1. Nvidia Commits $26 Billion to Build Its Own AI Models SEC filings reviewed by Wired show Nvidia plans to spend $26 billion developing open-weight AI models.

2. Hacker News Bans AI Comments While Companies Let Bots Run Job Interviews Hacker News updated its site guidelines last week with a blunt addition: "Don't post generated/AI-edited comments. HN is for conversation between humans.

3. OpenAI Ships a Sandboxed OS for AI Agents A developer building an AI agent that installs Python packages, writes scripts, and reads output files can now do all of that inside OpenAI's infrastructure.

In Brief

• Anthropic Launches the Anthropic Institute Anthropic announced a new organization called the Anthropic Institute. The company has not yet disclosed the institute's full mandate, funding, or leadership details.

• Rakuten Cuts Mean Time to Resolution by 50% with OpenAI's Codex Rakuten deployed OpenAI's Codex coding agent across its software engineering workflow. The company reports halved bug-fix times, automated CI/CD review processes, and full-stack builds delivered in weeks instead of months.

• Wayfair Automates Product Catalog and Support Triage with OpenAI Models Wayfair integrated OpenAI models to classify support tickets and enrich millions of product attributes automatically. The system handles customer-facing triage and backend catalog accuracy at scale.

• Google Deploys AI to Screen for Heart Disease in Remote Australia Google launched an AI program targeting cardiovascular health in remote Australian communities. The initiative focuses on populations with limited access to specialist cardiac care.

• Study Finds Reasoning Mode Unlocks Factual Knowledge LLMs Otherwise Cannot Recall Chain-of-thought reasoning helps LLMs answer simple factual questions they get wrong without reasoning enabled. The effect applies to single-hop queries requiring no multi-step logic, suggesting reasoning activates latent parametric knowledge.

• MM-Zero Trains Vision-Language Models from Scratch Without Any Seed Data Researchers introduced MM-Zero, a multi-model system that bootstraps vision-language model training without seed images or paired data. The approach removes the data dependency that typically gates VLM self-improvement.

• New Benchmark Exposes Consistency Failures in Long-Form LLM Story Generation Researchers released ConStory, a benchmark measuring how often LLMs contradict their own facts, character traits, and world rules across long narratives. Prior benchmarks measured plot quality and fluency but left internal consistency untested.

• CoCo Replaces Natural-Language Planning with Code for Text-to-Image Generation A new framework called CoCo uses executable code as chain-of-thought reasoning during image generation. Code-based planning handles spatial layouts, structured elements, and dense text more precisely than abstract language prompts.

• LoGeR Scales Dense 3D Reconstruction to Minutes-Long Video Without Post-Optimization Researchers presented LoGeR, an architecture that reconstructs 3D scenes from long video sequences by processing chunks with hybrid memory. The design bypasses quadratic attention costs that block existing models from handling extended footage.

Read the full edition →