Important Trends in AI: How Did We Get Here, What Can We Do Now, and Where are We Headed? - Jeff Dean Princeton University CS Distinguished Colloquium: Jeff Dean, February 10, 2026 What makes it special is that the one and only Jeff Dean walks us through important milestones in the evolution of ML/AI and offers his perspective on where AI is headed in the near future.
std::hive is a new sequential container introduced in C++26.
std::have automatically manages its storage in multiple memory blocks and it can reuse the memory locations of erased elements.
Internally, std::hive is a linked list of blocks. Each block carries a skipfield: a small integer per slot that tells the iterator how many erased slots to jump over.
From performance point of view, std::vector is still the most efficient data structure. std::hive performs better than std::list, but is not as efficient as std::vector.
Cloudflare’s Big Pineapple platform stores over 250 billion DNS cache entries at any given time.
At that scale, even 1 extra byte per entry costs more than 250 GB of memory!
Big Pineapple is implemented in Rust.
The post explains how analysis of the usage patterns, careful choice of the data structures e.g. Box<[T]> vs Vec and through other application optimizations, they were able to reduce the aggregate working-set memory across the fleet by about 100 TB.
Archil • aims to be the universal storage engine that sits on top of Amazon S3 • allows you to have infinite storage in your own S3 bucket, but still be treated like a local disk with POSIX compatible filesystem • introduces SSD cluster to speed things up • clients interact with the SSD cluster and the cluster intelligently stores data to your S3 bucket
Archil’s solution addresses the common issues with handling lots of small files over network file system
Archil also works with other cloud storage providers e.g. Cloudflare R2
Typical object store performance characteristics • high latency • high parallelism • dynamically scaled based on usage • high sufficient throughput with sufficient parallelism
Lots of proprietary and open source products are being built on top of object stores • Analytical databases e.g. Snowflake, Databricks • Virtual Disks e.g. DIS: blockDevice over Immutable Storage • Embedded databases e.g. SlateDB • Vector databases e.g. turbopuffer • Kafka compatible data store e.g. WarpStream • solution for mobile observability e.g. bitdrift