The Inktelligence - April 14, 2025

There has been a lot of hype recently about "vibe coding," a term initially coined by Andrej Karpathy in February 2025. Here’s what he said about it:

In practice, Karpathy meant a specific workflow where:

1. He would speak to AI coding assistants using voice recognition rather than typing.

2. He would ask for simple changes like "decrease the padding on the sidebar by half" because he was "too lazy to find it".

3. He would "Accept All" changes without reading the diffs.

4. When encountering error messages, he would copy and paste them to the AI without comment, usually resolving issues this way.

5. The resulting code would grow "beyond [his] usual comprehension" where he'd "have to really read through it for a while" to understand it.

I get it. Because I’ve done the same thing myself.

The term represents a significant shift from traditional programming, where the programmer's role changes from manual coding to guiding, testing, and refining AI-generated code. Karpathy specifically noted that this approach was "not too bad for throwaway weekend projects," suggesting he saw limitations to its applicability in more serious development contexts.

However, across social media, I've noticed many people promoting a different idea - that anyone can create complex applications with AI coding tools, even without programming experience. This misinterpretation shifts focus from understanding code to relying purely on intuition, as if technical craft no longer matters.

Improvisational Coding

Perhaps this would be a better term. It reminds me of jazz musicians. Top improvisers spent years drilling scales and practicing fundamentals before improvisation became second nature. Their spontaneity is built on deep technical knowledge, not shortcuts. "Vibe coding," when misunderstood, attempts to skip this essential foundation.

It suggests you can simply prompt your way to working software. But when systems inevitably break, troubleshooting requires the very knowledge that was bypassed. True mastery doesn't mean doing less. It means knowing enough to use tools properly. AI can enhance programming, but it works best as an extension of understanding rather than a replacement for it.

Recently I used Cursor to help me create an application that will take the LinkedIn posts from a specific creator and and show me a summary of their tone of voice. The left side shows all the names that are available and when I drag and drop the name to the right side, the saved analysis is displayed.

The application has the front end you see below and the backend is a Supabase database that stores all the names and analysis.

Doing this the old fashioned way before generative AI coding tools like Cursor came around would have taken me at least 10 times longer if not more.

The beauty of AI coding tools is that it enables the democratization of coding. We will see the long tail of software applications as general purpose SaaS become highly customized applications tailored to particular business segments and use cases.

Now instead of Hubspot, you’ll see industry specific CRM applications customized to the needs of each vertical. It opens up the ability to take something from idea to working software in a weekend even for those who are non-technical. But one must use it wisely.

Finding the Middle Ground

My experience illustrates why "vibe coding" has captured so much attention. The ability to rapidly transform ideas into working software represents a genuine breakthrough in productivity. For personal projects, prototypes, and many business applications, these tools offer tremendous value.

What concerns me isn't the technology itself, but rather the messaging that sometimes surrounds it. The notion that understanding code fundamentals has become optional can create false expectations and potential problems down the line.

In my LinkedIn analyzer project, I still needed to:

• Work out the complete workflow for what I needed it to do. (This started in my brain)

• Come up with a way to obtain the posts without scraping LinkedIn or using the API (Research and AI helped)

• Understand database relationships to properly structure the Supabase backend

• Identify and fix errors while testing (of course AI helped a lot in this regard)

My prior programming and troubleshooting knowledge, even if not specifically in Javascript, provided the foundation for effective problem-solving when things didn't work as expected.

A Practical Approach

Perhaps the most productive way to think about "vibe coding" is as a spectrum rather than an all-or-nothing proposition. The appropriate level of technical understanding depends on:

1. The complexity of your project

2. Its intended lifespan and maintenance requirements

3. Security and performance considerations

4. Whether it's for personal use or widespread deployment

For a weekend project or personal tool, embracing the "vibe" approach makes perfect sense. For mission-critical business applications or products with many users, a deeper technical foundation becomes increasingly important.

AI coding tools don't eliminate the value of programming knowledge but they transform how we apply it. They allow us to operate at a higher level of abstraction while still benefiting from foundational understanding when needed.

The true power of these tools emerges when we recognize both their capabilities and limitations, using them as partners in creation rather than complete replacements for technical understanding.

Midjourney v7

The early version of Midjourney v7 introduces a few improvements. I particularly like the speed of draft mode and also the ability for conversational voice prompting. Here are some of new features:

1. Draft Mode
A new feature called "Draft Mode" allows for rapid image generation at half the usual cost and approximately ten times faster than standard rendering. This mode is ideal for quickly iterating ideas and refining prompts before committing to high-quality outputs. ​

2. Personalization Profiles
V7 introduces personalization profiles, enabling users to tailor the AI's output to their aesthetic preferences. By rating around 200 images, users can train the model to align with their unique visual style. ​

3. Enhanced Realism and Detail
The new model offers improved realism, with better rendering of textures, lighting, and anatomical details such as hands and facial features. This results in more lifelike and coherent images. ​

4. Improved Prompt Interpretation
V7 features a more sophisticated understanding of text prompts, allowing for more accurate and contextually relevant image generation. This enhancement reduces the need for overly complex prompt engineering. ​

5. Compatibility with Previous Versions
The new version maintains full compatibility with V6.1's style reference codes (sref), ensuring that users can continue to utilize their existing style libraries without modification. ​

6. Voice Input and Conversational Prompting
Users can now input prompts via voice, facilitating a more conversational and intuitive interaction with the AI. This feature streamlines the creative process, allowing for real-time adjustments and refinements.

They still need to improve text generation. Here’s a comparison of the same prompt in Midjourney v7 and ChatGPT.

Happy 100th birthday to Bell Labs!

For a century, this legendary institution has been the quiet engine behind some of the world’s biggest breakthroughs such as transistors, lasers, information theory, Unix and much more.

I’m glad that I had the opportunity to spend part of my career at Bell Labs.

The work at the labs didn’t just shape technology - it shaped the future. Here’s to the brilliant minds who made it all possible, and to the next 100 years of bold ideas, wild experiments, and world-changing innovation.

Quote of the week

I’d like to hear from you!

If you would like provide me with any feedback, suggestions, thoughts, discoveries, or just say hello, please write to me (I promise that I will read every single email): [email protected]

Share this newsletter with your friends