AI Daily — May 7, 2026
2026-05-07
huggingface.co
vLLM V0 to V1: Correctness Before Corrections in RL
ServiceNow AI documents practical lessons from migrating RL training pipelines from vLLM V0 to V1, focusing on correctness issues that emerged during the transition. The post highlights subtle behavioral differences between versions that can silently corrupt RL reward signals, making validation critical before relying on V1 for production RL workloads.
openai.com
Uber uses OpenAI to help people earn smarter and book faster
Uber has integrated OpenAI models to power AI assistants and voice features across its driver and rider products, targeting a global real-time marketplace. The deployment spans both driver-side earnings optimization and rider-side booking acceleration, representing a large-scale production rollout of agentic and voice-driven AI interfaces.
rss.arxiv.org
Adapt or Forget: Provable Tradeoffs Between Adam and SGD in Nonstationary Optimization
This paper provides a rigorous theoretical separation between Adam and SGD under non-stationary stochastic objectives, decomposing finite-time bounds into initialization, objective drift, and hyperparameter-governed tracking error components tied to β1 and β2. The analysis characterizes burn-in times to reach Adam's irreducible tracking floor and proves high-probability stationarity guarantees under general L-smooth objectives, offering formal guidance on optimizer choice for continually shifting training distributions.