TLDR AI 2024-07-05

Intel has presented the industry's first fully integrated optical compute interconnect (OCI) chiplet designed to significantly boost high-bandwidth data interconnect performance. Essential for growing AI demands, the milestone represents a step forward in silicon photonics, combining integrated circuits with lasers to enhance data transfer rates and support AI infrastructure scaling. The OCI chiplet aims to overcome the limitations of current electrical I/O interconnects by enabling higher bandwidth and longer-distance data transmission.

Andrej Karpathy announced a new course that guides learners in building a Storyteller AI Large Language Model from scratch. It covers the fundamentals of web app implementation using Python, C, and CUDA. The content spans language modeling, deep learning frameworks, and AI deployment including multimodal applications.

Amazon has hired most of Adept's team, another aspiring OpenAI competitor, and licensed its technology, following a pattern that mirrors Microsoft's acquisition approach with Inflection. This move reflects a trend where big tech companies absorb AI talent and innovations without formal acquisitions to navigate antitrust scrutiny. Such strategies may become commonplace as the AI industry faces consolidation and high development costs.

CELLO is a new dataset with 14,094 causal questions to improve AI's grasp of causality beyond commonsense reasoning.

UIST is a novel approach that drastically improves click-through rate predictions by converting dense embeddings into compact, discrete tokens for user and item representations.

Method from Google to insert semantic objects into images with diffusion. Dataset and demo available.

Multi-Session SLAM is a new system that innovatively tracks camera movements across several unconnected video sequences under one universal frame of reference.

PTQ4SAM is a new framework designed to reduce the memory and computation demands of the large-scale Segment Anything Model (SAM).

OpenStreetView-5M is a significant open-access dataset with over 5 million geo-tagged street images from 225 countries designed to push the limits of computer vision by testing image localization capabilities.

Economist Daron Acemoglu's working paper on AI suggests a modest 0.06% annual increase in TFP growth, challenging more optimistic forecasts. He distinguishes AI's productivity impact through four channels but discounts deepening of automation and creation of new labor-intensive products, arguably underestimating AI's full economic potential. Critics find his approach overly narrow, potentially misrepresenting the broader macroeconomic implications of AI advancements.

This project introduces a method for enhancing image resolution on smartphones using a self-supervised learning model that improves reference-based super-resolution (RefSR).

SEMamba is a new speech enhancement system that leverages the Mamba state-space model to improve the clarity of speech signals.

Discover essential Apache Airflow recipes to kickstart your Generative AI projects. Check out the top 6 reference architectures to help get you started. Get the cookbook →

The small vision model runs entirely in the browser built on onnx and WebGPU.

FlexiFilm is a diffusion model designed specifically for generating long videos over 30 seconds with high consistency and quality.

OpenAI's ChatGPT macOS app, until a recent update, stored user chats in plain text, making them vulnerable to unauthorized access.