OpenAI Deploys NVIDIA Vera Rubin, Gemini 3.7 Flash and Personalized Vaccine Reach Milestones

Key Takeaways

1. OpenAI has started using NVIDIA Vera Rubin racks for training, expanding compute for next-generation frontier AI.
2. NVIDIA's AVO agent achieved a perfect score on ARC-AGI-3, while NeMo Switchyard and Shopee's Compass case show major inference efficiency gains.
3. Gemini 3.7 Flash tops AA-AnalystAgent and delivers strong ARC-AGI performance at very low cost per task.
4. Hugging Face open-sourced a new mixed-transformer model and a 100k-hour egocentric dataset.
5. Moderna/Merck's personalized cancer vaccine succeeded in Phase III, a milestone for AI-driven medicine.

1. AI Infrastructure and Compute

1. OpenAI has deployed its first NVIDIA Vera Rubin racks and is already running training workloads, expanding compute capacity for pretraining next-generation frontier AI. Greg Brockman and Sam Altman both highlighted the milestone. — via 1 2
2. NVIDIA's AVO, a general-purpose programming agent, scored 100% on the ARC-AGI-3 interactive reasoning benchmark, solving all 183 levels across 25 environments without instructions, rules, or explicit goals. This demonstrates autonomous agents can handle complex reasoning tasks end-to-end. — via 1
3. NVIDIA introduced NeMo Switchyard, a tool to route agentic workflow steps to the best model based on quality, latency, and cost. Separately, Shopee's self-developed Compass model on NVIDIA's AI factory stack grew monthly token volume from 3 billion to 340 billion (113x) in eight months, while fraud detection became 50x faster and processing costs dropped 90%. — via 1 2
4. swyx discussed NVIDIA's $6 billion acquisition of poolside's model factory, which is reportedly producing highly competitive models. This acquisition, if confirmed, could strengthen NVIDIA's position in AI model production. — via 1

2. Models and Developer Tools

1. Gemini 3.7 Flash ranks first on the AA-AnalystAgent data analysis benchmark, achieving the highest accuracy across 80 real tasks while being 60–90% faster than top models and 2.4x faster than the closest competitor. It also scores 84.6% on ARC-AGI-2 ($0.25 per task) and 95.5% on ARC-AGI-1 ($0.12 per task), underscoring its cost-efficiency. — via 1 2
2. Runway released Runway Ruby, a model that converts SDR video to up to 16-bit HDR, supporting ProRes and EXR sequences, and compatible with any uploaded or generated video up to 30 seconds. This expands professional video production capabilities. — via 1
3. Hugging Face open-sourced SenseNova-U1.5-8B-MoT, a mixed Transformer that uses different weights for text and image tokens and denoises in pixel space, with benchmarks matching Nano Banana 2. — via 1
4. Hugging Face and LightwheelAI released EgoSuite-Open100K, a 100,000-hour egocentric dataset with over 15,000 tasks and scenarios, including hand/body pose and wrist-camera views. The first 10,000 hours are available now, with commercial training permitted, covering 7 environment categories and 128 scene types in LeRobot format. — via 1 2
5. Perplexity Agent API now offers access to 41 frontier models from 9 providers via a single endpoint, with built-in tools for web search, financial search, scraping, and sandboxed code execution. This simplifies building multi-model agent workflows. — via 1

3. AI in Science and Medicine

1. Moderna/Merck's personalized cancer vaccine succeeded in a Phase III trial of 1,137 patients, with each dose individually customized to the patient's tumor; ML-based neoantigen prediction proved effective in a blinded trial. This is the first positive Phase III individualized neoantigen therapy and mRNA cancer treatment. Yann LeCun also highlighted that therapeutic mRNA vaccines are showing efficacy across cancers, with 2026 seen as the year they move from science fiction to reality. — via 1 2

4. Key Signals and Insights

1. Kai-Fu Lee announced his new book "AI NATIVE," arguing that 95% of companies still fail at true AI transformation. The ebook/audiobook launches September 15, with hardcover on November 3. — via 1
2. ChatGPT added an Apple Messages plugin that lets users search messages, view conversations, and draft/send replies on Mac, making ChatGPT more integrated into desktop workflows. — via 1
3. In swyx's "Kill My SaaS" challenge, 68 participants completed; his team used Claude to redesign a scheduling tool, replacing a poor interface with a Trello-like experience and releasing production-grade scheduling 20x faster. This case illustrates how AI-driven development can accelerate product iteration. — via 1