GraphCast three years later: how Google's weather AI 'breakthrough' was inflated by PR and media

震惊！谷歌AI天气预报屠榜！三年之后回头看，有多少是真的？

When DeepMind released GraphCast in November 2023, Chinese tech media ran headlines like 'crushed,' 'dominated,' and 'beat the world's best forecast system.' Three years later, ECMWF's public forecast accuracy curve shows zero jump in 2023—it still gains roughly 0.15 days per year. GraphCast's claimed 90% win rate across 1,380 metrics came from counting 6 variables × 37 pressure levels × multiple lead times repeatedly; the hard end metric of 'effective forecast days' didn't budge. None of the 15 Chinese articles quoted an independent meteorologist; in English media, only New Scientist interviewed one, who noted GraphCast doesn't do data assimilation—the most compute-heavy step—and simply feeds on pre-processed data from other systems. Later releases GenCast and WeatherNext 2 repeated the same playbook, with win rates climbing to 97% and 99%, but the baseline quietly shifted from ECMWF's operational system to the team's own previous model. The post suggests three checks when seeing 'crushed' headlines: find the end business metric, count how many quoted sources are truly independent, and verify whether the field has public long-term monitoring data.

Why it matters: A fact-check piece using public data to dismantle the narrative that GraphCast 'crushed' traditional weather forecasting in 2023. ECMWF's forecast accuracy curve shows zero jump that year; the claimed '90% win rate on 1,380 metrics' was inflated by counting 6 variables × 37 pr...