Newsletter Archive

xAI's Ethan He, who built Grok Imagine in 3 months, argues that video models derive their intelligence primarily from LLMs rather than video data alone, and predicts the next breakthrough won't be better video generation but video agents capable of planning, editing, and iterating across creative tasks—following the pattern that shaped AI coding.

Latent Space · 67 min read Community

OpenAI breaks ground on a 1GW data center in Michigan as part of the Stargate initiative, substantially expanding compute capacity for training and inference workloads. The project aims to increase AI accessibility while creating local economic impact.

OpenAI · 1 min read Hardware

Alphabet is raising $80 billion through equity offerings, including a Berkshire Hathaway investment deal, to fund its AI infrastructure and compute buildout. This massive capital commitment underscores the scale of spending required to compete in frontier models.

Bloomberg Tech · 1 min read Market

Seven Chinese military-linked universities are pursuing Nvidia's H200 chips—the most advanced AI processors the US currently permits China to buy—signaling intense demand for cutting-edge computing power in defense applications and exposing gaps in export restrictions.

Bloomberg Tech · 1 min read Policy

Alphabet is raising up to $80 billion in equity (including $10B from Berkshire Hathaway) to fund AI infrastructure and compute scaling, joining hyperscalers projected to spend $750B+ this year on AI buildout. The move reflects unprecedented customer demand and the existential stakes of the AI race.

Axios · 2 min read Market

NVIDIA released JetPack 7.2 with agentic AI skills and brought NemoClaw (its agentic AI framework) to Jetson edge devices, enabling developers to run autonomous agents on robots, inspection systems, and industrial automation hardware. The update includes CUDA 13 support, performance gains on Jetson Orin, and multi-instance GPU support on Jetson Thor.

NVIDIA AI · 6 min read Tools

NVIDIA launched Cosmos 3 today—an omnimodal world model unifying language, image, video, audio, and action in 16B Nano and 64B Super variants with new SOTA open-weight image/video generation. Concurrently, Nemotron 3 Ultra (550B, A55B efficient) emerged as the strongest US open-weight LLM, while RTX Spark—a 1 petaflop personal supercomputer—was previewed with Microsoft as launch partner, signaling NVIDIA's aggressive push into open physical AI.

Latent Space · 11 min read Industry

Researchers red-teamed 11 LLMs across 690 clinical scenarios and found that aggregate accuracy hides catastrophic failures: top models (GPT-5, Claude Opus 4.1, X-BAI) scored 0.97+ overall but produced complete breakdowns on safety-critical tasks, with 10-20% error amplification on equity-related prompts when demographics changed.

arXiv NLP · 3 min (abstract) Research