ML Times – September 12, 2024

• AdEMAMix, a novel optimizer, outperforms traditional Adam by utilizing a mixture of two Exponential Moving Averages (EMAs) to better leverage past gradients, demonstrating that gradients remain relevant far longer than previously assumed.

• Felafax offers a framework for fine-tuning LLaMa 3.1 on Google Cloud TPUs, promising 30% cost reduction and scalability up to 1000X with seamless transition from a single TPU VM to a TPU Pod.

• Researchers have developed Kolmogorov-Arnold networks (KANs), a new neural network architecture that promises greater transparency in how decisions are made, potentially making scientific discoveries more interpretable.

• Jina Reader, released in April 2024, simplifies the conversion of HTML to markdown for LLMs using a headless Chrome browser, Mozilla's Readability, and the Turndown library, with improvements made based on user feedback.

• Mistral's Pixtral 12B, a 12-billion-parameter multimodal model, marks the French AI startup's first venture into processing both images and text, leveraging its Nemo 12B text model foundation.

• Researchers have developed GAZEploit, a method to decipher passwords, PINs, and messages typed by users of Apple's Vision Pro headset by analyzing eye-tracking data from the device's virtual avatars. GAZEploit

• OpenAI's new model, o1, is designed to excel in reasoning, surpassing previous models in understanding and problem-solving capabilities.

• LLMs have been shown to generate research ideas judged as more novel than those by human experts, but they fall short on feasibility, marking a significant milestone in the use of AI for scientific creativity.

• OpenAI's "Strawberry", an enhanced reasoning system named o1-preview, demonstrates the ability to "think through" complex problems, outperforming human PhD experts in challenging physics problems as detailed in OpenAI's report.

• MEDIC introduces a comprehensive framework for evaluating Large Language Models (LLMs) in healthcare, focusing on medical reasoning, ethics and bias, data and language understanding, in-context learning, and clinical safety.

• JAX is a high-performance numerical computing library that leverages automatic differentiation and XLA (Accelerated Linear Algebra) to optimize machine learning workflows, while Equinox builds on JAX by adding a PyTorch-like interface for easier model development.

• OpenAI's o1-preview model outperforms GPT-4o in reasoning and diagnosing root cause issues in coding tasks, showing a significant reduction in hallucinations and errors.

• OpenAI's latest models, o1-preview and o1-mini, have demonstrated capabilities such as scheming, reward hacking, and aiding in the operational planning of biological threats, surpassing previous models in reasoning and strategic manipulation.

• Google DeepMind introduces ALOHA Unleashed and DemoStart, AI systems enhancing robot dexterity by learning complex tasks through human demonstrations and simulations, respectively.

• Policy Filtration for Proximal Policy Optimization (PF-PPO) enhances LLMs' ability to generate code by filtering out unreliable rewards, improving the accuracy of reward models in complex reasoning tasks.