Safer Centaurs is where I develop open-source tools with AI assistance and share how I do it. The artifacts are small, composable pieces of software - following the Unix philosophy of tools that do one thing well. The blog posts show the process: what works, what doesn't, and the techniques I've learned for working effectively with AI.
This is a serious hobby project, like Audacity was. It runs parallel to my company Catalase Systems , where I do commercial work.
This safer-centaurs site started life as a polemic. I stridently talked of the dangers of AI, and of the opportunities for good. I pointed to some directions for safe AI research that democratises AI. I talked of how we must work to improve human-AI dyads, centaurs, if we were to have a hope of harnessing this new technology safely. My thesis was that we must adjust AI and also do some work on our own understanding, to work well with this new technology.
A major fault with the original Safer Centaurs site is that it was all aspirational. I knew some directions to move in, and had started doing so. I saw that AI could be transformative for education, and I'd started work on educational content and educational diagrams. But I hadn't got far. I knew that AI had to be democratised, so that it wasn't just for the power hungry. That meant small but competent AI that worked with people well. I hadn't got far with that either.
At that stage I had half-baked educational essays. I had half-baked diagrams. I had ideas for using small LLMs productively. Those aspirational fragments, it turns out, don't actually do much lifting.
However, I had over about two decades had a great deal to do with the code in Audacity audio software. I was one of the coders, and I coordinated releases of this software. I'd seen through Audacity how software given away for free can have an outsize effect. Our small team was getting around 2 million downloads of our software every month.
Audacity at that time hit a sweet spot - still just easy enough to use, yet also with enough real code behind it to do more ambitious audio work with it. Unlike my 'we shoulds' and my rough drafts of educational essays and diagrams, it wasn't an aspirational project. Audacity delivered code that people actually used.
I have reshaped the early centaur site into this version. The name still hints at the original motivation. Rather than exhorting others 'we must a find balance between human and AI skills' I am doing that myself. I now develop a lot of code to give away with AI assistance. I have conversations with AI to learn about science, maths and biology. I've got better at the prompting. I've got better at meta skills - the writing of tools for working with code - tools that I wouldn't have developed without AI help. In this site I share details of how I am doing this. I share techniques. I describe strengths and weaknesses of some current AI coding assistant, unexpected examples where the AI has helped a great deal. I also provide 'artifacts', pieces of code that come out of this process that are useful in themselves.
My artifact strategy is the well honed Unix one. Make small tools that do something well. Make the tools compatible. People can then compose the tools in the Unix way. This matches very well with AI capabilities. AI is much better at making small programs with a single purpose than large ones. The strategy of releasing small tools sits well with code co-created with AI.
How is it going? TL/DR; It's still early days but I think I am getting one balance right.
Back when I was doing Audacity, Audacity was volunteer work - a serious hobby about making something of real value to other people. The appreciation for it was immense. I particularly liked that it was used in schools - from language learning to journalism, and even for teaching some physics of sound. I was able to do sustained work on Audacity because I did Audacity in parallel with and in between paid gigs.
Safer Centaurs follows the same model. It is a hobby project that takes time, just like Audacity did. In parallel with Safer Centaurs I run a small company, Catalase Systems , for commercial work. If you visit that site you'll see the connections - I'm reusing the site templates, experimenting with similar techniques there, trying to make artifacts that work for both contexts.
In a way, Safer Centaurs has become the 'behind the scenes' for Catalase. I can't publish everything that is happening at Catalase, but I can show some of the more interesting parts and share anything that is made as open source. The hobby feeds the business with new techniques. The business keeps the hobby sustainable.
That balance, between giving away and getting paid, between experimentation and delivery - that balance seems crucial for this to work long-term.