A KAIST team uses optimal transport over video-foundation-model embeddings to learn robot rewards from a handful of preference labels, winning a top-0.7% ICML 2026 oral slot.
Julia 1.13 hit its first release candidate on April 29, adding mutually recursive types and moving default constructor generation out of the language's Scheme bootstrap layer.
CVE-2026-35273, a CVSS 9.8 missing-authentication flaw in Oracle PeopleSoft PeopleTools, was exploited as a zero-day by UNC6240 before Oracle shipped mitigations and CISA set a June 15 federal deadline.
Snap is preparing a 2026 consumer launch of its standalone Specs AR glasses, having formed a dedicated subsidiary and signed a multi-year Qualcomm Snapdragon agreement.
A joint directive from China's industry ministry and state-asset regulator pushes humanoid robots from demos into paying jobs, targeting 10,000 commercial units and 100-plus applications by year-end.
In May 2026, solar supplied 12.8% of US electricity versus coal's 12.2% — the first month solar has outgenerated coal, according to Ember.
Announced at Build 2026, Coreutils for Windows packages a Rust uutils build of coreutils, findutils, and grep as a single native binary installable via WinGet, though the GitHub repo still labels it a preview.
A Penn State-led study in Physical Review Letters argues that the highest-energy cosmic rays, including the Amaterasu particle, may be nuclei heavier than iron that lose energy slowly enough to reach Earth.
Five-year overall survival was non-inferior when surgeons omitted completion axillary dissection, while severe arm problems fell from 12.6% to 3.6%.
Paramount Skydance launched an in-house games division on June 5, folding Skydance Interactive and Skydance New Media into one studio and unveiling a TMNT: The Last Ronin game at Summer Game Fest.
SIG-etcd released v3.4.45 alongside v3.5.31 and v3.6.12, marking the end of support for the v3.4 series and pushing remaining users to upgrade.
Moonshot AI released Kimi K2.7-Code, an open-weight 1-trillion-parameter MoE coding model under a Modified MIT License, claiming roughly 30% lower thinking-token usage than K2.6 on self-run benchmarks.