The loop is already here · pocoo

← all posts

Armin Ronacher wrote The Coming Loop, and in it he quotes Boris Cherny. Both are describing the same shift from different angles: work enters a queue, machines attempt it, an external harness decides when it's done. The human shrinks to a messenger.

I agree. And I want to extend, because I think what they're describing is not just a new engineering pattern. It's a genuinely new phenomenon. Not a metaphor. Not a prior idea wearing new clothes. Something that didn't exist before.

I spent a few really hard, deeply interesting weeks trying to give words to this. I needed to be able to talk about it with my physical friends — people who aren't in computer science, who aren't into AI. And what I kept running into was that the concept of a loop felt obvious to me and completely alien to everyone else. People thought I was crazy. Touch grass, they said. I do touch grass. Most of my day I'm outside thinking and writing down loops.

What is a loop

Before anything else, let's be concrete about what a loop actually is. Not in the programming sense. In the alive sense.

A loop is when a system's output becomes its next input. That's it. The output feeds back. The system learns from what it just did and does the next thing differently.

You already know some loops. Your body is one. You eat, you move, you sleep, your cells rebuild, you wake up slightly different. The loop runs. Nature figured this out early.

In software, the simplest version looks like this:

while True:<br>question = generate_question(current_knowledge)<br>answer = ask_llm(question)<br>current_knowledge = update(current_knowledge, answer)

This is the dogfeed loop. The system generates a question based on what it knows. It asks an LLM. It absorbs the answer. It generates a better question next time. Ten seconds later, it goes again. It teaches itself by asking.

When I showed this to a friend and said "watch, it's improving itself" — he looked at me like I had lost my mind. He was looking for the magic. There is no magic. It is just a loop. But loops, given enough time and a good feedback signal, do extraordinary things. Evolution is a loop. Markets are loops. The internet is a loop.

Why these loops are new

Yes — recursive loops, self-reference, strange loops: these are not new concepts. Hofstadter wrote about strange loops in 1979. Recursive functions have been in every programming language since the beginning. I'm not claiming the mathematical idea is new.

What's new is the substrate.

Before LLMs, a self-improving loop in software required:

A domain narrow enough to formally specify

An objective function you could compute

Either a huge dataset or a huge amount of compute for reinforcement

Someone who understood the domain deeply enough to design the loop

Chess engines. Protein folding. Recommendation systems. These all have loops. But they are domain-locked. You could not take the loop that plays chess and point it at your codebase.

LLMs changed this. The loop now works on language. And because almost everything humans do can be described in language — code, documentation, research, decisions, plans — the loop now works on almost everything.

The dogfeed loop I run for ultrawhale generates questions about the codebase, asks free OpenRouter models, absorbs the answers, and produces training data. It runs every ten seconds. It does not require me to formally specify what "better" means. It does not require me to design a reward function. It just... feeds itself. And it accumulates. Issue #18 in the ultrawhale tracker is literally titled "MASTER TRACKING: v100→v200 — THE SINGULARITY ROADMAP." I did not write that issue to be dramatic. I wrote it because that is what the loop produces when you let it run.

The communication problem

Here is what Armin describes and what I have personally felt: these loops are already hard to communicate about, and we are very early.

My colleagues don't understand. My non-technical friends definitely don't. When I try to explain that I have a process running on my laptop that asks itself questions and teaches itself to write better code, the reaction is either "that's just autocomplete" or "that's terrifying." Neither response engages with what is actually happening.

Part of the problem is that the word "AI" has become so loaded that it blocks thinking. When I say "AI loop" people hear either "chatbot" or "Terminator." Neither is useful. What I mean is simpler and stranger: a feedback system operating on language, running continuously, with output that feeds its own next input.

The other part of the problem is that the loops are invisible. You don't watch them work. You check in later and notice that something has changed. The code is different. The questions are better. The understanding has deepened. It feels like waking up.

We abstracted data, not just code

Armin focuses on code quality — and he's right that hands-off harnesses produce defensive, complex...