https://peertube.watch/videos/watch/f3fb69ca-5820-4aad-b55d-531b52dc25ca Peace Call. Mr. Peace is an AI-made character; voice is AI-generated; script written with AI assistance. No affiliate links in this video. How this site is run: https://getpeacecall.com/verify/?utm_source=peertube&utm_medium=video&utm_campaign=underdog-dory Important context: About 62% of the 8.39-point average gap comes from one AA-Omniscience non-hallucination metric (79.38 vs 17.39). Excluding that metric, the other 11 scores average approximately 37.07 for Dory and 33.56 for Qwen under the paper's equal-weight method. These are author-reported benchmark results, not a measure of overall model superiority or equal inference cost. In a preprint by NVIDIA researchers (Fishman, Sun et al.), Dory1.0-4B - a 4-billion-parameter "looped" model that can re-run its middle block to spend more compute per answer - averaged 40.60 over 12 benchmarks at 4 loops, vs 32.21 for Qwen3.5-9B (Table 2, equal-weight average). At 1 loop it averaged 35.17. On SWE-Bench Verified it went from 50.3% (1 loop) to 61.8% (4 loops), vs 61.0% for Qwen3.5-9B. The catch: Qwen3.5-9B still won on hard science and knowledge questions (GPQA Diamond 80.93 vs 72.16; HLE 15.43 vs 10.43). The average includes a non-hallucination score where Dory was far ahead (79.38 vs 17.39). Qwen3.6-35B-A3B still had the top average (42.30). Scores were measured by the authors (NVIDIA), and more loops means more compute per answer. Source: Fishman, Sun, Acharya, Ladhak, Chmiel, Akle Serano, Banner, Ginsburg, Puvvada, 'Dory1.0-4B: A Looped Normalized-GPT', arXiv:2611.03998v1 - PREPRINT (Nov 2, 2026), not yet peer reviewed. https://arxiv.org/abs/2611.03998 . Numbers from Table 2. Names are trademarks of their respective owners; Peace Call is not affiliated with or endorsed by them. Charts drawn by code from the source's numbers. Paper: https://arxiv.org/abs/2611.03998 (version v1, posted Nov 2, 2026). AI Disclosure: Mr. Peace is a fictional AI character using a synthetic voice. This video is AI-assisted and does not depict a real person speaking. Research Disclosure: Based on a research preprint that has not been peer-reviewed. Results apply to the reported benchmarks and experimental conditions, not overall model superiority. Trademark Disclosure: Names are trademarks of their respective owners. Peace Call is not affiliated with or endorsed by them. Sat, 07 Nov 2026 15:49:39 GMT https://validator.w3.org/feed/docs/rss2.html PeerTube - https://peertube.watch https://peertube.watch/lazy-static/avatars/3b284627-b75f-409e-ac2e-10f4c2946f43.png https://peertube.watch/videos/watch/f3fb69ca-5820-4aad-b55d-531b52dc25ca All rights reserved, unless otherwise specified in the terms specified at https://peertube.watch/about and potential licenses granted by each content's rightholder.