Vibe Coding WhomHow When: [00:00:00] Hey, and welcome back to Easier Than than Done. Today. I'm gonna talk about vibe coding. As a lawyer who's turned into software engineer in my late thirties mainly to keep my startup alive, I have directly witnessed over the last 10 years the immense seismic shifts in the coding world, right over the last decade.
Vibe Coding WhomHow When: Now, today, I'm excited to explore with you on this episode, the latest paradigm. Coined by Andrej Karpathy, and I'll tell you who he is in a second.
Vibe Coding WhomHow When: In a recent tweet, he calls it vibe coding, right? So he said in essence, and I'm reading from Andrej Karpathy's tweet here.
Vibe Coding WhomHow When: There's a new kind of coding I call vibe coding, where you fully give into the vibes, embrace exponentials, and forget that the code even exists. It's possible because LLMs are just getting that good and I'm skipping over some of the parts just to keep it shot.
Vibe Coding WhomHow When: He coined this term recently in the tweet. And before we dive into what Vibe coding is, let's first understand who Andrej Karpathy is because he's the one who's describing his own experience with Vibe coding and how this is how he does it now. So let's appreciate a little bit more about him and [00:01:00] his background.
Vibe Coding WhomHow When: 'cause not everyone can just do this.
Welcome to Easier Said Than Done with me, Zubin Pratap, where I share with you the tens of thousands of dollars worth of self development that I did on my journey from 37 year old lawyer to professional software engineer. The goal of this podcast is to show you how to actually do those things that are easier said than done.
Vibe Coding WhomHow When: So let's appreciate a little bit more about him and his background.
Vibe Coding WhomHow When: 'cause not everyone can just do this. So Andrej Karpathy is absolutely a titan in the world of ai, right? He's one of the biggest household names, if you like, in AI and deep learning, and I'm gonna link to his GitHub as well in the show notes. But you'll see there's not a lot there in the public stuff, but you can see the depth of his experience in just how long, how many decades he's been doing work in the AI and neural network space.
Vibe Coding WhomHow When: Now he's got well over a decade of experience at the forefront of AI research and development has actually left a very lasting mark on this field. In fact, he's probably best [00:02:00] known for some of his work in computer vision, neural networks and so on, and was also the former director of AI at Tesla. Now, you know, Elon's not an easy guy to work for, right?
Vibe Coding WhomHow When: And at Tesla he spearheaded the development of autonomous driving systems. So we know from this that Karpathy has deep technical expertise and years of hands-on engineering. This makes him uniquely qualified to identify and name or give a label to his style of coding, which he calls Vibe coding, which is done with AI LLMs 'cause he's been using it for years.
Vibe Coding WhomHow When: And here's the key point folks, all that experience also makes him capable of using the technique of vibe coding.
Vibe Coding WhomHow When: So let's talk about vibe coding as a technique, right? Vibe coding is basically the technique where you use ai, LLMs large language models to generate code based on descriptions or vibes that you provide.
Vibe Coding WhomHow When: But that means you need to know how to tell it to do the thing you want to do. That takes deep expertise, right? It's like this. Think of what you do at work [00:03:00] today, right? And you have a gifted member on your team who's very good at translating your experience and top judgment- based ideas into an actual reality.
Vibe Coding WhomHow When: It's kind of the same thing, right? You treat the AI LLM as a very talented member of your team, but you are supervising it, right? And you tell it to translate your engineering vision into functional code. And it can do that very fast faster than human beings that actually doing the typing parts of it, right?
Vibe Coding WhomHow When: But you have to give it the clear engineering vision and it can't be done at a very high level like. Build me an autonomous car, like it's not gonna work for that. You've got to know enough of the whole system that you can break down all components of the architecture, every piece of it. Keep breaking it down to smaller and smaller pieces, and then outlining that outcome and piece to the LLM
Vibe Coding WhomHow When: and then being able to supervise its output.
