Search the archive

ULTRASOUND IMAGE CLARITY

RDSTN is a new network that enhances ultrasound images through arbitrary-scale super-resolution, addressing the trade-off between image quality and field-of-view.

TEXT-BASED PERSON RETRIEVAL

UFineBench is a new benchmark in text-based person retrieval. It uses detailed descriptions to help AI better understand and find people in images.

IMPROVED OBJECT SEGMENTATION

SegRefiner, a new model-agnostic solution, improves object mask quality in various segmentation tasks by interpreting refinement as a data generation process. It refines coarse masks pixel-by-pixel using a discrete diffusion process, enhancing segmentation and boundary metrics.

CODEL

Fully autonomous AI Agent that can perform complicated tasks and projects using terminal, browser, and editor.

EFFICIENT SPATIOTEMPORAL FORECASTING

VMRNN cell is a new architecture that combines Vision Mamba blocks and LSTM for enhanced spatiotemporal forecasting.

DETR ENHANCES OBJECT DETECTION

This project introduces Salience DETR, which leverages hierarchical salience filtering to refine query selection in object detection, offering a balance between computational efficiency and accuracy.

MATHEMATICIANS USE AI TO IDENTIFY EMERGING COVID-19 VARIANTS

University of Manchester mathematicians have developed an AI tool, CLASSIX, to rapidly identify emerging COVID-19 variants from genetic data, potentially streamlining the tracking of viral evolution and aiding in proactive responses like tailored vaccine development. The method, which is computationally less demanding than traditional phylogenetic analysis, processed 5.7 million sequences in 1-2 days on a standard laptop, demonstrating its efficiency in managing large-scale genomic data. The research signifies a step towards quicker, resource-efficient identification of significant pathogen strains.

TRENDS IN SYNTHETIC BIOLOGY & AI IN DRUG DISCOVERY IN 2024

2024 is set to be a monumental year for AI in drug discovery, marking considerable advancements in synthetic biology. This article highlights key trends, including generative AI's influence on science and the synthesis of modular biological components. AI's integration with synthetic biology is accelerating the drug discovery process, as demonstrated by Insilico Medicine's AI-powered candidate's progression to Phase II clinical trials.

LLMS HAVE SPECIAL INTELLIGENCE, NOT GENERAL, AND THAT'S PLENTY

Anthropic's new AI model Claude-3 surpasses the average human IQ and outperforms other models, including GPT-4, in advanced cognitive tasks. Despite this achievement, Claude-3 still struggles with basic challenges like puzzles that humans find easy. LLMs may possess a 'Special Intelligence' rather than human-like general intelligence. They could be reflecting our knowledge back to us in innovative ways.

Install the Kit CLI

IOS 18 REPORTEDLY WON’T FEATURE APPLE’S OWN CHATGPT-LIKE CHATBOT

Apple will reportedly not debut its own generative AI chatbot in its next major software updates, including iOS 18, but is discussing potential generative AI partnerships with companies like Google, OpenAI, and Baidu while planning to enhance various apps with generative AI features for user assistance.

ARC (PRODUCT)

AI-powered remote jobs marketplace to hire software developers from around the world.

AI SAAS COMPANIES WILL BE MORE PROFITABLE

AI software companies might initially have higher expenses, but their profitability could surpass traditional SaaS companies due to AI's deflationary effects across marketing, sales, operations, and software development.

SORA: FIRST IMPRESSIONS

A compilation of Sora content generated from visual artists, designers, creative directors, and filmmakers.

OPEN INTERPRETER O1 LIGHT

The 01 Light is a portable voice interface that controls your home computer. It can see your screen, use your apps, and learn new skills. 01 is the open-source foundation for a new era of AI devices.

CHARACTER VOICE FOR EVERYONE

Character Voice is a suite of features that allows users to hear Characters speaking to them in 1:1 chats, taking the Character.AI experience to the next level. It is the first step in the company's larger plan to build a multimodal interface, which will facilitate more seamless, intuitive, and engaging interactions.

GENERATING REALISTIC SHADOWS IN IMAGES

This study introduces a new method for creating realistic shadows in image composition, overcoming previous challenges with shape and intensity accuracy. The researchers significantly improved shadow generation in images by enhancing ControlNet with intensity modulation modules and expanding the DESOBA dataset.

PEOPLE RECOGNITION FROM DRONE AND GROUND CAMERAS

IMAGE GENERATION WITH FLEXIBLE SIZES AND ASPECT RATIOS

ElasticDiffusion is an innovative decoding method that allows text-to-image diffusion models to create images in various sizes and aspect ratios without additional training.

IMAGE SEGMENTATION WITH PSALM

PSALM is an extension of the Large Multi-modal Model (LMM), which introduces a mask decoder and a versatile input schema to excel in various image segmentation tasks. This approach not only overcomes the limitations of text-only outputs but also allows the model to understand and classify complex images effectively.

CEREBRAS CS3 CHIP

Cerebras' new wafer chip can train 24T parameter language models. It natively supports PyTorch.

IMPROVED IMAGE PERSONALIZATION

Researchers have developed a new method to improve how AI creates personalized images, addressing overfitting issues. This approach ensures a more balanced and diverse representation of concepts in the images.

