# Nov 14, 2024

- **Francois Chollet**, creator of **Keras**, is leaving Google, marking a significant transition in his career while continuing to support Keras's development and roadmap.

- **Watermark Anything** enables the embedding of **multiple localized watermarks** into images, enhancing digital content protection and ownership verification.

- **Gwern Branwen**, an anonymous researcher, was among the first to predict the **scaling of large language models (LLMs)**, emphasizing that intelligence is fundamentally a search over Turing machines, which has profound implications for AI development and understanding.

- **BERTs, specifically DeBERTa, can perform generative tasks** through a simple inference technique, showcasing that **masked language models** possess in-context learning capabilities akin to causal models like GPT.

- **A novel AI method** developed by Professor Markus Buehler uncovers hidden connections between **science and art**, enabling the prediction of innovative materials through generative AI and graph-based reasoning.

- The article explores the **missing metric tensor** derived from the **Stein score**, highlighting its potential to define **data-driven geometries** that adapt to the data manifold, which is crucial for understanding relationships in high-dimensional spaces.

- **Undetectable backdoors** can be inserted into ML models using **digital signatures** and **random features**, enabling malicious manipulation of outputs without detection, even under white-box conditions.

- **Vision-language-action (VLA) models** are pivotal for advancing general-purpose robotics, yet their **performance varies significantly** across tasks, with GPT-4o leading in consistency due to effective prompt engineering.

- **LLMs can play chess surprisingly well despite not being designed for it**, demonstrating an unexpected ability to interpret game states from textual move lists, hinting at their potential for broader applications in complex tasks.

- The introduction of the **first "graph foundation model"** demonstrates that **graph neural networks (GNNs)** can effectively scale with both data and model size, offering robust generalization across various domains and efficient fine-tuning capabilities on new datasets, as detailed in the [arXiv paper](https://arxiv.org/pdf/2407.11907).

- The article discusses how to derive an **efficient metric tensor** for data manifolds using the **Stein score**, highlighting its significance in the context of **differential geometry** and machine learning.

- **SelfCodeAlign** enables a **7B model** to enhance coding skills through self-generated tasks, achieving a **67.1 pass@1** on HumanEval+, outperforming larger models like CodeLlama-70B-Instruct.

- **NVIDIA Blackwell** achieves **up to 2.2x more performance per GPU** on LLM benchmarks, showcasing its capability to efficiently train large language models like GPT-3 and Llama 2 with fewer resources.

- **Japan is advancing healthcare through sovereign AI**, enhancing drug discovery, genomics, and robotics to address a projected shortage of 500,000 healthcare workers by next year, leveraging platforms like [NVIDIA BioNeMo](https://www.nvidia.com/en-us/clara/bionemo/?srsltid=AfmBOopLc2vEXb7KXJhbbEPVWBSgDYlKIf5wJWxgFedVDNuDnMp0y_Wh) and [NVIDIA Holoscan](https://www.nvidia.com/en-us/clara/holoscan/?srsltid=AfmBOoq_gsE-SSNs9mkiUUlWghhjgBZiuDanALaTqICWxR7jEFpsPT_3) for real-time insights.

- **Japan's tech leaders are leveraging NVIDIA's AI tools** like NeMo and Omniverse to enhance domestic AI capabilities, focusing on culturally nuanced language models for sectors such as healthcare and finance.