This project is an experiment in large-scale engineering with language model (LM) agents.

It serves two purposes.

• The first is as an experiment, testing the hypothesis that agents can be coerced to write high-performance software. Coercion here roughly means with disciplined context management through tools, including programming languages and systems.
• The second is as a source of high-performance programmable software, for usage more generally (beyond the purview of seeing if agents can write this stuff: we want it to be useful! To laypeople, to scientists, etc)

In any case, it should be abundantly clear ahead of time that agents are being used to write this software. This project does not seek to hide that, it is the entire point of the project.

We dream of a world where users can quickly spin up customized high-performance software for their open access computing devices, without having to reinvent the universe (indeed, this project might serve as such a universe, or as source for your own universe).

If you find our work useful, you might cite the project from the citation key at https://a-tiny-project.github.io/

Agents are now capable enough to allow anyone to rapidly spin up new software. This is awesome! Software deserves to be democratized! Assume that anyone can write any piece of software, what is the value of the software produced by a project such as this?

Answering this question is analogous, in many ways, to the existential crisis confronting mathematics. In general, anyone can spin up a piece of software ... but not everyone can do it well.

(whether you want “well” depends drastically on what you’re doing, of course!)

At least for today's agents, what you get out remains a function of what you put in. Here, we (and presumably also you) want someone who is motivated to carefully understand what they are doing at the helm ... a.k.a someone who is in it for the love of the game (the game in this case is some combination of PL and systems, etc).

We love the game, and want to understand it! We’re also interested in the idea that we might teach agents “the game” (can you teach agents good PL / systems design?)

At the end of the day, we enjoy fast programmable software, and wish to see more of it in the world. This endeavor seems worthy of our tokens.

All work, unless otherwise specified, is licensed under AGPL v3.0 (which, as is abundantly clear, is mostly symbolic these days -- but we hope it warms the hearts of a few)

If the idea of this project proves valuable, we fully expect people to reverse engineer or clean room reconstruct things they want. We've done the exact same here! There are countless projects that we've drawn upon (and we collect and propagate their licensing, and maintain references and commentary on influences in the READMEs throughout our systems).

We can't really stop people going wild with agents. That's fine: use this project if you feel like you can trust the contributors here to steer the agentic traces in good directions.

We like to have a bit of fun with our software: we hope you find a bit of playful chaos in some of our releases. In the words of The Little Schemer, bon appetit!