What this is
Tamerix is where I keep the long-form essays that don't fit in a video. Each issue takes one machine-learning concept (the transformer, fine-tuning, soon embeddings and RL and diffusion) and walks through it with real math, real numbers, hand-drawn diagrams, and zero hand-waving.
Every formula on this site matches its source paper. Every default hyperparameter is a documented production value. Every code snippet is a real, runnable HuggingFace API. Every interactive panel runs on a deterministic toy model so you can refresh the page and get the same numbers.
If you spot something wrong, please tell me. The contact section below has every way to reach me.
Who writes this
I'm Monal Singh. By day I work on AI systems; outside of that I make videos on YouTube as AGI Tamer and write the essays you'll find here.
The brand name "Tamerix" is a small play on the channel name and the idea that modern AI is a wild thing worth understanding from first principles, not just consuming as a product. I draw each diagram by hand because doing so is how I actually understand it; I publish it because other people might benefit from the same drawing.
How to use the site
- The collection is the homepage. It lists every essay and every video, organized into rows by category.
- Each Issue is a single concept walked through end to end. They're long and meant to be read, not skimmed. Sticky table of contents on the left lets you jump between sections.
- The Videos section links out to YouTube. Most issues have a video companion; most videos have an issue companion.
- Every interactive panel (sliders, heatmaps, live training) is deterministic. Refresh and you'll see the same numbers. Move sliders and the math updates.
Get in touch
- YouTube: AGI Tamer
- LinkedIn: Monal Singh
- Mistakes / errata: if you find a wrong number, formula, or claim, send a note via LinkedIn. I'll fix it and credit you in the issue.
Colophon
Hand-built. System fonts only (Iowan Old Style for display, system sans for body, system mono for numbers). No frameworks, no trackers, no cookies, no CDN beyond Cloudflare's edge. Every page is one HTML file you can read.
Source: a private GitHub repo. Hosting: Cloudflare Pages. Domain: Cloudflare Registrar. Total annual cost: the price of one nice coffee. The point of doing it cheap is that the writing is the work; the infrastructure shouldn't be.