Alibaba says Qwen has run recursive self-improvement for a month ahead of Qwen 4.0
teortaxesTex · x · 2026-10-11
At the Yunqi conference, Qwen lead Liu Dayiheng revealed Alibaba has already been running RSI (recursive self-improvement) continuously for over a month on Qwen 3.8 Max, and the entire Qwen 4.0 lineup (Max/Plus/Flash/27B) will adopt it deeply.
The pipeline: the model mines real logs for weaknesses, generates its own training data, validates via smaller models over multiple rounds, then feeds results into flagship training — with near-zero human intervention. Reported results: 33 effective iterations over a month, with Qwen3.8-Max's AA intelligence index climbing from 40 to 45 in 30 days.
Poster teortaxesTex remains skeptical, quipping "yeah… after a manner…", suggesting the marketing framing may overstate the actual gains.
More from Models
- Claude iOS app's unkillable pseudo-notifications draw fire from ex-OpenAI safety VP — Miles_Brundage · 2026-10-11
- Google's EmbeddingGemma 2 runs free and fully local on a Mac — Saboo_Shubham_ · 2026-10-11
- User says Google AI Mode still hallucinates after a year without seeing AI errors — burny_tech · 2026-10-11
- User burneda $17 Claude bill before a task even finished — lxfater · 2026-10-11
- Self-funded AI user: cost per task is the only benchmark that matters — victor_explore · 2026-10-11
- Mathematicians reviewing OpenAI's Navier-Stokes proof find the Lean formalization sound — burny_tech · 2026-10-11