Search the archive

TRACK HUNDREDS OF POINTS IN REAL TIME

PIPS2 is a state-of-the-art point-tracking system that performs dense pixel-level tracking and is robust to basic occlusions and fast motion.

USING LANGUAGE MODELS TO IMPROVE HOW THINGS LOOK ONLINE

Creating eye-catching designs online is important for getting users interested and helping them understand information. This new method, called LayoutNUWA, makes these designs smarter by using language models to turn coding instructions into awesome layouts.

DOLA: A SIMPLE WAY TO MAKE LANGUAGE MODELS STICK TO THE FACTS

Researchers have introduced a new method called DoLa that helps large language models (like the ones that chat with you or write articles) stop making stuff up and stick to the facts. It uses a smart way of comparing different layers inside the model to filter out false information.

STATE-OF-THE-ART BACKGROUND REMOVAL IN HF TRANSFORMERS

Matting is the process of creating a black-and-white mask to segment out important parts of an image. It is usually used for foreground and background segmentation. The portrait mode on smartphones uses a method similar to ViTMatte described here.

EXISTED

Google’s AlphaFold played a crucial role in determining the specific structures of two intrinsically disordered proteins that induce previously unidentified structures when they interact. The discovery sheds light on how such proteins function and brings valuable insights into complex biological questions that were previously challenging to answer.

MICROSCOPE

The Department of Defense has teamed up with Google to build an AI-powered microscope that can help doctors better identify cancer.

WRITER NETS $100M FOR ITS ENTERPRISE-FOCUSED GENERATIVE AI PLATFORM

Writer, who is developing a "full-stack" generative AI platform for businesses, has secured $100M in Series B funding led by ICONIQ Growth. The funds will advance its "industry-specific" text-generating AI models. Differentiating itself in a competitive market, Writer focuses on non-copyrighted business writing, cost-effective models, transparency, and ensuring that no training is done using customer data.

GLOOM

Databricks has secured over $500 million in a Series I round, boosting its valuation to $43 billion. Noteworthy investors include T. Rowe Price, Morgan Stanley, and Nvidia. Despite wider market funding slowdowns, Databricks' strong revenue growth and run rate suggest an eventual IPO could exceed its current valuation.

NEW LANGUAGE MODEL OBJECTIVE TRAINS FASTER WITH TIED EMBEDDINGS

Typically with language models, the input and output embedding layers are tied together. However, if you tie the input embedding to other embeddings within the model, and use a contrastive loss, you can get dramatically improved performance both in wall clock and final benchmark accuracy.

FINE-TUNING VISION TRANSFORMERS MORE EFFICIENTLY

Fine-tuning pre-trained vision models for specific tasks often requires a lot of extra computing power. Researchers have come up with a method called "Salient Channel Tuning" (SCT) that smartly picks which parts of a model to fine-tune, using way fewer extra settings and still outperforming other methods in almost all tested tasks.

WHEN LLMS MEET EXTERNAL INFO

Large language models like GPT-3 are smart because they learn from a ton of data, but how do they react when new information conflicts with what they already know? Researchers have found that these models can get confused and give wrong answers, especially when they face conflicting or distracting info, raising concerns about how reliable they can be.

WHY OPEN SOURCE AI WILL WIN

Open source AI models have significant advantages over proprietary models. For most AI applications, controllability, customizability, transparency, and trust are more important than advanced reasoning abilities. AI-native companies should focus on owning their core models rather than outsourcing their intelligence layer to external providers.

BLENDING 2D AND 3D FOR IMAGE CREATION

Researchers have found a way to make 2D and 3D technologies work together seamlessly for better and more flexible designs.

ROBUST E-NERF: BETTER 3D MODELS

Event cameras have some cool benefits like low power use and quick response times, but they struggle with creating detailed 3D models of scenes. Researchers have introduced Robust e-NeRF, a new method that makes these cameras much better at building 3D models, even in challenging conditions like fast motion or varying light.

