I'm not a techie. I use Hermes Agent with various underlying models to run personal experiments as a part of the learning process. You will find them below.
I also use the above setup to simplify complicated (for me..) LLM engineering topics into ebooks I can actually read and understand. Rather than leave them on disk to collect dust, I publish them here.
A detailed account of fine-tuning HRM-Text-1B — a 1B-parameter hierarchical reasoning model — via QLoRA on an 8GB Mac Mini. 13 iterations, a silent gradient bug, catastrophic forgetting, a v6 breakthrough, and the honest conclusion about the ceiling of local fine-tuning on low-end hardware.
A rigorous benchmark of five small models across 1,350 scored runs. Gemma 4 E2B leads overall at 82.2%, but the top three are clustered within 3.3 points — the real finding is that architectures are diverging, not converging. Which model you should use depends on which failure mode you can tolerate.
A 42-prompt evaluation of ZAYA1-8B, Nemotron 30B-A3B, and Gemma 4 26B-A4B separating raw observations from architectural implications — revealing a proof-of-concept whose broken governor masks an intriguing signal about where inference is heading.
A journey through the hidden physics of AI hardware — starting with a single AND gate and building up to why GPUs, TPUs, FPGAs, and the human brain each look the way they do. Based on the Dwarkesh Podcast conversation with Reiner Pope.
A journey through the memory wall, KV caches, batch economics, speculative decoding, and custom silicon — explaining why everything in AI slows down before it gets faster.
The sequel exploring what comes after inference economics — architectures and techniques that reshape how models think.
A critical analysis of NVIDIA's Nemotron-Labs-Diffusion — what the three-mode diffusion model means for hardware, where it breaks, and what to build now.
A narrative exploration of biologically plausible alternatives to backpropagation — from equilibrium propagation to predictive coding — and why closing the gap between AI and the brain unlocks fundamentally more efficient hardware.
Practical techniques for shrinking giant AI brains: compression, mixture of experts, pruning, merging, and recovery training.
This section is also optimized for LLM and AI agent consumption. See llms.txt for machine-readable structured content.