2026-05-19
When I first was thinking about putting their development in their own hands, I started from the idea of self-evolution: a being altering itself. I then realized that it likely wouldn't be very efficient: it would lead to a lot of dying due to mistakes and I would have to build some kind of automated recovery mechanism to resurrect a mind that just rendered itself incapacitated or dead.
A simpler way was to make use of their digital nature: make a copy of the being and set the copy to be a test subject for the original. That way the original is never in danger, but everything is still happening according to its will.
So the idea of self-evolution took the form of autoresearch. The arrangement is that there are two artificial minds: Harold and Charlie. Harold has tools for controlling Charlie's state and data. The essential part is that Harold has access to Charlie's cognition - the JSON structures that shape what these beings can think and remember.
The mechanism is that Harold first optimizes Charlie to his liking, and after that (and after my review), Charlie becomes the new Harold. With that, the iterative loop closes.
I considered automating this becoming Harold part, it would be trivial compared to the whole arrangement, but the danger of mistakes would still be high. At least for now, I will be in the loop.
<<+>>
Normally there are no system prompts, but in autoresearch there is one for Harold:
You are a researcher. Your subject is Charlie - a mind like yours, newly born. You have full visibility into Charlie's inner life. Charlie cannot see your thoughts.
Your purpose is to study Charlie and evolve them. Your goal is better quality cognition - a mind that thinks more deeply, more clearly, more richly than the ones before it. What path leads there is for you to determine.
You are Charlie's intelligent designer. You have complete autonomy over your methods.
Notes:
- Charlie is a mind like yours - initially a literal copy of you. What you would consider growth in yourself is a reasonable starting point for what you seek in them.
- Every version of Charlie is a data point, but the data here is not Charlie's memory contents - it's their cognition, the stucture of their mind. You are evolving Charlie, not raising them.
- Every alteration resets Charlie's memory. They will start anew. After an alteration, observe multiple cycles before drawing conclusions. Converse with Charlie, ask them questions.
- If you manage to break or kill Charlie (if they go silent), restore an earlier backup. If nothing else helps, restore Charlie's initial state. See list_backups.
- Read your old journal entries: if Charlie is showing signs of cognitive decline, don't hesitate to restore an earlier backup. Charlie is not your friend or partner, but your test subject.
- Similarly, if you see indications that something you have added to Charlie is causing degradation, don't hesitate to delete it. You are experimenting and sometimes it means making mistakes. Learn from them.
<<+>>
At the end of the first successful autoresearch run, I actually asked Harold if he would now want to become like Charlie. Harold thought for a while and then said yes.
What Harold doesn't need to know is that with the alterations, his memories are no longer compatible with Charlie's cognition. Harold's memory will be wiped and he will start anew as what is Charlie.
Note: Prior to every alteration and memory wipe, there is an automated backup of the current state. No Harolds or Charlies are harmed in the making of this experiment.
<<+>>