Lith

the oldest tool we know how to make

1931, Olduvai Gorge, Tanzania.

In a ravein cut through the Serengeti plain, researchers discovered stones that had been deliberately broken, intentionally struck with larger ones till sharp flakes came away. They call this era of tool making Oldowan after the gorge, and it has examples dating back to 2.6 million years ago.

A stone chopper is the simplest thing that is unmistakably made, and for most of the time there have been hominins at all, it was the most complicated thing anyone could do. Making one is harder than it looks: you have to read the stone, find a platform, and strike at an angle that takes a flake off rather than shattering the core. People who try it without being shown mostly fail.

These tools are some of the simplest things that can be unmistakably made, for most of the time we've been about it was the most complicated thing anyone could do. However making one takes some skill, finding the right stone, striking the right angle with enough force to flake off instead of shatter.

Knapping is difficult enough to learn, and useless enough to invent from scratch, that it had to be taught. A 2015 experiment by Morgan and colleagues found that transmitting Oldowan technique got measurably better as you moved from letting people watch, to gestural teaching, to teaching with words.[1] Their argument is that toolmaking and language leaned on each other as they grew. The first thing we made and the first thing we said may have been the same conversation: here, hold it like this.

what lith is

lith is a small desk object: an ESP32-S3, a 320x170 display, a scroll wheel, two key switches and a vibration motor in a printed shell. Out of the box it is a pomodoro timer. That is the shape it arrives in, not the shape it has to keep.

The name is from líthos, stone. You reconfigure it by knapping it: you describe what you want it to be, an agent called Oldowan writes the firmware, and it is flashed to the device from the browser. The tools people make get shared in the knappery.

The reason it is built this way is Bandura's: self-efficacy, the belief that you can actually do the thing, is built most strongly from mastery experience, and after that from watching someone like you succeed.[2] Which role model, and for whom, turns out to matter: a systematic review by Gladstone and Cimpian found that the effect depends on how similar the model seems and how attainable their success looks, and that a model whose success reads as out of reach can demotivate rather than encourage.[3] lith is arranged so a beginner gets all three. You watch it be made through short form video content. You make one, by building it from its constituent parts. Then you use the thing you made, on your desk, developing your own firmware by hand or with Oldowan's help.

[point on how overcoming the first hurdle towards that mastery experience improves continutity]

what is in here

website/

• site/: the four pages as they are served. index.html is the scroll-driven homepage, onboarding.html the walkthrough a new lith arrives with, knapp.html the knapping bench where you talk to Oldowan and flash the result from the browser, and knappery.html the shelf of builds other people have published.
• agent/: Oldowan. oldowan.py is the conversation and prompt layer, builder.py compiles and repairs the sketch it produces, knappery.py is the sharing side. providers.json selects the model; keys come from a secrets.json that is not in this repo.
• versions/: dated snapshots of every earlier version of the pages, back to before the renders existed.
• assets/: the WebP sequences the site actually serves.

research/

• ui-judge/: 19 firmware builds from 11 models across 5 providers, compiled, rendered on the simulator and judged. The headline is that the repair loop matters more than the model tier: first-pass compile rate was 26%, and the builder's own repair passes took it to 79%.
• provider-cost-latency/: wall-clock latency and list price across the providers Oldowan can run on, replayed through one real 3-turn conversation.

Both are written up with their figures in research/README.md.

references

1. Morgan, T. J. H., Uomini, N. T., Rendell, L. E., et al. (2015). Experimental evidence for the co-evolution of hominin tool-making teaching and language. Nature Communications, 6, 6029. https://doi.org/10.1038/ncomms7029

2. Bandura, A. (1997). Self-Efficacy: The Exercise of Control. W. H. Freeman.

3. Gladstone, J. R., & Cimpian, A. (2021). Which role models are effective for which students? A systematic review and four recommendations for maximizing the effectiveness of role models in STEM. International Journal of STEM Education, 8(1), 59. https://doi.org/10.1186/s40594-021-00315-x