I’ve just realised something: I’ve been using AI (LLMs on someone else’s computer, I know, lol, it’s just what everyone calls it in normal speech) in entirely the wrong way.
Problem
| WRONG | RIGHT |
|---|---|
| AI2AI Use AI to build tools that themselves use AI | AI2TOOL Use AI to build tools that run on plain, programmatic algorithms |
Let’s see why the wrong path is actually wrong, and why the right one is right.
AI2AI
Imagine you build something using AI.
And that something then goes on to use AI itself.
See where it falls apart? It’s a vicious circle. That’s the Ouroboros: the snake eating its own tail, forever.
Every run of your tool is a new dice roll. Every answer costs a token, a request, a bit of trust. You never quite know what it will do, because you built a thing that doesn’t know either.
AI2TOOL
Now picture yourself building a genuinely useful tool using AI.
And that useful tool runs on algorithms that are baked in and deterministic, ready to go at any moment. Deterministic code doesn’t have good days. That’s a compliment.
No reliance on APIs. No LLM cranking up your RAM and spinning up your fans at night when your unsatisfied wife is asleep. Just code that does the same thing today, tomorrow, and on a plane with no Wi-Fi.
The AI did its job: it helped you write the thing. Then it got out of the way.
Conclusion
I’m not pretending to be someone really smart. In fact, I might even be a bit slow in the head, having worked this out only now after 3 months of working with it. But I’m 100% sure that simply knowing it makes my life a lot simpler and leaves me with far fewer things to worry about.
There’s a README in that repo. It was written by something that has never run the code, for someone who will never open the file. Not being sure of what we’re even talking about is the worst part. AI will write tons of documentations that you never really read. And that’s exactly why AI2AI is doomed to fail. Use AI right.