From the archive
Wednesday, 14 October
Folded from 15 sources that day.
Sorted by quality and by how many sources covered it, the same for everyone.
-
Integrating Netflix's Foundation Model into Personalization applications
Netflix compares three ways to use its central personalization foundation model in production (embeddings, subgraph and fine-tuning), with monthly pretraining, daily refreshes and embedding stabilization.
-
Measuring the Impact of Personalized Recommendations
Netflix presents a behavioural model that separates members' underlying tastes from the incremental effect of recommendations, validated with small A/B tests that nudge title salience.
-
The Self-Organizing Data Stack
Argues that agents make generating pipelines cheap while refactoring stays risky, and proposes treating data assets as stable logical data products with the orchestrator maintaining the dependency graph.
-
Google DeepMind and Singapore Government collaborate to build fine-tuned AI model for scam detection
A Gemma model fine-tuned on Singaporean scam conversations runs on-device, warning users before they share money or credentials, without server-side scanning of messages.
-
AI systems are improving rapidly at Slay the Spire and its sequel
Epoch's new SpireClimb benchmark finds frontier models now beat Slay the Spire at base difficulty at least 90% of the time. At maximum difficulty they reach win rates of about 17-20%, still far below top humans.