Harsha Johny

My 80s AI photo took about 3 seconds to generate. It also used somewhere between 0.01 and 0.29 kilowatt-hours of electricity, a range that exists because nobody, including the people building these models, has settled on one honest number. Since early September, ChatGPT’s viral 1980s portrait trend has been everywhere, big hair, denim, film grain, a fake 1985 date stamp. Millions of people generated one. Most, myself included, generated five or six before picking a favorite. I did the math on that afterw...

So I’ve spent the last 10 days going deep into how AI actually works, starting with building a neural network completely from scratch. No PyTorch, no TensorFlow, none of the usual shortcuts, just NumPy, a lot of math I half-remembered from college, and honestly, a lot of staring at the wrong output wondering what I broke. That was just the starting point though, from there it kept going. I wanted to write down what I actually learned, in plain language. Not the textbook version, the version I’d explain t...

I promised a deeper dive on backprop in my last post, so here it is. This is the part of building a neural network from scratch that genuinely broke my brain for a day or two. Every explanation I found either skipped the actual math, or threw so much notation at me that I lost the plot halfway through. So this is my attempt at explaining it the way I wish someone had explained it to me, building up the math slowly, with actual numbers where it helps. The one idea to hold onto through all of this: it’s th...