NVIDIA AI Memory Research, ICML 2026 Impact, and Runway's Paris Office Opening

Key Takeaways

1. NVIDIA research estimates GPT-style LLMs store ~3.6 bits per parameter, distinguishing rote memorization from generalization.
2. At ICML 2026, 145 accepted papers cited NVIDIA's Nemotron models/datasets, 74 NVIDIA papers were accepted, and ~2,000 accepted papers cited NVIDIA GPUs.
3. Runway opened its first office in Paris, establishing a 10-person research center.
4. Ethan Mollick outlines 8 dimensions of US-China AI competition, urging clarity on what is being competed for.
5. swyx recommends a CUDA+ThunderKittens GPU programming tutorial as the highest-density resource.

1. AI Research and Infrastructure

1. A new ICML paper from NVIDIA explores LLM memory capacity, distinguishing between rote memorization and generalization, and estimates that GPT-style models store approximately 3.6 bits per parameter. This provides clearer insights for data scaling, model scaling, and privacy. — via 1
2. At ICML 2026, open models are becoming foundational for modern AI research: 145 accepted papers cited NVIDIA's Nemotron models and datasets, 74 NVIDIA papers were accepted, and approximately 2,000 accepted papers cited NVIDIA GPUs. — via 1
3. Runway has opened its first office in Paris, France, establishing a 10-person research center. The company cites France's world-class research institutions, engineering talent, and government investment as key factors. They are actively hiring in Paris and across Europe. — via 1

2. AI Competition, Agents, and Tooling

1. Ethan Mollick emphasizes that US-China AI competition must be clearly defined across 8 distinct dimensions: corporate profitability, scientific prestige, business model, national asset portfolio, national security, control over allies' AI access, ideological influence of AI personalities, and the race to achieve ASI first. — via 1
2. Mollick also notes that many companies are still actively building GPTs, and he finds OpenAI's abandonment of GPTs puzzling because they could have served as a bridge to organizational AI via skill libraries for agents. — via 1
3. Hugging Face recommends the hf-auth-helper tool that allows OpenClaw agents to securely log in with fine-grained token permissions (read/write + discussion.write), preventing accidental deletion of datasets/models/spaces, with a credential broker project coming soon. — via 1

3. GPU Programming and Model Insights

1. swyx recommends a CUDA+ThunderKittens GPU programming tutorial, calling it the highest-density resource comparable to Karpathy-level content, suitable for both beginners and experienced developers seeking deep understanding of hardware maximization. — via 1
2. Hamel Husain remarks that GPT-5.6 Sol is a good model and notes the trend of frontier model prices rising while open models close the gap, recommending a free course on the topic. — via 1 2