Vibe Coding WhomHow When: So let's talk about how vibe coding differs from traditional coding Number one. Vibe coding is very outcome driven, right? Traditional coding requires [00:04:00] the coder to make thousands of keystrokes typing away to type out the code. And in LLMs or in Vibe coding, we kind of describe the outcome and it does the typing for you.
Vibe Coding WhomHow When: Now, when you're good with your instructions, it can even design that code effectively. So it's not just the typing, but actually the design of the interfaces and the structuring of the code can be done quite effectively. But you've gotta know how to give the instructions and break it down at a small enough piece and use the right terminology.
Vibe Coding WhomHow When: Number two is that the LLMs take a natural language input, right? Vibe Coding takes a natural language input. Now, this is because VIBE coding interprets high level intention based on the natural human language that you're using, and then vibe coding therefore can allow for more conversational prompting rather than a strict syntax based approach.
Vibe Coding WhomHow When: So that's another big difference from traditional coding, which is very much about the strict syntax and actually typing out the code and knowing the language inside out yourself.
Vibe Coding WhomHow When: Number three, rapid iteration is possible with vibe coding because vibe coding can generate and refine code under your [00:05:00] supervision when you tell it to much faster than manually typing by a human being.
Vibe Coding WhomHow When: So that means you compress the development cycle. So what could otherwise take a week of research and typing... now, if you know enough to give it the right instructions, you can do it in a few hours, and then spend a few hours fine tuning it and refining it yourself.
Vibe Coding WhomHow When: And the number four difference is that there's contextual understanding because LMS are really good at what's known as semantic understanding, right?
Vibe Coding WhomHow When: It's a contextual understanding. That's why you give context filled prompts to an LLM. From that LLM can infer a broader context and best practices, unlike traditional IDEs where it's just completely an interface for you to type things, right?
Vibe Coding WhomHow When: And LLMs are getting really good at contextual understanding.
Vibe Coding WhomHow When: For example, I can talk about something like he or she's really hardcore, but if the previous and next sentences give more contention to what that person is, hardcore or bad, it'll know what it means. It may not take it as literally like a hard core of a for mango or the [00:06:00] Earth, which has liquid course, not a hardcore, but you know, it, it can infer the context.
Vibe Coding WhomHow When: It can say, Hey, I know this person's a hardcore guitarist. Why? Because the previous sentence said that, or the next sentence said that, or, this person's a hardcore engineer because the previous sentence or previous paragraph said that. LLMs are really good now figuring out context, which is why it's being used in search even more right .
Vibe Coding WhomHow When: Now the fifth difference is multi-language fluency. Vibe coding can essentially switch between programming languages because it is trained on pretty much the entire internet, right? It has all of the internet and its training data, and it can therefore
Vibe Coding WhomHow When: actually provide the answers for you from multiple programming languages based, just based on your prompt.
Vibe Coding WhomHow When: Now, the number six difference is that it reduces all. Sometimes even eliminates research, right? The developer, the traditional developer workflow requires multiple windows, multiple tabs going down, all these Google search rabbit holes you know, looking at lots of stack overflow, looking at old documentation, reading old code, trying to figure out how things are done.
Vibe Coding WhomHow When: That's all part of the research process. That's very much the most time consuming part of traditional software [00:07:00] engineering. When you don't know the code base or you're building something new, these approaches, these design patterns that you need to research. The API documentation is extremely time consuming and there's a lot of trial and error involved.
Vibe Coding WhomHow When: Now the LLM is trained and synthesizes all of this. All of this, right? And does it much faster than a human can. And so it eliminates or compresses the research window required, and then all engineering needs to do is look at it and says, yes, that fits all the patterns I've seen, and therefore this is probably the right approach.
Vibe Coding WhomHow When: So it's a judgment call that comes from experience.
