Search the archive

ANTHROPIC SCALES TO 100K CONTEXT TOKENS

STABILITY AI TEXT TO ANIMATION

SDK released to allow users to create animations using stable diffusion models. You can generate these animations unconditionally, conditioned in an image, or conditioned on a video. The results are lovely, are somewhat computationally expensive, and have a very distinctive style.

BARD UPDATES

CREATING BETTER VIDEO CONTENT USING AI AND SKETCHES

This paper describes a new way to make video content using artificial intelligence (AI), based on sketches and text. The method, which combines two technologies called Text-to-Video Zero and ControlNet, can create high-quality, consistent videos that closely match what the user intended, as demonstrated by various experiments.

DIFFUSION-BASED 3D FACIAL RECONSTRUCTION FROM A SINGLE IMAGE

This paper introduces a method that uses diffusion models, which have been successful in creating images, to reconstruct 3D facial features from just one image. The process involves using a dataset of facial textures, simulating various lighting conditions, and then using the diffusion model to fill in missing parts of the texture and unknown reflective properties, resulting in a more accurate and consistent 3D face.

AN IN-DEPTH LOOK AT TEXT-TO-3D: THE NEW FRONTIER IN GENERATIVE AI

This paper is a comprehensive survey of the emerging field of text-to-3D, a part of generative AI that turns written descriptions into 3D models. It introduces different types of 3D data, foundational technologies, and how they're combined in recent works, as well as how text-to-3D is used in various applications like creating avatars and generating scenes.

HUGGINGFACE CHAT UI

HuggingFace has open sourced the codebase powering the HuggingChat app, a competitor to ChatGPT.

VIDEOCHAT: A NEW SYSTEM FOR UNDERSTANDING VIDEOS BETTER

This repo introduces VideoChat, a system that makes sense of videos by combining artificial intelligence techniques for understanding videos and language. The researchers also created a large set of videos with detailed descriptions and conversations to help train the system to be better at understanding the sequence of events and cause-effect relationships in videos.

EVA

EVA is designed for supporting database applications that operate on both structured (tables, feature vectors) and unstructured data (videos, podcasts, PDFs, etc.) using deep learning models.

EU CONSIDERING TOUGHER AI RULES

European lawmakers came a step closer to passing new rules regulating artificial intelligence tools such as ChatGPT, following a crunch vote on Thursday. The European Union's highly anticipated AI Act is set to be the first comprehensive legislation governing the technology, with new rules around the use of facial recognition, biometric surveillance, and other AI applications.

WILL AI BECOME THE NEW MCKINSEY?

In this article, Ted Chong dives into the question of AI’s future in the economy, specifically its potential negative impacts on the workforce.

MANAGEMENT

META ANNOUNCES GENERATIVE AI FOR ADVERTISERS

REGULATION

FLOWISE

Drag & drop UI to build your customized LLM flow using LangchainJS.

HUGGINGFACE AGENTS

GOOGLE RELEASES MUSICLM

GOOGLE'S AI CHATBOT BARD GETS MULTILINGUAL WITH NEW UPGRADES

ENHANCING LANGUAGE MODELS WITH ALGORITHMS FOR SUPERIOR PERFORMANCE

Recent research shows that incorporating algorithms into large pre-trained language models (LLMs) can significantly improve their capabilities, without the need for fine-tuning. An example of evidence-supported question-answering demonstrates a 6.4% improvement over the standard baseline, suggesting the potential advantages of this innovative, cost-effective method.

BUILDING AFFORDABLE BUSINESSES WITH LARGE LANGUAGE MODELS

MODELS

This paper systematically evaluates methods for generating images from diffusion models and their benefits for data augmentation in downstream tasks. The study reveals that personalizing diffusion models towards target data is effective, but using the training data of the diffusion model alone through a simple nearest neighbor retrieval procedure results in stronger downstream performance, highlighting the potential and limitations of diffusion models in data augmentation.

PRE-TRAINING WITHOUT ATTENTION

Attention is a way for a language model to perform computation over words. It is pretty expensive. What if we could use 1d convolution state space models for BERT training? It turns out to perform quite well, all the code, weights, etc is available here in Jax.

CHAIN-OF-THOUGHT-HUB

Chain-Of-Thought-Hub benchmarks large language models' complex reasoning ability with chain-of-thought prompting.

LLM CHAIN

LLM Chain is a powerful rust crate for building chains in large language models allowing you to summarize text and complete complex tasks.

