Save the OSS Commons with a Format
If you accept the premise of my last post, OSS has moved to a Dark Forest, then open source software is highly vulnerable and possibly past saving with the advent of very capable AI models like Mythos. You may not be happy with where that leaves us — the answer to the problem of shared code is: don't share code.
Don't share code is simple except for the fact that we all benefit from the sharing of core pieces of software. So what do we do? The history of our industry is full of stages of innovation. A defining feature of every step is moving up the ladder of abstraction. Today is no different except that the step feels a little bigger than usual.
So, what is the step: this "move up the ladder of abstraction"? Well, I think it is an idea that is not particularly novel for those who have been coding fully with AI since the beginning of the year: specs and evals.
That is it. Here is the move. Share the specs and evals. Generate ALL of your code. Don't share code!
I submit the Commons Format to you as a candidate move.
It has the basics of what I think are needed:
-
A self describing format that is true to itself and even allows for bootstrapping its own tooling.
-
A language agnostic format. No constraints on implementation.
-
A mergeable format. As with code libraries and modules, shared specs need to be composable so we can define what we build on.
-
Evals as a first class member. Specs without evals are less than half the battle. Evals carry more of the contract than the prose does, but unstructured text is the easier entry point for humans (and LLMs?).
-
References. Just as code refers to libraries, a Commons Format module can refer to other modules via canonical URLs (e.g. GitHub links).
So, simple. Yet often simple implies a great deal of change. A new world of token economics, of security by shrinking and fragmenting the attack surface, of codegen everywhere. It seems to me this is a large change that isn't obvious at first. But once you see it, it seems hard to unsee.