From the archive
Monday, 26 October
Folded from 13 sources that day.
Sorted by quality and by how many sources covered it, the same for everyone.
-
Making Failure Diagnosis Reliable with Pensive AI
Netflix's Pensive AI adds AI-assisted error extraction to its rule-based failure diagnosis for data and ML workloads. AI extraction now covers 100% of failures, including ones the legacy system left unknown.
-
Multi-Ratio Video Resizing for Mobile Feed and Marketing
Netflix describes an automated pipeline that uses shot detection, active speaker detection and face saliency to produce 9:16, 4:5 and 1:1 crops from 16:9 footage in a single pass.
-
Working with academic communities to build scientific understanding
OpenAI shares model results on the Kakeya conjecture, the Exponential Time Hypothesis and quantum chromodynamics, and explains how researchers can verify and build on them.
-
Import AI 478: Math for alignment; ProgramBench; and Mirendil shows what RSI means
Resolution used AI agents to autoformalize Hironaka's resolution of singularities, producing about 600k lines of Lean in roughly a week for an estimated $135,000 in token spend. The company argues that cheap formalization could speed safety research.
-
ChatGPT's expanding world of work
OpenAI Signals data covers 165 work activities across nearly 100 countries, showing how people use ChatGPT for work and how that usage changes over time.