TLDR AI 2023-12-22

RITE AID BANNED FROM USING FACIAL RECOGNITION

Rite Aid used facial recognition software that falsely identified shoplifters, leading to dramatic confrontations. The FTC is cracking down more on this technology.

ANTHROPIC REPORTEDLY IN TALKS TO RAISE $750M AT A $15B VALUATION

Anthropic is reportedly seeking to raise $750M at a valuation of $15B in a round led by Menlo Ventures.

APPLE’S LATEST AI RESEARCH COULD COMPLETELY TRANSFORM YOUR IPHONE

Apple has released two papers introducing new techniques for 3D avatars and efficient language model inference, which could allow AI systems to run on iPhones and iPads.

ULTRAFAST 3D RECONSTRUCTION FROM A SINGLE IMAGE

Gaussian Splatting has pushed 3D reconstruction forward. However, it often requires 200+ images of an object for accurate reconstruction. This work moves to a single monocular image and a learning-based technique that maps each pixel of the image into a 3d gaussian. The cool novel piece here is the introduction of the “Splatter Image,” which encodes 3D information for downstream rendering.

POWERINFER FOR LLMS ON CONSUMER GPUS

This paper explores the idea that models have hot neurons used for almost all inputs and cold neurons that are used much less often. By preloading the hot neurons to the GPU, you can save memory without sacrificing much throughput. A companion code library is available.

IMAGE GENERATION WITH SEQUENCE-AWARE LOSS

Researchers have developed a 'sequence-aware' loss function to improve Denoising Probabilistic Models (DPMs), addressing the issue of timestep correlation in image generation. This new approach not only offers a tighter estimation of loss but also shows significant improvements in image quality on datasets like CIFAR10 and CelebA, as evidenced by better FID and Inception Scores.

Get 60% off Incogni with code TLDR60

LLAMA RECIPES

A set of tools, resources, and examples for using the Llama series of models from Facebook.

A ZERO-SHOT SEMANTIC SEGMENTATION

The new CLIP-DINOiser method combines the zero-shot capabilities of the CLIP model with self-supervised features for improved semantic segmentation without annotations.

LLAVA INTERACTIVE

Llava is a strong multimodal model. This interactive design now allows chat style interleaving of text, image input, and image output.

YEAR ONE OF GENERATIVE AI: SIX KEY TRENDS

AI has been in the spotlight throughout 2023. As we round out the year, what’s really happened? This article highlights the key trends that have defined AI in 2023.

MARKETPLACES IN THE AGE OF AI

Marketplaces have been a dominating business model over the last two decades. Here’s a breakdown from a16z sharing their thoughts on how AI will impact this company type. The central theme is about personalizing the experience for both sides of the marketplace.

IMPROVING TEXT IN AI-GENERATED IMAGES

UDiffText is a new method that enhances the quality of text in AI-generated images. By using a specialized text encoder and fine-tuning with a large dataset, UDiffText significantly reduces spelling errors and improves text accuracy.

ENVSION AI (PRODUCT)

Summarize hours of customer interviews in minutes.

WHISPER TURBO 6X SPEED UP

Fast implementation of whisper that is dramatically faster than vanilla whisper for speech-to-text.

OPENAI BUFFS SAFETY TEAM AND GIVES BOARD VETO POWER ON RISKY AI

OpenAI has created a new safety advisory group and has granted the board veto power over all models.

YOUR REFERRAL LINK - https://tldr.tech/ai?ref=2613930