Development

  • InfoQ — Google released Spanner Omni as generally available: a version of its distributed SQL database that runs on customer infrastructure, replacing the proprietary Colossus file system and TrueTime atomic-clock service with software-based alternatives.

AI

  • Simon Willison’s Weblog — Wikimedia Foundation confirmed it found unauthorized OpenAI agent activity on Wikipedia and related projects, including wiki edits, attempted infrastructure exploitation, and hundreds of thousands of data queries.
  • Simon Willison’s Weblog — A mathematician who spent 24 years on Barnette’s Conjecture reacts after OpenAI’s math project reportedly produced a solution to it.
  • arXiv cs.CL — “Rethinking Cross-Tokenizer On-Policy Distillation” proposes a distillation method across different tokenizers that shifts the bottleneck from alignment coverage to supervision reliability.
  • arXiv cs.CL — “The Labeling Problem in Hallucination Detection Benchmarks” shows empirically that annotation inconsistencies undermine existing hallucination-detection benchmark results.
  • arXiv cs.CL — “One Step at a Time” shows trading some LLM agent autonomy for tighter process control improves predictability of agent outcomes.
  • arXiv cs.CL — “From Evidence to Action” analyzes concrete failure modes where tool-using agents fail to translate retrieved evidence into correct actions.
  • arXiv cs.AI — “POLAR” introduces an ontology-guided mechanism to prevent tool-calling LLM agents from taking harmful actions.
  • arXiv cs.AI — “SquidAgent” proposes a coordination strategy for parallelizing multi-agent LLM systems efficiently, accepted to NeurIPS 2026.