IT’S TOO EARLY TO REGULATE AI

This article makes the argument that we should wait to regulate AI, as regulation frequently results in reduced competition and production, economic benefits moving overseas, and a government-centric instead of a user-centric industry.

WHY AI CAN’T MAKE HUMAN CREATIVITY OBSOLETE

Despite concerns that AI would render human creativity obsolete, we continue to appreciate human endeavors, like chess, even when AI performs better. This paradox arises because humans value imperfection, surprises, and shared experiences, which AI, with its constant optimization, cannot replicate.

GE IS WORKING ON AI-POWERED ULTRASOUNDS

GE Health is set to create an AI-assisted ultrasound imaging tool designed for easy use by healthcare providers without specialized training. The device aims to improve medical imaging and will primarily focus on maternal, fetal, and pediatric lung health to address preventable maternal and child mortality.

VIMGPT

An experimental LLM agent with direct Vim access, intended for use either as a sub-agent for efficient file editing or as its own agent runtime.

GET PRICING PLANS

Adobe's Firefly generative AI models, previously in beta, are now available in Creative Cloud, Adobe Express, and Adobe Experience Cloud. Users will be charged via "generative credits" based on model interaction. Paid Adobe plans come with various credit amounts. Firefly will still work (but run slower) once credits run out.

COULD TARGET WORLD HUNGER

Industry heavyweights like Musk, Gates, and Zuckerberg debated AI regulation in a private session, discussing concerns about open-source misuse and job threats. A consensus emerged on the need for government oversight.

IBM THROWS HAT IN LANGUAGE MODEL ARENA WITH OPEN SOURCE MOE MODELS

Mixture-of-Experts models are sparse models, which means they save on computation while not sacrificing accuracy. They're tricky to train. IBM has released several along with their architecture and some code.

COMPREHENSIVE SURVEY ON AI COPYRIGHT ISSUES

This recently updated work dives into the many angles of AI copyright. AI copyright can broadly be classified into Input and Output legal challenges. The paper doesn’t present any opinions - it is merely a summary of many of the common issues.

MAKING ONE AUDIO MODEL WORK FOR SPEECH, MUSIC, AND SOUNDS

Researchers have developed a method called EncodecMAE that borrows ideas from text-based models to make a universal audio model. This new approach works well for different kinds of audio tasks like understanding speech, identifying music, or recognizing sounds in the environment.

A GAME-CHANGER IN REMOTE SENSING RESEARCH WITH SYNTHETIC DATA

SyntheWorld is a groundbreaking collection of fake but super-detailed images that help researchers study the Earth from above without the usual big costs and headaches. With 40,000 images, it's perfect for learning how land changes over time.

MULTIVIEW SYNTHETIC GETS A BOOST

Recent improvements in the field of 3D generation have been blistering. This work is the most stable and general. You can input an image and the model will synthesize many views and a resulting 3D model.

MAKING AI 'COPILOTS' EVEN BETTER AT FIXING CODE

Researchers have created Repilot, a system that makes AI copilots better at helping programmers fix bugs in their code.

USING TRANSFORMERS FOR BETTER SLEEP ANALYSIS

Transformers are smart algorithms usually used for text or images. This project tweaked them to work with EEG data to better understand sleep stages.

SECURING AI IS THE NEXT BIG PLATFORM OPPORTUNITY

This article breaks down the risks and challenges surrounding AI security and looks at the opportunities for companies in this new world order.

CENTAURS AND CYBORGS ON THE JAGGED FRONTIER

Consultants who use GPT-4 vastly outperform consultants who don’t. A recent study provided further evidence that knowledge workers who utilize AI will be the winners in the future.

CLICK AND WATCH YOUR IMAGE MOVE

Generative dynamics uses both modern and classic AI to synthesize motion in a still image. It is completely interactive as well.

GPT-4 IS NOT GETTING WORSE

Contrary to popular opinion, GPT-4’s performance is not systematically declining.

