The code listing shows hand-authored weights that Astra wrote directly into Qwen for a Pong task, reportedly reaching a mean score of 517.4, versus PPO baselines of 15.8 after 100M steps and 25.3 after 500M.
After tiring of YouTube's exaggerated thumbnails and repetitive recommendations, Karpathy had Opus 5.5 build HackerTube, a YouTube-style site drawing on 6,454 videos recommended across 40,000 Hacker News comments from 2007–2026, assembled in roughly two hours.
HackerBooks is a Show HN project that surfaces a daily-reshuffled selection from 16,543 books recommended on Hacker News, ranked by all-time recommendation score. The page lists titles with authors and counts of HNers recommending each.
The post argues Astra's weight-reading/writing claims are misleading: only small 17K–172K-weight experiments read raw weights alone, while the headline Qwen math improvement, raising MATH-500 from 7.9% to 48.1%, uses Python and up to 50 practice batches.