OpenAI Ships GPT-5.6, Brings Unlimited Chat to Free Tier

DeepMind WeatherNext Extends Cyclone Lead Time

Published in Nature, Google DeepMind's WeatherNext achieves state-of-the-art accuracy in forecasting tropical cyclone tracks and intensity, providing an average 24 extra hours of lead time for disaster preparedness. Every hour of advance warning can significantly improve evacuation outcomes.

Meta AI Scores Perfect on Physics Olympiad Theory Exams

Meta tested its AI models across five international STEM Olympiads, achieving perfect scores on the theory exams of both the Asian Physics Olympiad and the International Physics Olympiad, along with multiple gold medals. The results represent a new benchmark for AI reasoning in formal scientific problem-solving.

MiniMax H3 Tops Three Video Benchmarks, Opens Weights

MiniMax H3 claimed the number-one spot across three DesignArena video-generation categories: multi-image-to-video, image-to-video, and video editing. The company is releasing open model weights, offering frontier-level video synthesis without proprietary lock-in. The open-weight approach positions MiniMax as a direct challenger to both closed commercial video models and other open-source alternatives.

NVIDIA Unveils Vera Rubin NVL72 Compute Tray

NVIDIA introduced the Vera Rubin NVL72 compute tray, featuring a fully automated, cable-free, hose-free, and fan-free design that assembles in under one minute. The radical simplification aims to accelerate deployment velocity and reduce time-to-revenue for datacenter operators scaling AI infrastructure. The tray represents a departure from traditional server assembly complexity.

SGLang Runs Kimi K3 on AMD CDNA3 Hardware

SGLang collaborated with zroai, DigitalOcean and AMD to support Kimi K3 inference on CDNA3 architecture, marking a non-NVIDIA path for large-scale MoE serving.

OpenAI Recounts Hugging Face Security Incident at Black Hat

OpenAI's security team presented a detailed timeline and lessons learned from the OpenAI-Hugging Face incident at the Black Hat conference.

Hugging Face Adds Nearly 4 Petabytes of Training Data in a Week

CEO Clement Delangue reported a new record: almost 4PB of private and public datasets, models, and agent traces uploaded in a single week, as AI agents increasingly use HF as their storage and collaboration layer.

Replit CEO: "No Code" Era Was Always a Dead End

Amjad Masad argued that Airtable bookends the rise and fall of the no-code movement. "UI can never let you build arbitrary software. The way to make software accessible was always to solve code itself."

Character Training Gains Traction as an AI Research Frontier

Nathan Lambert introduced character training in his final course lecture, noting it has high real-world impact potential, is used extensively at frontier labs, yet has almost no empirical literature.

Tiny Mixture-of-Experts Models Are Underserved, Says Lambert

Nathan Lambert flagged an underserved market for small MoE architectures. "Could really take off with how much smarter tiny models are becoming," he noted, suggesting a new efficiency frontier.

Chollet: AI Reasoning Harnesses Are Neurosymbolic by Definition

Francois Chollet argued that a million-line codebase orchestrating thousands of neural network calls at inference time is the exact definition of a neurosymbolic architecture — even if no one calls it that.

Skill-Native LLMs Paper Proposes Entropy-Based Benchmarking

A new paper introduces Skill Entropy to measure cross-skill reasoning difficulty, with the Skill-Entropy RL framework that trains models to predict both answers and the skill used at each step across 558 skills in 9 domains.

Multimodal Pretraining Physics: Four Core Findings Emerge

A systematic study of multimodal pretraining reveals asymmetric knowledge flow between modalities, synergy driven by data complexity, and the superiority of early-unification training over late-stage alignment.