AI is not a magic wand and not a threat. It’s a tool. And like any tool, you only see what it’s worth when you start building with it. We used it to build two things that run on IPFS: Abundomy and 1CoinH. This is what that resulted in – and where things go wrong if you just let AI do its thing.
What we use AI for
At Brainfusion two projects are running on the decentralized web: Abundomy and 1CoinH. No rental server from an American giant, no platform that will change the rules tomorrow. The content is up IPFS, spread over nodes that we manage ourselves.
That sounds nice. It’s also work. Configure nodes, publish CIDs, set DNSLink records correctly, keep an eye on whether everything is still running. That’s exactly where AI proves its worth — in three ways.
1. Write code
IPFS, IPNS, libp2p and OrbitDB are not technologies for which you simply take a course. The documentation is fragmented, examples become outdated quickly. An AI assistant who knows the context of your project will write a working outline in minutes that would otherwise take you an evening with half Stack Overflow answers. Not because the AI is smarter — but because it doesn’t get tired of doing the research.
2. Check whether it really works
This is the part that people underestimate. Writing code is half the battle; proving he does what you think is the other half. AI is excellent at figuring out which tests you would skip: what if the gateway is slow? What if the DNS record has not yet been propagated? What if the node restarts during a publish?
We don’t let a change go live without post-verification — a curl check, a status code, an actual page loading. AI not only proposes these checks, it also carries them out and reads the outcome.
3. Health monitoring programs
The most boring work in management is also the most important: know that something is broken before a visitor notices. Disk space, memory, certificates expiring, backups silently stopping, a IPFS node eating up too much RAM.
You normally never build that kind of monitoring — it doesn’t yield anything visible, so it stays there. With AI you can write such a script in half an hour, including a daily reporting email. For us, this now runs every morning at 7:00 am. No email? Then something is wrong. That’s the whole system, and it works.
And why IPFS?
Just without jargon. The regular web works on address: you request a file from a server. If the server is turned off, your file will be gone. IPFS works on content: you request a file based on what it contains. Whoever has it can give it.
- Not one point of failure. If one node is not running, another one supplies. Content does not disappear because one machine fails.
- Content is verifiable. The address (the CID) is a fingerprint of the content itself. If one byte changes, the address changes. Unnoticed tinkering with published content is not possible.
- No platform that pulls the plug. No account will be suspended, no algorithm will push you away, no conditions will be different next year.
- It scales with interest. Popular content is located on more nodes, and therefore becomes faster — exactly the opposite than with a regular server, which collapses under pressure.
- It’s yours. Own node, own domain, own rules. We host everything ourselves. That’s not a principle for principle’s sake: it’s the difference between owning and renting.
Abundomy and 1CoinH run on that basis. No intermediary stands between you and your own data.
But then those weird AI videos
Now the other part of the story. Because you’ve probably seen them before: AI videos that promise something moving or spectacular and that fall apart at the first look. Hands with too many fingers. A face that changes shape halfway through a sentence. Audience clapping with arms that do not belong to a body. Lips that say something different than you hear.
Below is such a clip. The idea behind it is beautiful — an emotional performance, a message about what you leave behind when you’re gone. Look at it for a moment and then pay attention to the details again.
Isn’t it loading? Then the gateway is busy for a while or your browser blocks the embed —
then open it directly.
If you have your own IPFS node running, you can retrieve it using only the CID:
bafybeib6mwio6najbvtgbqscvyuctepgxeekblwuvfi3ky3g3lbyb6eg4qWhat this clip shows
The point isn’t that the video is bad. The point is that hesimultaneously impressive and unfinished is. Five years ago this was unimaginable. Today someone makes it in an afternoon. And that is precisely where the pitfall lies: it is good enough to convince on a cursory glance, and not good enough to be correct.
That’s the same pitfall as with code. AI produces something that looks convincing, neatly formatted, with neat comment lines. But you won’t know if it’s correct until you check it. A generated function that looks good and silently returns the wrong value is more dangerous than one that immediately crashes — just as a six-finger AI video is less harmful than one you don’t notice.
And then the real danger: a tool that works
You see six fingers. You see a crash. Those are the friendly mistakes — they report themselves.
The real risk lies in what does work. Someone without a technical background has AI build a tool: a customer portal, a booking form, a membership administration. It comes out, it looks professional, it does what it’s supposed to do. Users come in, fill in their details, everything is up and running. Nothing happened for a year.
Until it turns out that this tool has been passing on data all this time to parties that no one ever chose.
That doesn’t happen because the AI wants evil. It happens because an AI model draws on what is most common on the internet — and those are the standard building blocks of the big tech companies. Ask for a form and you’ll get one with a font loaded from an external server. Ask for statistics and you will receive a tracking script that each visitor passes on. Ask for storage and you get a database in a country with very different privacy laws than ours. None of those choices are an error in the code. They are choices — only no one made them.
And that’s exactly the point. Each of those parts is free, and free almost always means your users’ data is the payment. Anyone who can see which visitor was on which page when, knows more about your customers than you do. That knowledge is a commodity. There are plenty of parties that pay good money for it.
The person controlling the AI did not choose that. He didn’t know that there was a choice. There was no vision — and then the internet average fills the void. AI has no idea about where your customers’ data should end up. It has an average. And the average of the internet is big tech.
This is why we host everything ourselves. Not out of stubbornness, but because “it works” and “it’s safe” are two different things. A tool that works fine today could turn out to be the problem a year from now — and then it’s your customers whose data is floating around somewhere, and your name attached to it.
The difference between a six-finger AI video and working software is not the model.
It is the person who checks afterwards whether it is correct – and who knew in advance what he wanted.
How do you use AI?
- Provide context, not command. “Write a script” provides generic garbage. “Write a script that on this server, with this IPFS version, checks these five values every morning and emails them to this address” yields something useful.
- Always have it verified. Don’t just ask for code, ask for proof that it works. A command with an outcome that you can read yourself.
- Put AI on the work you’re skipping. Monitoring, documentation, testing, backup checks. Not the fun work — the work that gets left behind.
- Always ask where the data goes. For each part that represents AI: who can see this? Is it on my own server or somewhere else? Is something being loaded from an address I don’t know? This is the question that AI never asks on its own.
- Stay the last check. AI gets to write it. You decide whether it goes live. Always. Without exception.
- Look twice. In videos to the hands. In code to the edge cases. In both: what happens when things go wrong.
Finally
AI is not the reason Abundomy and 1CoinH run on IPFS. The reason is that we do not want to be dependent on parties that will apply different rules tomorrow. AI is the reason why feasible was for a small team: the code got there faster, the checks were actually built instead of postponed, and the monitoring now runs every day.
So tools. Excellent tools. But they are still tools — and the carpenter is you.
Want to set up something self-hosted yourself? We build and manage our own infrastructure — IPFS nodes, mail servers, password vaults, payments in crypto. Everything in our own hands, nothing with big tech.
Please contact us.