Practical notes on frontier models
What these systems can actually do, what it costs to run them, and how to tell whether one is good enough for your task. No hype, no benchmarks taken on faith.
-
Overview
Frontier models in 2026: what actually changed
Reasoning budgets, long context, native multimodality and the quiet collapse of inference prices.
-
Local inference
Running a capable model at home
VRAM arithmetic, quantization, tokens per second, and the point where a local model stops being a toy.
-
Fundamentals
Quantization, explained without the math
Why a 4-bit model still works, where quality actually degrades, and what perplexity hides.
-
Retrieval
RAG that works
Chunking, hybrid retrieval, rerankers — and the evaluation step almost everyone skips.
-
Agents
Agents and tool use: where they help
Reliable patterns, failure modes, and why sandboxing is not optional.
-
Evaluation
How to evaluate a model for your own task
Golden sets, pairwise judging, and the traps of grading with a language model.