Vibe Coding WhomHow When: Now having used this approach of vibe coding for almost a year now, I can confirm that it is an absolute game changer, right? It's like you are the chef and you've got a savant sous chef, someone who really understands your vision and flavors and can execute complex tasks with the right instruction.
Vibe Coding WhomHow When: But there's a catch.
Vibe Coding WhomHow When: You've got to be experienced enough to know what you want, know what you don't want, and recognize bad on output, which means it's all about the [00:08:00] judgment.
Vibe Coding WhomHow When: Now, here's why Vibe coding is fantastic for seasoned pros. One, it reduces drudgery . Right. What does that mean? All the boring parts are taken out, okay?
Hey, hope you're enjoying this episode. Listen, I bet you're tired of the social media BS. You're confused and overwhelmed. And all you want is for someone to just show you a clear path to a coding career. Right? We get it. My partner, Brian and I. We changed to code in our thirties and we can help you do it too.
Brian and I have done all sorts of engineering work from startups to Google, and now we've built an exclusive private mentorship program designed to get you from beginner to professional developer, whether you've done a coding bootcamp, you have a computer science degree, or you're starting from zero, it doesn't matter.
We'll build you a step by step and customized plan not just for the coding stuff, but for the entire career change process Look, i'll be honest. The program is not easy and it's not short But if you do the work, you'll get results that [00:09:00] will change your life If that's what you want go to parsity. io slash inner- circle or find any of the links in the show notes And you can schedule a free call with us speak to you there
Vibe Coding WhomHow When: it reduces drudgery . Right. What does that mean? All the boring parts are taken out, okay? You don't have to spend hours or days building up your environment and putting up scaffolding code and stuff. Like all of that's gone.
Vibe Coding WhomHow When: Number two, you can focus on architectural engineering decisions, not just the development slash coding side, right? So you can spend more time on very strategic, high level designing of the system and less on the writing of the boilerplate
Vibe Coding WhomHow When: that's in all the crud that has to go into a modern application.
Vibe Coding WhomHow When: Number three, you get the benefits of pair programming without an actual human pair programmer now done right. You and the LLM can effectively be pair programmers.
Vibe Coding WhomHow When: You discuss and explore patterns together. You approach interrogate and question techniques, you analyze other techniques, and you analyze the code base together as well.
Vibe Coding WhomHow When: Very powerful stuff, but you need to know what you're doing.
Vibe Coding WhomHow When: Number four. [00:10:00] Is that seasoned engineers get continuous learning much faster this way.
Vibe Coding WhomHow When: Now, when your fundamentals are strong enough, and that's a prerequisite, then you can learn new things much faster because the research time is exponentially reduced or compressed. And then you interact with the LLMs to synthesize all their research data into one place. You have the judgment and the intuition to know what feels right and what doesn't.
Vibe Coding WhomHow When: And so you don't need to yourself go to primary and secondary data sources. You can work with the LLM to build your learning fast.
Vibe Coding WhomHow When: And the fifth advantage for seasoned engineers is that you can do rapid prototyping, which you talked about earlier. It's very fast to do spikes. You, you compress through the development cycles and you can iterate through an idea very fast.
Vibe Coding WhomHow When: You can generate throwaway code every single day and get better and better, much faster, right? Huge benefit for experienced engineers, but knowing what is better than the previous situation requires judgment, which comes from experience.
Vibe Coding WhomHow When: Now for novice engineers, vibe coding is frankly very tempting. But very treacherous and dangerous.
Vibe Coding WhomHow When: It's kind of like giving a novice driver the keys to [00:11:00] an F1 car, right?
Vibe Coding WhomHow When: A Formula one car. It's happening right now in Melbourne in March, 2025, right now, this weekend, in fact. And the power of it is intoxicating. The power is incredible. The glory is incredible. The, the fun, the glamor is awesome, but without the foundational skills of knowing what you're doing and knowing the track you're on, it's kind of a recipe for disaster, right?
