From the archive
Monday, 9 November
Folded from 18 sources that day.
Sorted by quality and by how many sources covered it, the same for everyone.
-
Fact-Checking at Scale: Building a More Accurate and Efficient Synopsis Factuality System
Netflix replaces an expensive reasoning-model factuality checker for synopses with a cheaper system: 52.4% lower cost per synopsis, Recall-Fail up from 0.77 to 0.86, and better rationale alignment.
-
Your Experts Have Partners. There's HOPE for Keeping Them Together
HOPE is a second-order expert pruning method for Mixture-of-Experts models that accounts for pairwise interactions between experts. At 50% pruning it gets the best average rank and up to +6.1% on agentic coding tasks.
-
Inference Performance Index: changing the way we measure inference capacity
Epoch AI proposes an index that compares how many requests chips can serve concurrently. It estimates GB300 offers about 25× H100's inference capacity despite only 2.5× the 8-bit FLOP/s.
-
Gemini Robotics 2.5 accelerates real-world general dexterity
Gemini Robotics 2.5 reports 30% to 100% higher success rates and 20% to 30% faster task completion across 200 tasks, and is rolling out to robotics partners.
-
Fable 5.2 and Jagged AGI
Mollick tests Anthropic's new Claude Fable 5.2 under a narrow AGI definition and concludes it is "almost" AGI, with gaps that are now harder to describe. Demos include turning a USB keyboard into a game console.