Aug 28, 2024

• Sapiens: Foundation for Human Vision Models
  Sapiens models, designed for human-centric vision tasks like 2D pose estimation and depth estimation, leverage self-supervised pretraining on over 300 million human images for enhanced performance.

• Diffusion Models Are Real-Time Game Engines
  GameNGen, the first game engine powered entirely by a neural model, enables real-time interaction with complex environments, simulating the classic game DOOM at over 20 frames per second on a single TPU.

• DisTrO – a family of low latency distributed optimizers
  DisTrO significantly reduces inter-GPU communication requirements, enhancing distributed training efficiency by three to four orders of magnitude.

• [R] Playable 20FPS Doom via a finetuned SD1.4 model from Google research team
  GameNGen, the first game engine powered by a neural model, enables real-time interaction with complex environments, simulating DOOM at over 20 frames per second on a single TPU.

• Tesla's TTPoE at Hot Chips 2024: Replacing TCP for Low Latency Applications
  Tesla's TTPoE protocol aims to replace TCP for low latency applications by simplifying state machines and removing wait states, targeting microsecond scale latencies.

• Splatt3R: Zero-Shot Gaussian Splatting from Uncalibrated Image Pairs
  Splatt3R introduces a pose-free, feed-forward method for 3D reconstruction and novel view synthesis from uncalibrated stereo pairs, predicting 3D Gaussian Splats without needing camera parameters or depth information.

• "Writing in the Margins (WiM)" - a better inference pattern for long context LLMs that solves the Lost-in-the-Middle problem
  Writing in the Margins (WiM) introduces a new inference pattern for Large Language Models, optimizing long input sequence handling through chunked prefill and segment-wise inference.

• Launch HN: Bucket Robotics (YC S24) – Defect detection for molded and cast parts
  Bucket Robotics introduces a novel approach to defect detection in manufacturing by transforming CAD models into custom defect detection models, aiming to improve upon the traditional 80% human success rate and manual or traditional ML-based methods that rely on real-world sample imaging.

• OpenAI shows 'Strawberry' to feds, races to launch it
  OpenAI's new AI model, Strawberry, is designed to solve complex problems accurately on the first attempt, leveraging a technique known as process supervision.

• The Mamba in the Llama: Distilling and Accelerating Hybrid Models
  Linear RNN architectures like Mamba can rival Transformer models in language tasks, benefiting from efficient deployment characteristics and the ability to distill large Transformers into more deployable forms using existing linear projection weights.

• When A.I.'s Output Is a Threat to A.I. Itself
  A.I.-generated data's indistinguishability poses a risk of creating feedback loops where A.I. trains on its own output, leading to degraded performance and model collapse.

• PyRoboCOP: Python-Based Robotic Control and Optimization Package
  PyRoboCOP is a Python-based package designed for robotic control and optimization, specifically addressing manipulation tasks and collision avoidance through handling complex complementarity constraints.

• Show HN: Repo2vec – an open-source library for chatting with any codebase
  repo2vec is a modular library designed for chatting with codebases, simplifying the process of understanding and integrating code without manual code review.

• [D] Image segmentation converges to all zeros when masks are too small
  Image segmentation models, such as UNet, struggle with small masks, leading to convergence on all-zero outputs, particularly in breast cancer image segmentation tasks.

• [D]Exploring the Potential of Edge Computing/Federated Learning in Continuous Training for GPT/LLMs
  Exploring the potential of Edge Computing and Federated Learning could revolutionize the way GPT and Large Language Models (LLMs) are continuously trained, by decentralizing the process.