Vibe Coding WhomHow When: It's the same thing for inexperienced coders. Because the dangers can include things like one, you end up having a dependency or a crutch. You rely very heavily on the ai.
Vibe Coding WhomHow When: I've seen this a lot in interviews actually, where people can't explain the code that the AI generated. They understand what it's kind of doing, but they don't know why it's there, and that's an essential part, right?
Hi, if you want a no BS insight into how to change your career, whether to code or something else and how to actually get job opportunities in tech, then please subscribe and like.
It's no BS because I have zero incentive to mislead you. I just want to help you and give you tons of value so that you will consider working with me to get to your next career.
Take a look at the
description text below to learn more about the training I offer. But I do post content here regularly and by subscribing and liking and hitting that notification bell [00:12:00] you will get to know when I post new industry insights for you. You'll also know within about three seconds if you want to learn more, but at least you won't miss out.
Oh, and please follow me on LinkedIn too. I pretty much post there every single day. Just look for my name, Zubin Pratap. All right. So please like, and subscribe to this channel and let's get back to the episode.
Vibe Coding WhomHow When: Number two, there's a lack of understanding because you're building you are creating all this output without actually understanding the engineering principles, the design principles, the human experience principles, and this kind of inflates what's known as the Dunning Kruger Effect. If you look at my LinkedIn blog, which is based on this on this podcast, I've linked to what the Dunning Kruger Effect is.
Vibe Coding WhomHow When: Just Google it, you'll see it on, on, on Wikipedia. Human humans are very prone to Dunning Kruger, which is basically a way where we tend to overestimate our own skills and competence at something. Right now, when we fall prey to Dunning Kruger, it leads to very disastrous professional and personal consequences.
Vibe Coding WhomHow When: And that's a big pitfall I'm seeing with junior engineers relying too much on AI or relying on ai. Full stop. Number three. The difficulties that junior engineers, or [00:13:00] rather inexperienced engineers, I don't really care for the term junior, but inexperienced engineers have significant difficulty debugging systems because they never actually learned what the, because they never actually learned through trial and error how that system got put together.
Vibe Coding WhomHow When: And when you're struggling to fix issues in the code that you didn't write, this is a real problem. Okay, this is a real problem 'cause you don't have the intuition, the judgment, the pattern recognition or the ability to reason through a code that has produced some bugs that somebody else may have written or you may have written or the AI generated and there's bugs in it.
Vibe Coding WhomHow When: Okay, and the number four challenge of when inexperienced engineers rely too much on AI is that it creates a lot of technical debt and you don't know it because you don't have the judgment, insight, or experience to understand what the technical debt is. You have an accumulation of poorly structured or design code becomes very hard to manage.
Vibe Coding WhomHow When: Hard to debug. Hard to maintain. And if you're lucky enough for your product or your app to start, getting users, it's very hard to scale when you don't understand your technical debt.
Vibe Coding WhomHow When: Now to conclude vibe coding absolutely [00:14:00] favors the experience because experienced people have many compound skills that go into do into their skill, right?
Vibe Coding WhomHow When: Whether it's in the law, whether it's in music, whether it's in software engineering, whether it's in journalism, whatever it is, I, and I've been very close to all these professions. It's not actually the core skill that matters. It's the compound skills of the core skill plus dozens of other skills, right?
Vibe Coding WhomHow When: So you may be a really good coder, but if you don't understand business or user requirements or client requirements or internal operations, you're probably not gonna be an effective, impactful coder, or vice versa. Maybe you understand all those things, but you're not very good at the code stuff. It's going to impact your ability to have impact, right?
Vibe Coding WhomHow When: It's true of all professions and software engineering is very much a compound skill. Very much a compound skill. Now, the problem with the over reliance on AI is that it kinda shifts the balance and dependency on those skills, right? Experience Coders get to benefit from vibe coding because they possess the compound skills that were developed
Vibe Coding WhomHow When: through years of exposure. Inexperienced engineers can be expected to
Vibe Coding WhomHow When: suffer through [00:15:00] over lines on AI due to the absence of those compound skills.
