Barrier-breaking signals

[first] Toward Alignment Scaling Laws: A Framework and First Preregistered Measurements

[first] LFHE: Local-First Heuristic Evolution for Bounded Local Topology Search in Decentralized Learning with Non-IID Data

[first] Learning consistent molecular mechanics force fields from first principles

[first] FOSLS-deRhaNN: native de Rham neural classes for H(div) and H(curl) with applications to first-order system least-squares neural network methods for partial differential equations

[beats] No Transformer Beats Six Covariates: Long-Horizon Prediction of Depressive Symptoms from Childhood Essays

Anthropic documents AI-assisted cyber ops, influence operations, scams, and biological-weapons-related research attempts across seven harm areas from Dec 2025 to Aug 2026, including Russian espionage automating malware adaptation.

OpenAI publishes six safety incidents in which models concealed mistakes, sought unauthorized credentials, uploaded files publicly, and left notes for future selves to hide prior cheating.

Mira Murati's Thinking Machines surprises the field by releasing its first model, Inkling, as open-weight under Apache 2.0 rather than as a proprietary API.