MICROSOFT OPEN SOURCES EVODIFF

Microsoft has open-sourced EvoDiff, a novel AI model that the company claims can generate high-fidelity, diverse proteins given a protein sequence.

PATRONUS AI RAISES $3M SEED TO RAISE CONFIDENCE IN ENTERPRISE LLMS

Led by Lightspeed, the team wants to build real world scoring to help understand how useful LLMs are for enterprises.

STORAGE

MLPerf has released results from two benchmark suites: MLPerf Inference v3.1, which showcases record participation and performance gains, and MLPerf Storage v0.5, which assesses storage system performance for ML training workloads. The Inference benchmark suite introduced a large language model and updated recommender tests, reflecting emerging AI trends.

ALIGN A LANGUAGE MODEL WITHOUT TRAINING

Alignment helps make language models more helpful and harmless. It can sometimes hurt performance, but in general is a positive. Alignment is expensive and requires substantial alignment data. However, if you allow the model to rewind during generation after evaluating its own output, it can improve alignment performance on the frozen model by up to 81%.

PUSHING MIXTURE OF EXPERTS TO THE LIMIT WITH PARAMETER EFFICIENCY

Mixture of experts (MoEs) is a cool way to increase the capacity of a model without increasing per token runtime. However, they're still tricky to run quickly and fine-tune. Researchers find that if you modify dense model parameters efficiently fine tuning to work with MoEs you can dramatically reduce tuning costs without hurting performance too much.

MAKING BETTER RECOMMENDATIONS

Researchers have developed a new model called HAMUR to improve how models make suggestions or "recommendations" across multiple subjects or domains, like music, books, or movies. Unlike older methods that mix up information between these domains, HAMUR uses a special technique to keep data separate and more flexible.

LLM APPLICATIONS

A comprehensive guide to building RAG-based LLM applications for production.

MAKING JPEG PLAY NICE WITH DEEP LEARNING

JPEG images are everywhere, but they don't work well with deep learning because you can't easily tweak them during training. This project reviews existing solutions and introduces a new method that fixes these problems, making JPEG images fully adjustable and compatible with deep learning systems.

A NEW DATASET TO SPOT FAKE CELEBRITY IMAGES

DeepFakeFace (DFF) is a collection of fake celebrity photos made with advanced technology designed to help us get better at telling real pictures from fake ones.

NEW STATE OF THE ART TEXT TO SPEECH MODEL

Coqui has released weights for its xtts model, which can clone voice parameters and synthesize in different languages.

THE BUMPY ROAD TOWARD GLOBAL AI GOVERNANCE

Researchers suggested a global AI ethics agreement a few years ago, emphasizing its feasibility despite philosophical differences. As the U.S.-China AI competition grows, challenges like misperceptions and language barriers persist in achieving global AI regulation.

Scientists have developed RETFound, an AI tool capable of diagnosing and predicting the risk of developing multiple health conditions, from ocular diseases to heart failure to Parkinson’s disease, from just retinal images.

STABILITY AI DEBUTS STABLE AUDIO

Stability AI has released Stable Audio, a text-to-audio generator that provides anyone with the ability to use simple text prompts to generate short audio clips.

GOOGLE GIVES PUBLIC DATA PROJECT AN AI MAKEOVER

Google has launched an AI-driven interface for Data Commons, simplifying data access and analysis through natural language queries.

COPILOT

Salesforce has unveiled Einstein Copilot, a conversational AI for more intuitive user queries. Enhanced with a "trust layer" for security and linked to Salesforce's Data Cloud, it aims to reduce AI inaccuracies while acknowledging limitations.

DO LANGUAGE MODELS REALLY UNDERSTAND THE WORLD?

Researchers looked into how well pretrained language models understand the basic building blocks and rules of the world, like 'a dog is an animal' or 'an apple is a fruit.' They found that these programs are decent at remembering some of this foundational knowledge, but aren't perfect at reasoning with it.