Resources

Research, technical writing, and engineering references.

Papers

Publication version

35 min

After the universe

Models of meaning should be built after the universe they are modeling.

Three moves: empirical convergence across LLM embeddings, quantum computation, and neural population geometry; containment over analogy; a single Lorentzian substrate viewed through four geometric tiers, with falsification criteria.

Read the publication version

Books

First edition

14.3 hr · 172k words

Working with Large Language Models

How they actually work, and how to ship with them.

A practitioner's book on what large language models actually are, what they actually do, and how to ship with them. Twenty-five chapters across five parts: the reality, the interface, agency and augmentation, the crucible of production, and bare metal. First-party benchmarks in the appendices, no hype.

Read the book

In development

Substrate prototype

In development

Engineering work on the four-tier architecture proposed in the paper. First milestone replicates the Nickel and Kiela hyperbolic WordNet result.

Engineering notes

All notes

Quantization Explained: VRAM, Quality, and the RAM Performance Cliff

What quantization does to model quality and speed, how Q4/Q5/Q8 compare on real hardware, and why running a model split between GPU memory and system RAM can be 5-30x slower. Third-party benchmarks, practical recommendations.

Benchmarks8 min read

Qwen 3.5 Benchmarks: 0.8B to 14B on RDNA4 Vulkan

Qwen 3.5 benchmarks across 80 prompts on AMD RDNA4 16GB with Vulkan. Unconstrained (2048-token) results show qwen3:14b leading at 0.594, with 9B pulling ahead of 4B when not artificially truncated. Includes analysis of how token budget constraints distort benchmark rankings.

Benchmarks10 min read

LLM Benchmark Comparison: 12-Source Rankings

Cross-referenced LLM benchmark data from 12+ sources. Consensus rankings for coding, math, reasoning, tool calling. Interactive charts, cost vs. quality scatter plots, and full methodology.

Benchmarks12 min read

How We Price: Aggregate Component Costs by Scrape

We publish the exact component costs, eBay search queries, retail sources, and markup formula for every Foundry GPU server. Prices are scraped from live market data and fed into an auto-pricing engine. This is the full breakdown.

Architecture6 min read

ROCm HIP Idle GPU Bug on RDNA4: Vulkan Workaround

When loading Qwen 3/3.5 models via Ollama using ROCm/HIP on AMD RDNA4 GPUs, the GPU becomes permanently stuck at 100% utilization. We trace the root cause to HSA runtime teardown and benchmark the Vulkan backend as a drop-in fix with comparable performance.

Tools8 min read