100 million tokens. 5 truths.

We spent around 100 million tokens in the last few weeks. 

Mainly with Claude Code, but also others (Kimi, etc).

Here’s what we found out:

1. AI can help you ship faster. But not always better. 

We shipped entire code repos in a matter of days.

But a lot of it was garbage. Tests that don’t really test anything. Duplicated code. 

Sometimes garbage is okay. 

Like, for example, if you are working on an MVP that will never go live. 

But for commercial applications, garbage piles up quickly.

Soon your product becomes unusable. 

Even if it doesn’t break anything, dead code will increase your token costs. 

So, if you are working on anything in production, with real users, my advice would be to slow down a bit. 

Do regular manual cleanups. Keep your code as modular as you can. Review all the critical code (sensitive business logic, data handling, etc).

In the frontend, it is worth investing in a design system. 

You don’t want your app to look just like another Anthropic frontend.

Ohh, by the way, have you done our new Senior technical assessment? Bogdan added brand-new questions. 

2. If you just “vibe-code” without any understanding of the fundamentals, you are flying blind.

We had a silent failure in our application. 

A form submission wouldn't work. But after checking the network tab, we saw the HTTP request was okay. Nothing was wrong.

It turned out that Claude Code set the wrong “max-age” in Cache Control. 

Data was persisted in the backend.

But you couldn’t see it in the frontend because the frontend would hit the cache instead of making a new requests. 

It took Bogdan 60 seconds to understand what’s happening. 

No way he could have fixed this bug without a deep understanding of caching and the web browser. 

AI codes faster. But it also creates bugs faster. 

All the productivity gains you get by not having to code can be lost on fixing an impossible bug it just added.

If you don't have a solid understanding of the tech stack you are working with, you are just playing the Claude Code slot machine. Good luck!

3. AI means more code. More code means more Software Engineers. 

We started building 1 application. Now we have 3.

All of them have to be deployed, maintained, and even extended in the future. Bogdan has been working 12-hour days for 14 days straight. 

And it is still not enough.

We are building things we had in the pipeline for years. This would not have been possible without AI.

This will happen at the industry level. Companies will realize AI allows them to do more, but AI can’t come up with what to build. Let alone debug it. 

Software engineering hiring is going to have a huge comeback :) 

Developers with strong fundamentals, Senior-level thinking, and who can use AI in the most efficient way possible will be the biggest winners. 

We can help you, see if you qualify here! 

4. The SAAS apocalypse is partially true. Partially.

Bogdan and I have been evaluating the software tools we use at theSeniorDev. 

Around 120 of them.

Some of them charge a lot but offer little value back. We cancelled them and used AI to integrate those functionalities with other tools we already use. 

Do we want to do this?

Of course not. We are software engineers. We know how much work it takes to build good software. And how much work it takes to maintain it. 

But many software companies give you no choice. 

They increase prices and add zero relevant features. At some point you are better off building a substitute.

(we don't plan to do this in the future again unless we really see no other option)

But if you work for a tech company whose added value is not clear, it might be a good time to think about switching. There's a bloodbath coming.

5. AI is a bubble. It will pop. Yet part of it will stay with us forever. 

LLMs write shitty social media posts and shitty emails.

(and shitty landing pages too)

They also suck at creating anything new. They are data-driven models, and data is by default backwards-looking. The GTA V CEO has a good talk on this. 

But code is not content. 

A lot of it is repetitive and functional. Some parts are simply direct translations of the requirements to the code needed. 

Even if the AI bubble pops, that part will still be done by AI. Even tested by AI. 

And maybe reviewed by a human dev.

Whether the LLM provider will be Anthropic, a cheaper Chinese model, or even local AI is still to be seen. But it will be AI. 

Yes, software engineering is going through a massive change. 

We can help you use this to your advantage and come out as a highly skilled, confident, and irreplaceable Senior engineer. 

Check it out here!

See you next week with more updates on this 100 million token adventure. 

Ciao,

Dragos