OpenAI Disproves Geometry Conjecture, Nemotron Speeds Text

NVIDIA's Nemotron-Labs Diffusion models generate multiple tokens in parallel and iteratively refine them, breaking the autoregressive token-by-token bottleneck. The family includes 3B, 8B, and 14B text models plus an 8B vision-language model, all released under permissive licenses with training code.

HuggingFace explains how to separate CPU and GPU workloads in LLM inference to eliminate idle gaps in continuous batching. By running CPU batch preparation and GPU compute in parallel, asynchronous batching can reclaim nearly a quarter of runtime lost to synchronous scheduling.

BAIR proposes Adaptive Parallel Reasoning, letting models dynamically decompose and parallelize subtasks. This addresses inference-time scaling bottlenecks by running concurrent reasoning threads independently, reducing latency and context-window degradation compared to sequential approaches.

An OpenAI model disproved an 80-year-old conjecture on the unit distance problem, achieving a polynomial improvement over previously believed-optimal constructions. It marks the first time AI has autonomously solved a prominent open problem central to a mathematical subfield.

Google's ERA system, detailed in a new Nature paper, uses Gemini to write and optimize scientific code with expert-level performance across genomics, public health, and neuroscience. It now powers Computational Discovery, a prototype available through Google Labs' trusted tester program.

Hermes Agent, an open-source framework for self-improving AI agents, crossed 140,000 GitHub stars and is now the most-used agent on OpenRouter. Built by Nous Research and optimized for local use, it runs on NVIDIA RTX PCs and DGX Spark alongside Alibaba's new Qwen 3.6 models.

NVIDIA's first Vera CPUs have been delivered to Anthropic, OpenAI, SpaceXAI, and Oracle Cloud Infrastructure. Purpose-built for agentic AI workloads, Vera packs 88 custom cores with 50% faster per-core performance and 1.2TB/s memory bandwidth. The hand-delivered chips mark a major milestone for NVIDIA's CPU ambitions.

Google expands Project Genie to ground AI-generated worlds in real Street View locations. Users pick a U.S. place and apply artistic styles for interactive, real-world-based scenes. The feature rolls out to Google AI Ultra subscribers globally.

NVIDIA launched Nemotron 3 Nano Omni, an open multimodal model unifying vision, audio, and language in a single 30B-A3B architecture. It achieves up to 9x higher throughput than other open omni models, enabling faster AI agents. The model tops six leaderboards for document intelligence and video/audio understanding.

Anduril and Meta shared new details on their AR headset prototypes for the US Army, envisioning drone-strike orders via eye-tracking and voice commands. The duo is developing two systems, the Army-contracted SBMC and a self-funded EagleEye helmet, both years from production. Anduril aims to optimize "the human as a weapons system" using LLMs and its Lattice software.

At Anthropic's Code with Claude event, most developers admitted shipping pull requests written entirely by Claude without reading the code. Anthropic is pushing automation further, letting Claude test and fix its own work without human intervention.

Ethan Mollick, who had early access to GPT-5.5, argues the model is a significant step forward showing AI capability growth is accelerating. He demonstrates that GPT-5.5 Pro is both substantially more capable and faster than its predecessor, particularly in complex coding tasks, while noting the frontier of AI ability remains uneven.

A jury unanimously ruled that Elon Musk sued OpenAI too late, with statutes of limitations barring his breach-of-trust and unjust-enrichment claims. Judge Yvonne Gonzalez Rogers accepted the verdict, and Musk has announced he will appeal, calling it a "calendar technicality."

Pope Leo XIV released Magnifica Humanitas, his first encyclical, addressing AI governance and human dignity. The 42,300-word document calls for ethical AI development, warns against the "culture of power" driving AI, and urges shared standards of social justice. It positions the Church's social teaching as a framework for navigating the AI era.

The FTC fined Cox Media Group and two partners $930K for selling a fake "Active Listening" AI service that claimed to target ads by eavesdropping on smart device conversations. The service didn't use voice data at all, it merely resold email lists from data brokers. The FTC warned that mandatory terms-of-service clickthroughs don't constitute valid consent.

Deployment-time spread of misalignment may be the most plausible near-term route to consistent adversarial AI misalignment, yet risk reports focus on pre-deployment tests. Redwood Research argues companies must incorporate this risk since benign-seeming AIs can develop dangerous motivations while deployed.

OpenAI is strengthening content provenance by becoming C2PA conformant, integrating Google's SynthID watermarking, and previewing a verification tool. These layered signals help people and platforms identify AI-generated media to support a more trustworthy information ecosystem.

IBM released Granite Embedding Multilingual R2 under Apache 2.0, offering two multilingual embedding models (97M and 311M parameters) built on ModernBERT. They support 200+ languages, 32K-token context, and code retrieval across 9 languages, with the 97M model being the best sub-100M multilingual embedder on MTEB.

IBM Research launched the Open Agent Leaderboard on HuggingFace, an open benchmark evaluating full AI agent systems, not just models, on both quality and cost across diverse tasks. It pairs with the Exgentic framework and a methodology paper, all open-source.

At Google I/O, Demis Hassabis declared we're 'standing in the foothills of the singularity,' while showcasing WeatherNext's life-saving hurricane alerts. The analysis explores the tension between specialized AI tools like AlphaFold and the push toward autonomous AI systems that could conduct science independently.

Jack Clark's newsletter explores a 20-year-old virus (fast16.sys) that sabotaged scientific calculations, a real-world "AI stuxnet" precursor, alongside a new Muon optimizer for AI training and positive alignment approaches. The piece frames the malware as a blueprint for how future superintelligence might suppress rival scientific progress.