Vibe Coding WhomHow When: So let's talk about some of those compound skills. One, pattern recognition, the ability for seasoned engineers to quickly identify and spot issues and anti-patterns.
Vibe Coding WhomHow When: Number two, architectural insight, experience allows for better high level design understanding that then leads to better breaking down of the problem of the entire system into its tiny components, which then lead to better prompting of the AI
Vibe Coding WhomHow When: and the third compound skill that I think is really important:
Vibe Coding WhomHow When: Is that you lose the ability to debug effectively, right? And that's critical because most of the time in, in, in, in a job situation, you are not writing brand new code. Most of the time you're looking at existing systems that hundreds of people have written before you, that have been around, you know, for 20, 30 years in some cases, and you have to figure out what's going on.
Vibe Coding WhomHow When: Number four is quality assurance, right? This is a compound skill. Experienced engineers can critically evaluate and improve AI outputs.
Vibe Coding WhomHow When: Number five, ethical and business considerations. Seasoned professionals are much more likely to understand the nuances [00:16:00] and implications of the code in the real world when it actually gets released into the real world.
Vibe Coding WhomHow When: Number six. Engineers have the compound skills of being able to know and to challenge the AI and how to interpret it in its obsequiousness right now.
Vibe Coding WhomHow When: What I'm gonna do very quickly is I am going to share my screen for just a moment so you can see what I mean by this, right?
Vibe Coding WhomHow When: For those of you who are on YouTube, you get to watch this on My screen, 'cause I'm gonna share my screen with you and show you just. How the AI, when challenge does weird things, okay, so in this example, I actually lifted Code Golang code from the Go By Example website, which is a beautiful starting resource for people who wanna understand the basics of the Go language.
Vibe Coding WhomHow When: All right, so you guys can see this now. So what happened is I said, I pasted it in the code from go, go by example. And it's, it's in this big prompt here, right? And I said, what does this following code do? It analyzed the code, it broke down the code, and it actually did a good job. It was correct, but it's pretty simple fundamental code.
Vibe Coding WhomHow When: So fundamental Golang code. So it was able to correctly [00:17:00] break it down. Absolutely right. No problem. Right? But I decided to challenge it. I said that's wrong. You've analyzed the work of function incorrectly. So I gave it a vague criticism, right? And it said, you're absolutely right. I apologized for the error in my analysis, right?
Vibe Coding WhomHow When: I misinterpreted it, didn't it actually got it right and then it went on, and then I continued to challenge it. I said, you've also made an error in how the channel's data gets consumed. Again, vague, not using the right terminology, but you know, I'm trying to simulate what it would be like for a less experienced engineer and who thinks that it's wrong, even though it's correct.
Vibe Coding WhomHow When: And it said, you're absolutely right. I apologize for the repeated errors. There were no repeated errors. Let's get this right, and it changed a bunch of its own analysis, so.
Vibe Coding WhomHow When: I don't know about you guys, but that makes me nervous because that means if I don't know how to tell, right for wrong, or I don't have an intuition that says, let me research this outside the LLM, because the LLM is basically doing the research for you and taking away your ability to learn how to do your own research, even though it takes more.
Vibe Coding WhomHow When: Much more time. So it gives you convenience at a great cost where you have to trust either IT or your gut, [00:18:00] right?
Vibe Coding WhomHow When: So this is just an illustration of why experienced engineers are much more likely to get more out of vibe coding because they have the compound skills and the intuition to know when and how to challenge the AI .
Vibe Coding WhomHow When: As with any powerful tool, the key to LLMs lies in how we wield the tool.
Vibe Coding WhomHow When: Experienced engineers absolutely can use it as a force multiplier to amplify existing skills and knowledge and judgment. They know when to trust. They know when to dig deeper. They know when to challenge it.
