Hugging Face AI Updates: August 14, 2026

1. Amazon Details a Record-Train-Deploy Robot Loop with Strands Agents, LeRobot, and HF Storage Buckets

Hugging Face. A post on the Hugging Face blog from Amazon describes an end-to-end robot learning pipeline that records demonstrations, trains directly from the Hub, and deploys policies back to hardware using Strands Agents, LeRobot, and Hugging Face Storage Buckets. Demonstrations are captured in LeRobot’s standard format, with Parquet shards for state and action data and MP4 shards for video, while the buckets use Xet content-defined chunking for byte-level deduplication that reportedly cuts repeated uploads roughly fourfold. A stream_dataset() function lets GPUs train straight from the buckets without downloading full datasets, and a trained checkpoint deploys to a physical SO-101 robot by switching a single mode="real" parameter. Source

2. Hugging Face Reproduces 2,226 ICML Papers and Finds 23% with Contested Claims

Hugging Face. Hugging Face published results from a community hackathon that used AI agents to reproduce papers from ICML 2026, drawing 1,221 participants who published 6,816 reproduction logbooks covering 2,226 papers, about 34% of the conference, and evaluating 35,908 individual claims. Roughly 51% of papers had at least one claim independently verified and 266 were fully reproduced, while 23% had falsified or contested claims and 242 papers showed opposite verdicts on the same claim. The write-up documents proofs with flawed steps, mismatches between theory and implementation code, and evaluation methods that inflated reported performance by about threefold, and argues that human oversight remained essential to steer the agents and catch subtle flaws. Source