Source note: Local copy of a Summer Into AI 2026 submission originally published on Substack. Screenshots, video, and audio found in the post body are mirrored locally when publicly accessible; profile and avatar images are intentionally not copied. Read the canonical post.
The whole field this week has been building AI tools to query federal datasets — ask a chatbot a question, it reads the data, answers. Useful. But I wanted to do the opposite, and the harder thing: not query the data — learn from it.
GLASS BOX is a neural network I wrote from scratch in one file — the forward pass, the backpropagation, the optimizer, all by hand, no TensorFlow, no PyTorch, no libraries at all — that teaches itself, live in your browser, on real numbers from the U.S. Department of Education: what graduates of ~865 American colleges actually earn. Press train and watch it learn to predict earnings from a school’s cost, size, selectivity, and type.
And here’s the punk part twice over: not only did the government collect this and never build you a model on it — the models that do exist (theirs, and the AI labs’) are black boxes. GLASS BOX is the opposite. You can see every neuron and every weight as it thinks:
the glowing web of neurons in the middle, its connections visibly rewiring as it learns
pulses of light streaming forward through the wires — that’s data flowing through the net
the cloud of test schools on the right snapping onto the diagonal as its guesses get right
the error curve dropping live
It reaches about R² 0.70 on held-out schools it never trained on — a genuinely decent prediction from a handful of features — and I don’t ask you to trust that: there’s a built-in gradient check (the numerical and analytic gradients agree to ~1 part in 10 billion), so you can prove the math isn’t faked.
Then ask it yourself. Slide the cost, the SAT, the school type, and it forecasts the median earnings — and whether that’s “strong value” or “pricey for the payoff.” A model you watched train, answering the one question every 18-year-old is actually asking: is this degree worth it?
On the theme. Go get that ‘murica data. The government has an extraordinary record of what higher education actually returns — and buried it in a spreadsheet nobody opens. The punkest thing I could do with it wasn’t a dashboard or a chatbot. It was to build a mind that learns it, from nothing, in the open, with the glass walls up. Thanks Eric for the Data.Punk theme — go get that ‘murica data was the perfect prompt for this.
A nod to a competitor. Jake Strait has been the most prolific builder this week — AI tools that read federal contracts, EPA data, FDA records and brief you on them. Great stuff, and the perfect foil: Jake built tools to query the data; GLASS BOX trains a model on it, from scratch, and shows you every neuron. Two very different answers to “go get the data.”
Open the glass box (it all runs in your browser — the model trains on your machine, nothing sent anywhere):
▶️ Live: https://myles-mf.github.io/summer-into-ai-2026/week-3/glassbox/
💾 Source (one file, read the backprop): https://github.com/myles-mf/summer-into-ai-2026/blob/main/week-3/glassbox/index.html
🗂️ Everything I’ve built for the contest: https://myles-mf.github.io/summer-into-ai-2026/
Also written up on my site: https://merlinforge.com/blog/glass-box-week-3/
Punk isn’t dead. It’s compiling. ★
Original source
Canonical Substack URL: https://merlinforge.substack.com/p/glass-box-everyone-asked-an-ai-about.