of all activities performed by about half of all occupations studied, and its use is no longer restricted to a small group of coding-related activities, as the share of traffic directed towards the ten most frequently performed activities decreased from 24% in November 2025 to 19% in February 2026, it means that the scope of paid work done using Claude is diversifying, not narrowing down.</p><p>Independent labor-market data: Upwork’s Future Workforce Index 2026, which surveyed 2,400 skilled workers in the United States along with data from the company’s marketplace, published July 2026, found that freelancers who incorporate AI into their delivery receive an average of 34% more pay per hour than those that do not. This does not imply that incorporating AI into the delivery increases income linearly; on the contrary, complex and judgment-driven tasks that are augmented with AI have led to a 45% Year to year increase in freelancer income, while high volume generative tasks that don’t require much differentiation saw a decrease in contract value of 13%, despite the number of these contracts increasing 90%. “It ain’t about what you think about using AI, it’s about using it to produce better results”, in other words. The same report stated that almost a third of skilled knowledge workers in the U.S. freelance at present, up from 25% last year, and over half of full-time employees are thinking of doing so.</p><p>In addition to the above, it seems that the business of Anthropic itself grew so much in a relatively short period that it provides an indication of how significant demand there is for Claude-powered products/services. According to Bloomberg and Axios reporting, its annualized revenue run rate exceeded $65 billion by the end of July 2026, up from approximately $9 billion at the end of 2025. As VentureBeat reported, Claude Code, Anthropic’s software development tool, reached $1 billion in annualized revenue run rate within six months of its general release in mid 2025 which implies, among other things, that “get paid to build software faster with Claude” is not just a good theoretical idea.</p><h4>Income paths the evidence actually supports ✅</h4><p>It’s clear the value lies with those who combine the use of Claude with professional domain skills and ownership of the output, not with those who sell “AI-generated” content as the product. Below are the strategies based on that principle, not on specific numbers (any site promising a guaranteed amount of money for a newbie freelance should be approached with a healthy dose of skepticism).</p><ol><li>Delivery of freelance services, powered by Claude. If you have a marketable skill (writing, research, analysis, accounting assistance, translations, paralegal work), Claude will be able to accelerate drafting and initial work, allowing you to take on more clients at the same quality level or spend more time on the high-value judgment calls that your clients will pay extra for. This is the underlying logic of the 34% hourly earnings difference in Upwork: it is not about AI writing better copy but about AI-literate freelancers doing more high-end work per hour. The Upwork statistics also show what needs to be avoided: a category of generic, low-effort “AI content” is becoming unprofitable. Position yourself as an expert in your field who uses AI tools, not “an AI freelancer.”</li><li>Software development and custom tools development using Claude Code. Claude Code is made exactly for that purpose! Turning plain English specification into working code, even for those who are not software developers. There clearly is a market for such services since Claude Code itself sees a rising number of customers — companies that are ready to pay for building internal dashboards, automatons, MVPs, small tools, when they do not have (or are unwilling to hire) an engineering department in-house.</li><li>Spreadsheet work, presentations, documents for businesses. Claude for Excel and Claude for PowerPoint target a very specific and concrete, if not glamorous, niche — small and medium businesses that need financial models, reports, presentations prepared or improved, but do not have the in-house resources to do that. It is largely overlapping with the traditional field of bookkeeping, financial analysis and business assistance services — Claude will only speed up the process.</li><li>Research, writing, and editorial work through Projects and Artifacts. The old school activities of content and research aren’t dead — the data in the Anthropic Economic Index shows that they’re still a fast-growing category, only they’re moving to higher-context and higher-trust form (structured reports, edited and fact-checked content, content based on the customer’s documents in the context of Projects).</li><li>Workflow consultation for small businesses through AI. Given the expansion into more mundane companies and individuals (supported by the data in the Economic Index about wage reduction for lower-wage tasks diffusion), the demand for setting up Projects, Skills, and Claude Code for customers too busy to learn how to use the platform themselves will be real and steady. It’s a service-oriented industry that hinges on your fluency in the platform itself, not a product you can package and sell forever and ever.</li></ol><h4>A realistic path to get started</h4><p>Select an existing skill you possess, not a new “AI skill.” Claude multiplies existing skills; this is a terrible foundation for itself.</p><ul><li><strong>Pick one skill you already have</strong>, not a brand-new “AI skill.” Claude amplifies existing expertise; it’s a weak starting point on its own.</li><li><strong>Learn the actual tools deeply</strong> — Projects for context, Skills for repeatable workflows, Claude Code if your work touches software, Artifacts for shareable deliverables.</li><li><strong>Find your first client or project through an existing channel</strong> — Upwork, your existing network, or a niche community in your field — rather than trying to build an audience from scratch first.</li><li><strong>Price for the outcome and your judgment, not for “using AI.”</strong> The data is explicit that undifferentiated AI-labeled work is losing value per contract even as volume grows.</li><li><strong>Track what’s actually landing.</strong> Treat the first few months as calibration: which services get repeat clients, which get one-off work, and adjust.</li></ul><blockquote>A caution worth stating</blockquote><p>Anthropic’s own report from its June 2026 Economic Index, based on interviews with 81,000 users of Claude, revealed that besides legitimate productivity benefits, there was legitimate concern about displacement — particularly prevalent among younger employees and in jobs where Claude is doing much of the work. Earning money through Claude makes sense, but it must be done within the context of a labor market being transformed by the same technology. Focusing on developing a sustainable niche based on judgment, relationships with clients, and expertise in certain domains makes much more sense than trying to compete on the speed of Claude-powered productivity alone.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=796c9c4d51c4" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>