THE LEVERAGE OF LLMS FOR INDIVIDUALS

THE AI TAKEOVER OF GOOGLE SEARCH STARTS NOW

SAM ALTMAN TO TESTIFY BEFORE CONGRESS

OpenAI CEO Sam Altman will testify to Congress for the first-time next week, the latest sign that policymakers in Washington are ratcheting up scrutiny of artificial intelligence as the technology booms in Silicon Valley.

INDIA’S RELIGIOUS CHATBOTS ARE CONDONING VIOLENCE

PRIVATEGPT

PrivateGPT allows you to Interact privately with your documents using the power of GPT.

TURNING A FOUNDATION MODEL INTO A CHAT ASSISTANTS

Star coder is an excellent new model from the big code project that can write 86 different programming languages. This post explores how to prompt the model to participate in conversations, how to use chat markup language, and even how to tune these models. Lots of interesting tidbits here for groups interested in taking an open source model and fine tuning it for their own chat-based use cases.

CLAUDE’S CONSTITUTION

META OPEN-SOURCES MULTISENSORY AI MODEL

ONE EMBEDDING SPACE TO RULE THEM ALL

LANGUAGE MODELS CAN EXPLAIN NEURONS IN OTHER LANGUAGE MODELS

CHATBOTS

This research introduces InternChat (iChat), a tool that combines powerful chatbots like ChatGPT with non-verbal cues like pointing movements, allowing users to directly manipulate images or videos. By adding gestures to language, iChat makes communication with chatbots more efficient and accurate, especially for tasks involving complex visual scenarios, making it a promising development for future interactive systems.

ENHANCING LANGUAGE MODELS WITH FEDERATED INSTRUCTION TUNING

The study introduces a new method, Federated Instruction Tuning (FedIT), to train powerful language models like ChatGPT and GPT-4, by learning from many different devices' data without violating privacy. This makes the models better at understanding and responding naturally in diverse situations, and the researchers have also created a tool, 'Shepherd', to help others use this approach easily.

FRUGALGPT: COST-EFFECTIVE APPROACHES FOR UTILIZING LLMS

This study examines the high costs associated with using popular language model APIs like GPT-4, ChatGPT, and J1-Jumbo, and suggests three strategies for reducing these costs: prompt adaptation, LLM approximation, and LLM cascade. The research introduces FrugalGPT, an efficient application of the LLM cascade strategy that can match the performance of top language models at a fraction of the cost or even improve their accuracy.

SUR-ADAPTER: AN UPGRADE FOR TEXT-TO-IMAGE MODELS

This study introduces a new way, called SUR-adapter, to help computer programs called diffusion models turn simple descriptions into detailed images. By teaching these models to understand stories better and using a new dataset of over 57,000 samples, the researchers made the models better at creating high-quality images from simple descriptions, making them easier and more fun to use.

MICROSOFT’S COPILOT IS GETTING MORE FEATURES

IBM UNVEILS WATSONX AI PLATFORM

IBM on Tuesday launched Watsonx, a new artificial intelligence and data platform to help companies integrate AI in their business.

WENDY’S TESTS AN AI CHATBOT THAT TAKES YOUR DRIVE-THRU ORDER

SPOTIFY HAS PURGED TENS OF THOUSANDS OF AI-GENERATED SONG

AUTOPR

Fix issues with AI-generated pull requests, powered by ChatGPT.

MILLION DOLLAR AI FINE-TUNING CHALLENGE

Think of this like Kaggle but for fine tuning a language model. We want to improve the safety of these models, and so we are relying on the community now to see if we can make progress. There are compute partnerships available and a reward model that you can use. The competition starts on June 10th.

GOOGLE TO ANNOUNCE PALM 2

FAIR USE AND MODEL TRAINING, GETTY IMAGES VS. STABLE DIFFUSION

LEARNING DECISION TREES WITH GRADIENT DESCENT

Decision Trees are a type of artificial intelligence that make decisions based on data, but the current method of creating these trees can sometimes give us less than perfect results. This paper talks about a new method using something called gradient descent, which helps us to build better Decision Trees and get more accurate results for tasks where the AI needs to categorize things.

COMPOSE OBJECTS LOCALIZED WITH ATTRIBUTES

Compositional reasoning helps us understand complex visuals by combining objects and their features, but current AI models struggle with this. Researchers created a benchmark called Cola to test and improve these AI models, finding that a lightweight multi-modal adapter helps the models reason about images and language better.

ENHANCING WEBPAGE UNDERSTANDING WITH A NOVEL ATTENTION MECHANISM

Webpages offer a lot of information for AI tasks, but they haven't been used to their full potential. The new Wikipedia Webpage suite (WikiWeb2M) helps study multimodal webpage understanding by introducing 2 million pages and a special attention mechanism that improves performance on tasks like generating descriptions, summarizing sections, and creating contextual image captions.