I demonstrate how Codex is reversible computing.
Janus is a reversible programming language, i.e. it supports deterministic forward and backward computation by local inversion.
A language like Janus for reversible non- deterministic inference is sorely needed.
Codex is non-deterministic forward and backward inference.
From the example
json above, the example
jq below was generated, which was able to
generate the following json, an exact replica
of the original.
Code was generated from output.
In another sense, the input was generated from the output, and the program was the prompt.
A new language is needed
A language which enforces types is needed. What is the type of the output? JSON. But what does that mean?
Imagine a situation where generated output was chimeric and had no known compiler which could agree on its type. What is actually needed is consensus, somehow.
Therefore, we need language detectors which we agree on. I’m talking about LMs on the blockchain.
type would then be a prompt to that LM,
which validates the type by binary
classification. I’ll call this an
imaginary type, but the LM (the source of truth)
should be agreed upon, so it should be on a
This is why I have invested in
Demo of above
If this article appears incomplete, it may be intentional. Try prompting for a continuation.