bergen, norwayvol. i · no. 22 · July 21, 2026rss feed

Hasan Arief

A lab notebook on agentic coding, open-weight models, and what they cost to run

Section

Benchmarks

Collected benchmark results from published sources: never my own runs; every number links to where it came from.

Note

Why agent distillation needs separate success, constraint, and cost metrics

A reading of workflow distillation and why task success, constraint adherence, trajectory quality, and deployment cost belong in separate reported metrics.

· 3 min read

Note

Karpathy's autoresearch, and its loop for parameter optimization

Karpathy's autoresearch runs an AI agent in a keep-or-discard loop. I examined its techniques, then applied the loop to parameter search where a grid stalls.

· 4 min read

Guide

Open-weight LLMs for coding in 2026: hardware and real costs

Open-weight coding LLMs as of July 2026, GLM-5.2, Kimi K2.7, DeepSeek-V4, with the VRAM math, GPU rental prices, and when self-hosting beats an API.

· 6 min read

Note

Fable 5 versus Opus 4.8 with /goal: pricing and persistence compared

Fable 5 returned at double the price of Opus 4.8, with days-long autonomy as the headline capability. A comparison with Claude Code's /goal loop, sourced.

· 3 min read