Search the archive

DIAGNOSIS

NEW FORETHOUGHT TOOL LETS YOU BUILD WORKFLOWS WITH NATURAL LANGUAGE

META IS BRINGING OUT ‘SASSY CHATBOTS’ FOR YOUNGER USERS

BING TO GAIN A NUMBER OF AI IMPROVEMENTS

SECODA SECURES $14M TO EXPAND ITS DATA CATALOGING PLATFORM

Researchers have introduced TCOVIS, an online video instance segmentation technique that prioritizes temporal consistency.

EMOTION-AWARE PROSODIC PHRASING FOR NATURAL TEXT-TO-SPEECH RENDERING

A NEW MODEL FOR GENERATING REALISTIC BODY SHAPES AND POSES

This paper proposes a model that can create super-realistic virtual humans in any pose you can think of. This is the first model of its kind trained only on fake data, but it can still make highly accurate human shapes that could be useful for science and video games.

This repository introduces FGFusion, a new method that combines data from cameras and Lidars in a detailed way for self-driving cars. Instead of just using high-level details, it captures both big-picture and fine details to create a more accurate 3D view of the surroundings.

EFFICIENT ML COURSE

Course on how to do low resource machine learning efficiently at the edge. Lectures and homework included.

FLASH ATTENTION 2 IN HUGGINGFACE TRANSFORMERS

A NEW RESULT THAT UNDERMINES LLMS

DALL-E 3 KILLS PROMPT ENGINEERING

ANNOUNCING MICROSOFT COPILOT, YOUR EVERYDAY AI COMPANION

Microsoft has unveiled "Microsoft Copilot", an AI companion integrated across Windows 11, Microsoft 365, Edge, and Bing designed to provide contextual assistance based on web context, user data, and current activities while prioritizing privacy. The feature will begin rolling out in Windows 11 on Sept. 26. The Windows 11 update will also offer 150+ new AI features in apps like Paint and Photos.

MUSK, GATES, AND ZUCKERBERG WALK INTO AN AI MEETING

Sounds like the start of a bad joke. Tech titans Elon Musk, Bill Gates, and Mark Zuckerberg met in Washington on Wednesday to talk about AI regulation. They were joined by founders from leading AI companies like Anthropic and OpenAI.

ARTIFACT CO-FOUNDER KEVIN SYSTROM DOESN’T BELIEVE IN AI DOOMERISM

Instagram co-founder Kevin Systrom, now with startup Artifact, believes AI technologies will make people "super-powered" rather than causing harm. Contrasting views of AI's existential risks exist within the tech community. Artifact uses AI to improve news consumption by summarizing articles, countering clickbait, and personalizing feeds, with aims to surface quality content over mere popularity.

MAKE STABLE DIFFUSION BETTER BY SCALING THE UNET

You can substantially improve generation quality in Stable Diffusion without any additional training by cleverly shifting around weights and scaling residuals in the UNet. Code coming soon.

MAKING SMARTER 'MINI' MODELS USING KNOWLEDGE DISTILLATION

This study explores how to make 'student' networks (smaller models) learn better from 'teacher' networks (bigger models) in situations they've never seen before. It introduces a new method called Weight-Averaged Knowledge Distillation (WAKD) that does just as well as current techniques but is simpler to use.

HOW WELL MODELS UNDERSTAND SOUND AND VISION TOGETHER?

This study introduces AV-SUPERB, a new benchmark that tests how well a trained model understands both sound and visual data across a variety of tasks.

ML ENGINEERING GUIDE

A thorough resource for improving ML engineering skills.

DREAMLLM: CREATE BOTH TEXT AND IMAGES TOGETHER

This report introduces DreamLLM, a groundbreaking tool that is good at understanding and making content that combines both words and pictures.

A FASTER WAY TO ANALYZE 3D FACIAL PHOTOS

Researchers trained a model that identifies important points on 3D pictures of faces, a process usually done by hand. Their method was almost as accurate as human experts, but much faster.

AI is looking to take on several dying markets and news is just another avenue. This article looks at how AI might be a path forward for news if generated/managed with AI. We’re already seeing this happen with smaller local TV affiliates with some success.

THREE IN FOUR AMERICANS BELIEVE AI WILL REDUCE JOBS

Most Americans think artificial intelligence will negatively affect the U.S. job market, with 75% saying it will decrease the total number of jobs over the next 10 years.

AUTOMATIC KNOWLEDGE GRAPH CREATION

Generating knowledge graphs from unstructured text was a holy grail of natural language understanding - now you can do it with 50 lines of code and an LLM.

OPENAI LAUNCHES DALL·E 3

DALL·E 3 understands significantly more nuance and detail than previous systems, allowing users to easily translate your ideas into exceptionally accurate images. It will soon be integrated natively into ChatGPT.