THE GPT-4 BARRIER HAS FINALLY BEEN BROKEN

GPT-4's dominance in AI benchmarks has been challenged by four new models from different vendors, each showing the potential to surpass GPT-4's capabilities. However, concerns arise as, amidst growing legal and ethical considerations, none of these models are open-source or transparent about their training data. The push for models trained on public domain or licensed content continues, highlighting the complexity of creating competitive AI without proprietary data.

CHINA PUTS TRUST IN AI TO MAINTAIN LARGEST HIGH-SPEED RAIL NETWORK ON EARTH

China's high-speed rail network has seen an 80% decrease in minor track faults and no major track irregularity warnings in the past year thanks to AI and machine learning technologies. The success in proactive safety and maintenance was supported by AI's analysis of extensive data from the railway's sensors. Despite challenges, such as a shrinking workforce and U.S. sanctions on AI chips, China continues to advance in specialized AI applications across various sectors.

GO, PYTHON, RUST, AND PRODUCTION AI APPLICATIONS

the blog

LOW-LATENCY GENERATIVE AI MODEL SERVING WITH RAY, NVIDIA TRITON INFERENCE SERVER, AND NVIDIA TENSORRT-LLM

Anyscale and NVIDIA announced a new partnership enabling customers to scale generative AI models into production. Through the integration, customers can tap into the combined power of Ray and Anyscale’s managed runtime environment to improve resource management, observability, and autoscaling.

AI TOOLS DIRECTORY

Discover the best AI websites and tools.

MICROSOFT TO HOLD A SPECIAL WINDOWS AND SURFACE AI EVENT IN MAY

Microsoft is planning an AI-focused event on May 20th featuring CEO Satya Nadella, who will discuss the company's AI vision in hardware and software.

EMAD RESIGNS AS STABILITY AI CEO

Emad Mostaque, a prominent and divisive figure, has resigned as CEO of Stability AI to pursue decentralized AI. Stability AI's COO and CTO are now Co-CEOs as the company searches for a permanent figure.

OPENAI IS PITCHING SORA TO HOLLYWOOD

OpenAI is pitching its video generation tool Sora to a number of Hollywood studios, talent agencies, and media executives.

GITHUB’S LATEST AI TOOL CAN AUTOMATICALLY FIX CODE VULNERABILITIES

GitHub introduced a beta for code-scanning autofix, merging GitHub’s Copilot with CodeQL to detect and automatically remedy security vulnerabilities in JavaScript, TypeScript, Java, and Python.

DEEPDIVE: USING CHATGPT TO BUILD C EXTENSIONS FOR SQLITE

This post contains a deep dive into how to use ChatGPT (or any language model) to write code in an unfamiliar language for a challenging task. Its author uses ChatGPT’s code interpreter to write, compile, and download new bindings for the popular database SQLite.

IMAGE SEGMENTATION WITH OPEN-VOCABULARY ATTENTION MAPS

Open-Vocabulary Attention Maps (OVAM) modify text-to-image diffusion models like Stable Diffusion by allowing the generation of attention maps for any word, bypassing previous limitations.

PRIVACY-PRESERVING TRANSFER LEARNING

HETAL introduces a new approach to transfer learning that ensures data privacy through Homomorphic Encryption, marking a significant step forward in secure AI training.

OFFICIAL MISTRAL FINE-TUNING CODE

Mistral recently held a hackathon. As part of it, the company announced v0.2 of its 7B model and released code on how to fine-tune its language models. The code is clean and concise.

OPTIMAL TRANSPORT FOR BIG DATA

Optimal Transport theory is a rich area of mathematics and computation that deals with distances between probability distributions. This repository contains an extremely comprehensive reading list for scaling Optimal Transport in both speed and accuracy.

ENHANCED IMAGE RESTORATION WITH ADALR

AdaIR introduces an all-in-one image restoration network that leverages frequency mining and modulation to tackle various forms of image degradation such as noise, blur, and haze.

UNTANGLING CONCERNS ABOUT CONSOLIDATION IN AI

The recent acquisition of Inflection's talent by Microsoft reignited debates around the biggest tech companies gaining too much control over AI development. While their resources allow rapid progress on core language models, there are valid worries about the concentration of power limiting innovation and transparency. This article explores the complex tradeoffs as the commercialization of AI moves forward.

TIPS FOR FP8 TRAINING

The Databricks Mosaic team has continued to innovate on their language model training. In this post, they discuss the training stack used for fp8 and some of the benefits that come from reducing precision.

COMPRESSING 3D GAUSSIAN SPLATTING DATA

This project introduces the Hash-grid Assisted Context (HAC) framework, which achieves over 75X compression of 3D Gaussian Splatting (3DGS) data, outperforming current standards.

collects evidence, and much more. Try an interactive demo

RESEARCHERS GAVE AI AN 'INNER MONOLOGUE' AND IT MASSIVELY IMPROVED ITS PERFORMANCE

A new method called "Quiet-STaR" enhances AI reasoning by giving systems an "inner monologue," allowing for improved performance on reasoning tests, as shown by its application to the Mistral 7B language model.

‘A LANDMARK MOMENT’: SCIENTISTS USE AI TO DESIGN ANTIBODIES FROM SCRATCH

Researchers have used generative AI to design new antibodies, potentially revolutionizing the therapeutic antibody market.

DEEPGRAM AURA (PRODUCT)

Less than 250ms latency text-to-speech API.