I've used omp quite a bit before. This was designed to be intentionally minimalist without tons of built-in tools and LSP and all. I personally found the output to be much better without all that. Just base Pi works better for me than omp so this is using the same philosophy of Pi's minimalism but in very performant manner.
The minimalism of Pi is quite an advantage which is what KISS follows because this allows the harness to be used nicely outside of a coding domain too. One example would be in the finance domain for doing agentic orchestration without the harness being very much biased towards coding/software eng.
There's a print-friendly stylesheet, so you can easily print each article. You could probably automated this using the RSS feed.
But I'm not sure I'd want a printable book. The articles are each standalone so you'd need to figure out some logical sequence. Plus once you print it you can't make improvements to existing articles.
Around the time I turned 11 or so, my dad bought be a "how stuff works" book; definitely written to a "newspaper" reading level. It was paperback, not a lot of pictures, and generally held my interest. I think it targeted older teenagers or adults.
What surprised me was halfway through there was a section on how hash is made!
It's on the page, if you click the little info icon in the upper-right. Here's the text but there's some nice graphics there too:
Snake Game, training entirely in the browser. Built on tinygrad: the rollout / targets / train graphs are TinyJits authored in Python, then compiled once to WGSL and replayed here under WebGPU.
Observation: flat 10×10 board (100) + 4-dim prev-action one-hot = 104 dims. fc_pi.weight is zero-init so the opening policy is uniform over the legal actions; fc_v uses tinygrad's default Kaiming init.
Per rollout: T=24 × N=384 parallel snakes (9,216 transitions), then K=3 epochs × 4 mini-batches of PPO updates. GAE γ=0.99, λ=0.95; AdamW wd=0.01; ratio clip ε=0.1; grad-norm 0.5; Huber value β=1, val_coef=1; entropy bonus 0.008333333333333333.
Action mask + value clip + KL early stop. The 4-dim prev_a obs tail lets fc_pi zero the U-turn logit (the env silently overrides same-axis reversals anyway). Value loss is max(huber(v_new−td), huber(v_clip−td)) at ε=0.2. Approx-KL is sampled after each epoch and breaks the loop at 1.5·kl_target.