GITHUB’S AI-POWERED CODING COPILOT IS NOW WIDELY AVAILABLE

GitHub is expanding the availability of its AI-powered coding chatbot, Copilot Chat. Previously in public beta for business users, Copilot Chat is now available to individual users in Visual Studio and Visual Studio Code.

TRI IS DEVELOPING A NEW METHOD TO TEACH ROBOTS OVERNIGHT

The Toyota Research Institute (TRI) is showcasing advancements in research that can teach a robot a new skill overnight. TRI utilizes traditional robot learning techniques and diffusion models similar to generative AI models. Its system has been trained on 60 skills, focusing on versatility in different environments, aiming to make robots that can operate effectively in diverse settings and adapt to new tasks.

A BETTER WAY TO UNDERSTAND HAND MOVEMENTS

This study introduces RenderIH, a massive collection of ultra-realistic images showing hands in various positions that tackles issues with older datasets that were too simple. It also presents TransHand, a new tool that reads these hand positions very accurately.

BOOSTING RECOMMENDATIONS USING LARGE LANGUAGE MODELS

This study looks at how LLMs, which are revolutionizing many tech areas, can make sequential recommendations (like suggesting what song or movie comes next) even better.

MAKING LLMS USEFUL IN MEDICINE

This research focuses on making large language models, which are already great at understanding and generating text, even better at answering medical questions.

A NEW APPROACH TO BETTER UNDERSTAND RGB-D IMAGES

DFormer is a unique way to study both color and depth in images. Unlike older methods, DFormer is designed especially for RGB-D images.

TURN EXPENSIVE PROMPTS INTO CHEAP MODELS

OpenPipe is a tool for model distillation based on few shot described behavior via prompting.

RETRIEVAL IS MORE THAN JUST EMBEDDING SEARCH

Language models regularly hallucinate details to factual queries. By retrieving the facts, we can reduce these inaccuracies. However, the naive form of Retrieval Augmented Generation (RAG) of embedding a query and searching against a vector store often performs poorly. This post discusses how to dramatically improve performance with query objects and information retrieval techniques.

GENERATIVE AI’S ACT TWO

Generative AI’s first act of developing foundational models and novelty apps is coming to an end. In its place comes Act Two, where generative AI solves human problems end-to-end.

HOW GOOGLE TAUGHT AI TO DOUBT ITSELF

Google's Bard chatbot now offers a "double-check" feature, which evaluates its responses by checking the web for corroborating or contradictory information when prompted by users. This addition aims to enhance accuracy and accountability but may still require users to verify search results and interpretations themselves.

HOW AI CAN SPARK BOTH TECHNOLOGICAL AND BUSINESS MODEL INNOVATION

Truly great innovation comes in pairs. Technology innovation and business model innovation work together to deliver outlier results. AI is the next technology innovation, but have we found its counterpart on the business model side yet?

TELLING AI TO “TAKE A DEEP BREATH” CAUSES MATH SCORES TO SOAR

Google has developed Optimization by PROmpting (OPRO), a technique to improve the math skills of LLMs using human-style encouragement, such as “let’s think step by step”.

Microsoft is planning a special event to unveil its vision for integrating AI into Windows, Microsoft 365, Surface, and more.

OPENAI LAUNCHES A RED TEAMING NETWORK

OpenAI has introduced the OpenAI Red Teaming Network, a group of experts that will assist in assessing and mitigating the risks associated with its AI models.

ANYSCALE OPTIMIZES OPEN SOURCE AI DEPLOYMENTS WITH ENDPOINTS

Anyscale has unveiled Endpoints, which allows developers to integrate fast, cost-efficient, and scalable LLMs into their applications using popular LLM APIs. The company also announced a new integration with NVIDIA to boost the performance and development of LLMs on Ray, Anyscale’s open-source framework for distributed machine learning.

BARD CAN NOW FIND ANSWERS IN YOUR GMAIL, DOCS, AND DRIVE

Google's Bard chatbot can now scan Gmail, Docs, and Drive to retrieve information and perform tasks. Google says these new integrations, called extensions, aim to save users time when searching for specific information within their documents and emails.

6T MULTILINGUAL TOKENS

Hot on the heels of Google's new dataset comes CulturaX. A cleaned dataset of trillions of multilingual tokens will be a great boost for multilingual language models.

IMPROVING TEXT-TO-IMAGE GENERATION

Turning long texts into images can be tough, especially when there are lots of items or details to show. This new method breaks down long descriptions into simpler steps and makes sure each detail fits perfectly in the picture, setting a new high standard for creating images from text.

MAKING DEPTH ESTIMATION MORE FLEXIBLE

Figuring out 3D depth from a single 2D image is tricky because many 3D scenes can look the same in 2D. This method separates camera details from the picture's content, making it easier to guess the depth in various situations.