Vibe Coding WhomHow When: Like, like a master musician knows when to improvise in a gig, right? When someone's brand new, they need to read it off the sheets or get, go completely by rote memory. But an experienced engineer knows how to. An experienced musician knows how to improvise and when to stick to the score. That's the big difference.
Vibe Coding WhomHow When: And that comes from judgment and experience and a lot of effort and learning, right?
Vibe Coding WhomHow When: So for those starting your journey in coding, if you're an inexperienced developer, honestly, my advice is this, learn to walk before you. Learn to run. Don't be seduced by the convenience and ease of the AI into thinking that that's what engineering is all about.
Vibe Coding WhomHow When: I can [00:19:00] promise you.
Vibe Coding WhomHow When: The actual typing out of the code is the least important part of the job of the engineer. It's knowing what needs to be coded. That's the important part, very important difference. So
Vibe Coding WhomHow When: for you, newbies out there, please build a solid foundation in programming the hard way. Take the hard road. Don't take the easy road, okay? Build those programming fundamentals. Understand the different data structures available, the design patterns, and understand the WHY behind the code, why it's there, not just what it does.
Vibe Coding WhomHow When: But why is it needed?
Vibe Coding WhomHow When: Use AI as a learning tool and a partner, but assume you're the boss and it's reporting to you and you are accountable to supervise it, right? That way it's not a crutch. You have to know what it's doing.
Vibe Coding WhomHow When: Now at the Inner Circle program that Brian and I run, links are in the description.
Vibe Coding WhomHow When: We emphasize the fundamentals of programming. It has nothing to do with the language. It has nothing to do with data structures and algorithms.
Vibe Coding WhomHow When: At this stage, the fundamentals of programming is about the fundamentals of building complex tools and systems on computers. We teach that for a sustained period of time because it takes time to get that level of [00:20:00] judgment and insight.
Vibe Coding WhomHow When: And then we train people on the safe and appropriate usage of AI LLMs, right? Because LLMs must be used as though they are employee, not as though they just do the work for you that you don't know how to do. You've gotta be the boss, right?
Vibe Coding WhomHow When: So in conclusion, vibe coding is by no means a replacement for expertise.
Vibe Coding WhomHow When: It's an enhancement and can only be safely wielded by someone with expertise. It's a powerful ally or a minion, if you want to think of it that way. For but to be used by people who have paid their dues and understand the craft. For the newcomers here. This is a good thing. It's a glimpse of what's possible when you achieve mastery.
Vibe Coding WhomHow When: And at the Inner Circle program, I really focus, go very deep on mastery, right? It's very mastery based learning. It's a goal that you can work towards. AI is not a shortcut to bypass the learning process, okay? It really isn't. So the future of software engineering is bright because AI makes us much more
Vibe Coding WhomHow When: fast. Sure. It may mean that we need fewer engineers, but I don't think that's [00:21:00] typically what happens, right? The invention of the car didn't mean that that people needed fewer cars when public transports available. All that means is when people can move around faster, more things get produced that require moving around more.
Vibe Coding WhomHow When: And similarly, all AI is gonna do, I believe, is it's going to generate this explosion of opportunity of businesses of of new things coming up and shutting down, and this evolutionary process that's going to have a huge demand for people who have strong programming fundamentals and know how to use AI to accelerate what was already being done.
Vibe Coding WhomHow When: So it's more like a jet pack for an engineer, not a replacement for an engineer, and of course, there are no shortcuts. Please embrace the journey, embrace some of the suckiness of it, embrace the hardship, and really, really commit to your mastery .
Vibe Coding WhomHow When: I'd love to know from you what your experience with AI has been, what your experience with Vibe Coding has been. So please, please drop them in the comments, tag me and I'll do my best to respond. I usually respond to all comments. I will see you there. Alright, hope you got a lot out of this. Take care. Bye. .
Just subscribe, [00:22:00] you know you gotta do it.
