We’ve all heard the term vibe coding by now. Instead of learning how to code, you just code vibes, aka, tell AI what you want and watch it replace your average web developer. Money saved. Business owner happy. Developer sad. But is it that simple?
Inception
AI coding expertise has come a long way. A few years back you had to manually copy-paste code snippets to-and-from AI chatbots and your IDE and either be surprised at how it could easily and quickly solve a problem that would take you hours while on the other hand, watch in sheer horror as it straight up hallucinated hot garbage with the confidence of a politician.
Nowadays there are AI IDEs of all sorts, from Cursor all the way to web based AI IDEs (like Bolt.new by Stackblitz). I’m not sponsored by either but I wish I was. Now you simply tell these AI agents what you want and it’ll cook up something for you. Heck, even ChatGPT, Claude and Gemini have mini-versions of AI agents in their own websites that can build and render basic sites and apps.
So what’s the catch?
I’m glad you asked. I knoe you didn’t but here you are so…
Anyway, at the core of any vibe codding tool or even any AI tool for that matter, are things called tokens. To put it simply, these are the words (and/or things, data basically) you give an AI, and AI in turn outputs tokens (again, word and/or things). Say you ask ChatGPT what is your name, your prompt is convereted into tokens, AI processes those tokens and its output is also tokens but converted into the data you want, which in this case would be “Hi, I’m ChatGPT.”
Now that you know about tokens, you should know about context windows. Different AI models have different context windows, which is basically the limit on the number of tokens it can handle (both output and input). Ever talked with an AI for long and realize it’s getting dumber? This is because the longer your conversation grows, the AI starts to “forget” the initial parts of the conversation intentionally. Say you have a studio room and can only host 20 people. 5 more guests come, 5 initial guests will have to leave. That’s basically it. But this is an oversimplification. (There are different tokenization methods and different models handle context windows differently).
No, for real, what’s the catch?
Now that you know the basics of how AI works, you can start to see where I’m going with this. Initially when a new model comes out, the creators are usually generous with their token limits (well, besides Claude, lol). But over time, as popularity increases, these companies realize they’re burning money on the free tokens, so they start doing rate-limiting.
Remember when these companies boasted how they can replace entire development teams? Well, it’s not entirely moot but these subscriptions for the AI tools can easily get crazy. Especially on big projects. Heck, even thanks to enshittification, even small projects can be barely completed without being rate-limited and then being offered to upgrade to an even higher tier, only to be rate-limited in a few hours. And sometimes, the rate limits can stay for days. I’m looking at you Antigravity.
Anyway, just like any subscription model, prices can only ever go up. The sky is literally the limit. You think $100-$200 subscriptions are crazy. Look around, people are 100x-ing that in token usage.
So what’s the future?
Either hope your SaaS/business makes enough money to pump your funds into these AI tools or maybe just maybe, it’s still not yet time to put down the software engineering career. But who knows? Maybe this is the future.
Emerald, out!