Training an LLM from scratch, and why retrieval beats a better model in coding agents

21 September 2026#

• LLM training. Found an LLM training tutorial from scratch with handwritten algorithms. Exploring this hands-on.

24 September 2026#

• Agent efficiency. Microsoft’s fastcontext project reports that in GPT-5.4 trajectories, reading and searching accounts for 56.2% of all tool use turns and 46.5% of main agent total tokens. Considering this, the largest efficiency gain in a coding agent is not a better model or better diff format, it is a smarter retrieval.