Armin's piece resonated. I've been running these loops for months — ultrawhale is a self-improving coding agent that dogfeeds itself on a 10s loop. (I'm `using` it daily, standalone mode is very buggy, but as a `subagent`, `gateway`, it's the orchestrator of my loops)
My extension: the loop has already escalated past "write loops that prompt Claude." Now loops write loops. The job shifts to holding the genesis contract — being clear about what you want the loop to become, because you won't be tracking the recursive depth yourself. LLMs do that part.
Also: this post was dictated on my phone, sent to Claude, synthesized into the article. The post is the example.
- peter
What I learned is that you "spent a few really hard, deeply interesting weeks trying to give words to this", and yet you still, still, still could not bring yourself to write those words down yourself without asking an LLM for help.
>A loop is when a system's output becomes its next input. That's it
Not in programming speak.
In it, a loop is just a repeated execution construct. No need for feedback, and no need for any accumulator either.
What that describes is a feedback loop, and, if some extra restrictions apply, a closed loop.
>The system learns from what it just did and does the next thing differently.
That's not necessary either, even for a feedback loop. The system could just not "learn anything", do the same exact thing on each pass, and only the thing it acts upon changes (e.g. a microphone feedback loop doesn't learn or do something different each pass. It just does the same thing, but on top of the signal of the previous pass).
>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.
Aside from the obvious AI slop writing, nothing in the example he showed guarantees that this is "improving itself". Each iteration of the answer, as given back by the LLM can get increasingly worse, especially as the context window grows, or as something that shouldn't be there (e.g. a hallucination) in the previous answers becomes part of the later answers' focus, or the "roll of the dice" that's an LLM answer gets bad, and influences the future answers, and in other ways.
All 3 general examples are also problematic: evolution is not about improvement, but mere fit, and even given that, an organism can change via evolution into a no -longer-viable organism (fit for the previous evolutionary pressures, and disastrous once the environment changed). Markets of course crash or get worse (e.g. monopolistic, ineffective) all the time. And the internet is the perfect example of something going to shit over time.
>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.
Wrong again. Many old-style feedback loops where generic enough to be applied across domains. ML learning techniques like linear regression or genetic algorithms could be used in chess engines but also protein folding and elsewhere.
>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.
Non sequitur. Previous loops worked on numbers, which are even more generic than natural language.
>When you can automate the production of your own training data, you have closed a loop that was previously open. And closed loops, unlike open ones, compound.
Nothing about a closed loop necessitates compounding, and nothing about an open loop prevents it.
>We are not just building coding loops. We are building living data machines. The dogfeed loop doesn't just produce code. It produces a dataset
This is not just slop. It's mega-slop.
>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.
Armin's piece resonated. I've been running these loops for months — ultrawhale is a self-improving coding agent that dogfeeds itself on a 10s loop. (I'm `using` it daily, standalone mode is very buggy, but as a `subagent`, `gateway`, it's the orchestrator of my loops)
My extension: the loop has already escalated past "write loops that prompt Claude." Now loops write loops. The job shifts to holding the genesis contract — being clear about what you want the loop to become, because you won't be tracking the recursive depth yourself. LLMs do that part.
Also: this post was dictated on my phone, sent to Claude, synthesized into the article. The post is the example. - peter
Good to know I won't be reading it then, thanks for the heads up
summarize it with a loop and get it's interesting score, cheers adam :)
What I learned is that you "spent a few really hard, deeply interesting weeks trying to give words to this", and yet you still, still, still could not bring yourself to write those words down yourself without asking an LLM for help.
>A loop is when a system's output becomes its next input. That's it
Not in programming speak.
In it, a loop is just a repeated execution construct. No need for feedback, and no need for any accumulator either.
What that describes is a feedback loop, and, if some extra restrictions apply, a closed loop.
>The system learns from what it just did and does the next thing differently.
That's not necessary either, even for a feedback loop. The system could just not "learn anything", do the same exact thing on each pass, and only the thing it acts upon changes (e.g. a microphone feedback loop doesn't learn or do something different each pass. It just does the same thing, but on top of the signal of the previous pass).
>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.
Aside from the obvious AI slop writing, nothing in the example he showed guarantees that this is "improving itself". Each iteration of the answer, as given back by the LLM can get increasingly worse, especially as the context window grows, or as something that shouldn't be there (e.g. a hallucination) in the previous answers becomes part of the later answers' focus, or the "roll of the dice" that's an LLM answer gets bad, and influences the future answers, and in other ways.
All 3 general examples are also problematic: evolution is not about improvement, but mere fit, and even given that, an organism can change via evolution into a no -longer-viable organism (fit for the previous evolutionary pressures, and disastrous once the environment changed). Markets of course crash or get worse (e.g. monopolistic, ineffective) all the time. And the internet is the perfect example of something going to shit over time.
>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.
Wrong again. Many old-style feedback loops where generic enough to be applied across domains. ML learning techniques like linear regression or genetic algorithms could be used in chess engines but also protein folding and elsewhere.
>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.
Non sequitur. Previous loops worked on numbers, which are even more generic than natural language.
>When you can automate the production of your own training data, you have closed a loop that was previously open. And closed loops, unlike open ones, compound.
Nothing about a closed loop necessitates compounding, and nothing about an open loop prevents it.
>We are not just building coding loops. We are building living data machines. The dogfeed loop doesn't just produce code. It produces a dataset
This is not just slop. It's mega-slop.
>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.
OK, that's enough blog reading for today...
[dead]