OpenAI catalogs ten math and theoretical CS advances [Xtra!]A concrete showcase of models contributing to formal research, not just benchmarks.
Are reasoning models right for the wrong reasons? [Xtra!]Chain-of-thought outputs may look like reasoning without tracking the actual computation.
Training on the best of K guesses [Xtra!]A practical technique for squeezing better signal out of noisy model outputs.
ByteDance ships Seedance 2.5 [Xtra!]Another entrant pushing controllable, reference-driven video generation.
Speculative KV replication for bursty LLM traffic [Xtra!]Addresses a real production pain point: inference cost spikes under bursty load.
Kimi K3 on AMD's MI355X beats B300 on perf-per-dollar [Xtra!]Concrete evidence AMD accelerators can undercut Nvidia on inference cost.