Jan 22, 2025

• Stargate Project: SoftBank, OpenAI, Oracle, MGX to build data centers
  Trump announces a partnership involving OpenAI, Oracle, and SoftBank to invest up to $500 billion in AI infrastructure, with initial funding of $100 billion aimed at building data centers in Texas, reflecting a significant commitment to advancing AI technology in the U.S.

• Tensor Product Attention Is All You Need
  Tensor Product Attention (TPA) leverages tensor decompositions to compactly represent queries, keys, and values, drastically reducing key-value cache size during inference, thus enhancing memory efficiency.

• Hunyuan3D 2.0 – High-Resolution 3D Assets Generation
  Hunyuan3D 2.0 is a cutting-edge 3D synthesis system that integrates a shape generation model and a texture synthesis model, enabling the creation of high-resolution, textured 3D assets efficiently.

• Infinigen
  Infinigen is a procedural generator of 3D scenes, developed by the Princeton Vision & Learning Lab, that creates diverse, high-quality training data optimized for computer vision research using real geometry without faked details.

• Flame: A small language model for spreadsheet formulas (2023)
  FLAME is a transformer-based model specifically designed for Excel formulas, achieving competitive performance with only 60M parameters and significantly less training data compared to larger models.

• Arm releases Chiplet System Architecture spec beta version
  Arm's Chiplet System Architecture (CSA) has released its first public specification, engaging over 60 companies to standardize chiplet designs, enhancing flexibility and reducing costs in silicon production.

• Show HN: BrowserAI – Run LLMs directly in browser using WebGPU (open source)
  BrowserAI enables local execution of LLMs in the browser, ensuring data privacy and eliminating server costs, making it ideal for developers and researchers alike.

• How to solve computational science problems with AI: PINNs
  Physics-Informed Neural Networks (PINNs) effectively integrate physical laws into neural network training, enabling the solution of complex partial differential equations (PDEs) with improved accuracy and efficiency.

• [R] Learning to Continually Learn with the Bayesian Principle
  Novel meta-continual learning framework combines the representational power of neural networks with the robustness of Bayesian models, effectively preventing catastrophic forgetting during continual learning.

• From Drafts to Answers: Unlocking LLM Potential via Aggregation Fine-Tuning
  Aggregation Fine-Tuning (AFT) enhances large language models by synthesizing multiple draft responses into a single, refined answer, demonstrating significant performance improvements over standard supervised fine-tuning (SFT).

• Into the Omniverse: OpenUSD Workflows Advance Physical AI for Robotics, Autonomous Vehicles
  OpenUSD workflows are enhancing physical AI capabilities, enabling robots and autonomous vehicles to understand and interact with the real world through advanced simulation environments that replicate physical dynamics and spatial relationships.

• Test-time regression: a unifying framework for designing sequence models with associative memory
  Test-time regression serves as a unifying framework for sequence models, revealing that associative recall is essential for effective performance, bridging the gap between various architectures like transformers and recurrent networks.

• AdaServe: SLO-Customized LLM Serving with Fine-Grained Speculative Decoding
  AdaServe is the first LLM serving system that enables SLO customization through fine-grained speculative decoding, enhancing token prediction accuracy and throughput.