Own your workflow

I don’t play that much video games anymore. Especially since I grew a bit older. I have less time, my precision declined, and the perspective of getting 360 no-scoped by some teenager playing 10 hours per day doesn’t appeal to me that much.

However, I found joy in playing the recently released Slay the Spire 2. I played the first opus a lot. Now with the second one, I can play with my friends, and the strategic multiplayer decision-making is surprisingly good and entertaining!

Slay the Spire, the OG, was written using Unity, a game engine that was very appreciated by developers for its ease of use. "was." Because they messed with a currency that shouldn’t be messed with: trust.

They proposed a change in their pricing, making developers pay in proportion to the number of Unity runtime. It means, the more you are successful, the more you have to pay them. Lots of developers found it unfair, but in the end, they can do whatever they want with their product.

That’s why Slay the Spire 2 is using Godot engine. A free and open source engine that no one can take away from you.

"We have never made a public statement before. That is how badly you f***ed up."

Said the developers.

Take that as a cautionary tale: this can happen to any tool you use. A company will use retention and vendor locking technics to increase their profits. Designers dealing with the Adobe suite know what it is about. You have been warned.

Most LLM companies are selling at loss

ChatGPT introduced ads. Your $17 monthly Claude subscription doesn’t cover the token cost of your intense vibe-coding sessions. The first hit is free. They want to get you hooked but at one point they’ll come for your wallet.

They are already trying to reduce their spending. People report instabilities in these models, as well as inconsistencies. One day the model is smart, the other day it’s stupid. Because serving a quantized version is a bit less expensive, so maybe they are testing waters: what are you willing to accept?

I don’t want my workflow to change in the middle of the night, undocumented. Period. I don’t want to wake up and suddenly one of my API provider changed their contract interface without telling me beforehand.

That’s the problem with renting tools. Same goes with services like Spotify and Netflix. Your favorite album or show might suddenly disappear.

Except regression on these models can have a big impact on your ability to deliver value, and therefore make a living.

Renting for life

Well, don’t throw the baby out with the bath water. I’m not telling you to stop using LLMs for coding at once. What I’m saying is: take a step back. How dependent are you on these tools? If the price tomorrow is doing x10, is it still viable to use these tools? How easy would it be to move to a different tool? If your cloud hosted model has an outage, do you have a fallback solution or do you transform into a brick?

If your answer to the last question is a variation of the famous "It’s compiling" XKCD comic, then you might want to dig a bit into the world of open-source and open-weight models!

These things have tremendously improved. I’m sorry I can only rely partly on other people’s testimony. I don’t have (yet?) a data crushing cluster at home. But coding leaderboards (for what they are worth), and users report great results with, namely, Qwen3.5, Gemma4, GLM5.1, etc.

And even if you want to try a model or another, you still have the possibility to roll back to the previous version if there is a regression. And that point alone, I find wildly understated.

So, as a company, buying your own hardware, using open source models, fine-tuning them, having a controlled RAG and MCP, and in the end having a reliable workflow that you own is not a stupid as it might seem. It’s an investment and a wager for the future. It’s a philosophy and a mindset.

Do you want to be a hacker and a tinkerer, or a consumer? Both are fine. Just make sure you know which one you are and be ok with it. Be aware of the risks and tradeoffs.

So, let’s all surf on the bleeding edge of open source AI. I want to use whatever the Linux equivalent in that world is. Maybe I’ll be wrong and in the end the only viable tools will be proprietary models. "You’ll own nothing and be happy" as they say. But at least I would have tried.