Cathedral was built by hackers
I think it’s important to make the following distinction: the cathedral exists because hackers or nerds decided to monetize the majority of the non-technical population. The cathedral is built by people who understand technology and can monetize it. The issue becomes that with wealth the ivory walls get built higher and higher.
There are societal issues we have today. Some cathedrals are impossible to topple, as toppling them would mean tech stops working globally. Take the AWS outage that happened a few weeks back.
Open source is catching up
We have all experienced proprietary software. Sam Altman’s ChatGPT is a fantastic example of a job well done. There are industries where toppling a cathedral is harder, but not impossible. Take Mistral AI, the first open-source project that beat GPT at its own game. Mistral is open sourced and is amazing! Open AI’s recent products have been lackluster. Mistral’s on the other hand, I think are groundbreaking, like their voxstral models.
I think in today’s complex world there are a couple of truths that will not change. macOS and Windows will persist as they are convenient, but beneath the bloat there exists a different compute revolution. We have more of it per user. As LLMs evolve and become more complex and need more compute to train, we saw a gradual tapering off in quality. But on the flip side, we see open-source LLMs rapidly catching up. We can run LLMs locally with the help of Jan AI, Ollama, and others. In some cases they are better than current black-box solutions.
Writing creative code
As this continues to evolve, my bet is on the open-source software world. There is something special in how this community writes code. AI code is not creative. AI code is very much basic. The idea is that 10 good developers working together will beat 1,000 incompetent ones using LLMs. The language, the architecture, the platform matters so much more now than ever before. AI is not incentivized to help a person code. It’s optimized to consume maximum tokens to bill you. Let’s take, for example, web scraping. AI will write a Selenium script if you ask it to scrape a website. It is going to do it in Python.
Now this doesn’t work. You go back to AI, ask it again, maybe now it gives you Playwright or Puppeteer. How many times will you go back and forth before it decides that actually it needs to use Rust or WASM. Now it refactored your code, changed language, and by now you’ve probably consumed a million tokens. Welcome to vibe coding. You probably will have an intermediate step where it goes from python to js to rust to wasm or something similar.
An experienced engineer with web scraping will make that leap faster. There are plenty of libraries that use Rust instead of Python. For example, spoofing user headers in Rust is easy and cheap so your requests look like real browsers, which reduces blocks and retries and raises effective throughput. Combined with Rust’s tiny per-request overhead and tokio concurrency, you typically see multiple times the usable throughput compared with higher-overhead stacks because fewer retries and lower latency mean more completed pages per second.
Next generation of cathedrals
I think open source will exist and persist. There are too many things in motion for it not to be. With growing distrust of blackbox solutions and more information being available to people there is a huge privacy push happening that wants to prioritize the individual rather than monetize them.
How does open source monetize itself? Well, most commonly, they have their own cloud for which they charge money. They give you the ability to self-host, but that’s hard in some cases. In most cases it is way easier to just use their cloud.
Here are some of my favorite open-source initiatives:
System76, Deno, LanceDB, Zed, uv, Nous Research, Jan AI, Agno, DSPy, OpenBB, Qdrant, and Polars.
I think that open source is good for the soul when you program. In a way the plan is to always have a part of Desync open source. Why? Because we want to make it easy for people to use us. At the same time we want to get better at coding, and there is no better way than having some random developer on the internet brutalize your codebase 100 different ways.
In the end I learned to code because builders shared their work in the open. It was Andrej Karpathy with his YouTube tutorials gave me the push to pick up a keyboard and start learning by doing. His blog (see his site) and his channel (see his YouTube) made complicated workflows feel accessible and real.