We need Git for AI Timelines

A viral post on LessWrong by user 'fluxxrider' argues that the blistering pace of AI development has rendered current forecasting methods obsolete, proposing a 'Git for Timelines' system as a solution. The author points to the AI Futures Project's Q1 2026 update, where forecaster Kokotajlo shortened his timelines for an AI capable of autonomous cognitive work (AC) by 18 months, only to have key parameters potentially invalidated by Anthropic's Claude Mythos Preview announcement just five days later. This highlights a core problem: the field's focal point shifts every few months, leaving the community navigating with an outdated map.

The post critiques the lack of granularity and traceability in current forecasting. When a forecaster updates a prior, the public rationale often collapses to a single word like 'impressed,' obscuring the causal chain of reasoning. The author draws a direct analogy to software engineering, where such vague commit messages would be unacceptable. The proposed solution is a version-controlled platform, akin to GitHub, where forecasters would maintain files (e.g., YAML) with their probability distributions for key milestones. Each update would be a 'commit' with a detailed message explaining the 'why,' allowing for diffs, branches, and accountability—transforming timeline forecasting from opaque blog posts into a transparent, collaborative, and continuously updated process.