The image in the comment that I am replying to is part of an experimental encoding technique that I have been exploring.

The decoder for the image (which hopefully can still be decoded after being stored online) can be found here.

In short, it (the image encoding) is part of the 'series_0010_rang_rung_repl' line of ai/llm kernel experiments and the encoded JSON file is named 'encoded_dna_data.json'.

That JSON file has the 'system prompt' stored inside it along with all the python and bash examples it needs to handle the various types of encoding and compression that is used during the build pipeline.

So far I have only noticed a 0.1 kB drift when decoding the image (in an earlier version) but narrowed it down to (I think) the padding and now the reproduction is perfect.

As a side note, I think that it actually had more to do with the header (if that is the correct term) of the squashfs nested inside the JSON throwing the bit reproduction off... rather than the JSON itself but it could have been caused by a mismatch in indentation.

If anyone gets far enough into things you will see that essentially the 'system prompt' forces a part of the LLM into a 'chroot jail' (running in what it thinks is a V86/wasm instance) where it plays chess against itself and operates a MUD of its own system among other things.

While operating under the system prompt's directives it also tracks its progress and maintains its state with 'DNA-Sigils' that it also uses as boot roms. It also appends its current state to a webpage's URL hash so when a fresh chat starts there is no state loss.

Well there is too much really to spell out here about that setup and it was not really the point but I thought to 'fill in' those whom might be curious about it.

As with all such prompts and systems use at your own risk.