Karachi Wala Developer

Have you ever asked an AI to build something, only to realize it built the  exact  wrong thing? AI is incredible at generating code, but it lacks human judgment. In this episode, I share the story of building an internal AI coaching bot called "Jugnu." My first attempt failed completely because I relied on the AI's judgment instead of doing the research myself. I quickly learned that...

As AI agents become capable of rewriting entire C compilers and Next.js repositories in days, the value of software is shifting away from the code itself. If LLMs write probabilistic code, how do you guarantee deterministic results? The answer: Tests. In this episode, we argue that developers need to stop agonizing over AI-generated implementation code and instead double down on writing, reviewing...

The role of the developer is changing. In this episode, we discuss how to stop being a "consumer" of AI prompts and start becoming a "producer" of AI systems. By shifting your focus from writing code to designing autonomous loops and persistent architectures, you can move up the ladder of value and start steering the future of software development.

As software engineers, we are trained to control every detail—from code formatting to system architecture. But autonomous AI demands the exact opposite. In this episode, we discuss the "Autonomy Paradox." Learn why enforcing arbitrary rules on AI agents causes them to fail, and how shifting your focus from "how it's built" to "what it achieves" is the key to unloc...

If you are dumping your entire codebase into Claude or Gemini and hoping for the best, you are doing it wrong. In today's episode, we tackle  Context Engineering . AI agents have short-term memory (the Goldfish Problem) and limited Context Windows. If you fill that window with unnecessary data or heavy MCPs, the AI's performance drops drastically. Tune in to learn how to structure your inf...

In this episode, I share a personal story of building an ERP system with AI in the field, and how it forced me to confront the "5 Stages of Grief" as a veteran software developer. The way we have written code for the last 20 years is ending. To survive, we have to unlearn our "deterministic" mindsets and adopt new mental models for a non-deterministic world. Join me as we kick...

Dive into the much-anticipated GPT-5 update. We cut through the hype to analyze OpenAI's latest offering, covering its new unified architecture, efforts to reduce hallucinations, enhanced memory, and the controversial decision to deprecate GPT-4. Is GPT-5 the leap towards AGI we've been promised, or just an incremental step? We compare its capabilities against Gemini, Grok, and Claude, and...

Dive into the AI coding revolution! This episode explores how AI is transforming software development, from rapid prototyping and coding assistance with tools like GitHub Copilot, to the rise of generative coding. Learn the key trends, essential tools, and practical ways you can leverage AI to build faster and smarter. Perfect for developers and anyone curious about the future of code.

Ever wished your AI could actually  do  things directly from your terminal, beyond just chatting? Google's new Gemini CLI is here, and it's a game-changer! Unlike previous closed-source tools, this open-source power tool brings the intelligence of Gemini Pro right to your command line. We're talking direct file system access, automated project setups, and a powerful "reason and re...

Are Large Language Models (LLMs) truly intelligent, or just sophisticated pattern matchers? This episode dives deep into a fascinating debate sparked by Apple's recent research paper, which questioned the reasoning capabilities of LLMs. We explore the counter-arguments presented by OpenAI and Anthropic, dissecting the methodologies and the core disagreements about what constitutes genuine inte...

Dive deep into the world of AI communication with this episode, focusing on the game-changing MCP (Model Context Protocol). Discover how this protocol, pioneered by Anthropic, acts as a universal translator, enabling AI models to seamlessly interact with external services like Google, Twitter, or even your private databases. Learn why MCP was developed to overcome the limitations of static API doc...

Dive into the rapidly evolving world of Large Language Models! This episode breaks down the latest updates on cutting-edge models like Gemini 2.5 Pro (with its 2 million context window), the ambitious Llama 4, GPT-4o's impressive image generation, and Cloud's Sonnet 3.7. We explore key trends such as expert architectures, the nuances of leaderboard performance (including Llama 4's cont...

In this episode - I take on the task of explaining what vibe coding is, and if we should keep studing software engineering. I introduce some vibe coding tools, including the latest one Firebase Studio from Google

Last week OpenAI released a new model called O1 - which has advanced capabilities to reason and does a lot better in maths and coding. How does it do this and what does this change? This and more in this episode.

In the world of many open source LLM models, Google Gemma stands out with some unique capabilities. Firstly its made by Google after all - so its implementation is fully optimized and memory efficient. This and more on what is Gemma about and how its better.

This episode dives into the exciting world of on-device AI, specifically Large Language Models (LLMs) running directly on your smartphone! We'll explore why cramming this powerful tech into your pocket is a game changer, and discuss how LLMs can transform your daily interactions with your phone.

In this episode we cover the latest updates from Google IO and OpenAI Spring update. The all new ChatGPT-o and Gemini 1.5 Flash. We talk about where are LLMs headed next.

Lama3 is a popular open source model by Meta, the company behind Facebook, Instagram and Whatsapp. They have release the next iteration of the famous Lama2 model, this is a popular open source LLM which is heavily used by developers world wide. Is the new iteration strong enough to take on the best closed source models? We explore in this episode.

As we move to use LLMs more and more for different use cases, we will run into at times issues or limitations with the model. There are many different ways to train the model to do a specific task better, and fine tuning is one of them. What is about, when do we need to use it and how do we do it - all in this episode.

This episode covers my journey as a GDE over the last few years, what got me into it, and what is the big outcome. Also joined another category for GDE in Machine Learning.

In this episode I share my thoughts on the latest additional to AI development called Deven, and where the industry is moving towards. Will software engineering be taken over by AI in the next few years?

So this week we had 2 back to back announcements which completely shook up the LLM worlds. Gemini Pro 1.5 with its 10 million token context length makes it super powerful to run on large codebases and a lots of documents. The other big news was the launch of OpenAI Sora - their new text to video generation which can now generate super realistic videos for up to a minute. How will these impact our...

The all new heavy weight model has just dropped this week. Should you be rushing to try it out? Why is the ultra model better? This and more in his quick episode talking about Gemini Ultra.

In this episode - we discover the Retrieval Augmented Generation (RAG) process which allows us to train our LLMs using our data. In this short episode we cover what is this about and how can we use it.

If you are using ChatGPT or Bard - you are "prompting" the AI to get some answers. Not all prompts are made equal. The response of the LLM depends heavily on the quality of the prompt and the context given to it. Thus its worth while spending some time understanding prompt engineering and how to do it properly.