AIs can now often do massive easy-to-verify SWE tasks and I've updated towards shorter timelines

A prominent AI forecaster has published a significant update, moving towards "substantially shorter AI timelines" based on recent breakthroughs. The core update is a near doubling of the estimated probability for full AI R&D automation by the end of 2028, from around 15% to just under 30%. Furthermore, the analysis predicts that by the end of 2026, AI systems will achieve 50% reliability on completing massive, easy-to-verify software engineering (ES) tasks—projects that would take human engineers years or even decades. This performance is specifically noted for tasks that don't require novel ideation (ESNI tasks), relying instead on existing knowledge and rigorous verification.

The forecast revision is driven by several concrete demonstrations. The performance of models like Anthropic's Claude Opus 4.5 and 4.6, and OpenAI's Codex 5.2 series, has consistently exceeded expectations on benchmarks. More compellingly, the researcher cites specific, non-contaminated projects where AI, with only "moderately sophisticated scaffolding," autonomously completed work of a scale previously thought impossible in the short term. Examples include an almost entirely AI-written C compiler and significant cybersecurity results. These demos suggest the current "50% reliability time-horizon" for such massive ESNI tasks, using public models, is already between one month and several years when cost is equated to human labor. The researcher also anticipates a major training compute scale-up in 2026, which is expected to accelerate progress further.