skip to content
lambda
Table of Contents

The beginning

The first time I heard about OpenAI was during The International in 2017 (before OpenAI was cool). Where they created a bot which beats one of the world’s top professional Dota 2 player Dendi. You can find the full video here.

Later on OpenAI made a bot for 5v5 matches, called OpenAI Five which beats world champion team and in an open challenge for anyone it had a win rate of 99.4%.

I was around 15 back then, and my mind was blown. How could anyone even start writing a bot for such a complex game? How do you program all the rules, heroes, and balance patches? Knowing mostly procedural programming at the time, it was hard to wrap my head around.

Roughly 5 years later they announced ChatGPT, which gave me similar mind blown feelings.

Perfectionist’s Trap

I’ve read a lot about the benefits of writing, so I’ve wanted to start a blog for a while. But whenever I tried building my personal website in the past, I’d fall down the rabbit hole of designing the perfect site. This meant adding features like syntax highlighting, MathJax support, and using the newest web framework.

Trying to do all this in one go often led to frustration, and with how bad the javascript ecosystem tooling is, I often ended up with obscure errors. Many web developers know the feeling. Eventually, this frustration made me give up on blogging. My need to understand everything under the hood didn’t help either.

Letting Go

But since LLM got very capable and hearing the hype about “vibe coding” I decided to give it a try its been great. I no longer feel like Im missing out on crucial knowledge. I know the eventually there will enough data on a particular topic and an LLM would be trained on it. Its liberating realizing I don’t need to master every detail upfront.

Thanks to LLM I was able to actually start this blog, I started with a astro based theme and was able to customize it as I wanted.

It’s honestly liberating to know that you can store ‘almost’ all digital human knowledge in a single file on a pen drive.1

Before LLM, this wasn’t really possible because the information was scattered everywhere. Platforms like Stack Overflow and forums held pieces of the puzzle. But you had to know exactly what to search for. That’s not the case with LLM. You can be vague, describing your goal instead of the exact steps, and still get a good starting point. From there, you can choose: dive deeper yourself on important topics, or just trust the LLM and embrace the ‘vibes’.

Treat it as a tool

I think of LLM as lossy, compressed versions of the internet that you can talk to in natural language. But of course they aren’t magic bullets.

It’s important to understand how they work at a high level to know their limits.

While some worry about reduced critical thinking or the erosion of core skills, these concerns often apply to any powerful technology. Like relying on compilers instead of writing assembly, or using GPS instead of paper maps, the key is mindful usage.2

Blindly applying LLM suggestions without understanding the fundamentals can lead to more debugging time than learning the concept initially. However if wisely used, LLM accelerate implementation.

Treat them as a tool, understand their limitations, and leverage their strengths without letting them replace fundamental understanding.

Footnotes

  1. I’m not sure how efficient is the single file approach compared to other ways to self-host LLM. Feel free to reach out to me if you know about any kind of benchmark between those. ↩

  2. Comparison with compilers is bit of a stretch since they are deterministic while LLM are probabilistic tools, but there are many things which are useful while relying on probability ↩