He spent 30–60 minutes trying to set up each task and treated a result as complete only if he would use it in practice.
Claude Code produced a functional replica of an interactive interface for Epoch's GATE economic model, but Ho found factual inaccuracies and missing components that made it unusable. Claude Opus 4.5 produced an analysis and article from supplied data that looked acceptable on the surface, but Ho judged its style and factual accuracy poor enough to rewrite. ChatGPT Atlas partly moved an article from Google Docs to Substack and Epoch's site, but struggled with footnotes.
Ho assigns a 5–10% chance that these tasks can be completed successfully on a first try by the end of 2026, with his 50% estimates ranging from late 2027 to early 2029. The article does not establish general automation timelines, model-wide capability levels, or labor-market effects: it reports one researcher's subjective assessment under a limited setup, including two feedback rounds for the writing task.