From the archive
Wednesday, 11 November
Folded from 25 sources that day.
Sorted by quality and by how many sources covered it, the same for everyone.
-
How Open Is Open? Iceberg, Catalogs, and User Choice
Argues that users need a definition of openness for table formats that they can verify without relying on vendors. Points to Iceberg's REST catalog protocol and calls for a production-ready baseline catalog.
-
AlphaGenome 2: Deciphering how cells read DNA to better understand disease
AlphaGenome 2 is a conditional DNA model that predicts how variants change gene expression in specific cell types. It was trained on over 100 million cell profiles and is available via API for non-commercial research.
-
Making the economic case for warehouse-first AI context
The post argues that storing agent context in the data warehouse instead of retrieving it dynamically cut token costs by more than 50%. It cites Deloitte data showing AI use cases take 2 to 4 years to pay back.
-
Token Pooling-Aware Training for Multi-Vector Embedding Models
Training the encoder to anticipate token pooling keeps retrieval quality under compression. With pool_factor=4, nDCG@10 retention rises from 94.6% to 97.8% at the same storage budget, with a Sentence Transformers example.