Search the archive

CHATGPT’S API MAKES MOST TEXT GENERATING AI OBSOLETE

CHATGPT IS NOW AVAILABLE IN AZURE OPENAI SERVICE

LARGE LANGUAGE MODELS ARE HAVING THEIR STABLE DIFFUSION MOMENT

This article dives into how Large Language Models are exploding in innovation and popularity similar to the explosion in generative AI following Stable Diffusions release in August 2022.

TOOLBOX

This GitHub Repo is a mix of machine learning guides and tools.

PAN-VARIANT COVID VACCINE DEVELOPED WITH ML

FOR CRM

RUN STABLE DIFFUSION IN THE BROWSER WITH WEBGPU

This project brings stable diffusion models to web browsers. Everything runs inside the browser with no need of server support. This is the first example of the pre-trained model running on a client’s GPU from the browser. It still has some rough edges and sometimes results in browser crashing, but it is an exciting step towards a more modular future.

HOW GOOGLE BECAME CAUTIOUS OF AI AND GAVE MICROSOFT AN OPENING

More than two years ago, a pair of Google researchers started pushing the company to release a powerful chatbot. The conversational AI they had developed could confidently debate philosophy and banter about its favorite TV shows, while improvising puns about cows and horses. The researchers told colleagues that chatbots like theirs, supercharged by recent advances in AI, would revolutionize the way people searched the internet and interacted with computers, according to people who heard the remarks. They pushed Google to give access to the chatbot to outside researchers, tried to get it integrated into the Google Assistant virtual helper and later asked for Google to make a public demo available. Google executives rebuffed them at multiple turns.

LARGE LANGUAGE MODELS DO IN-CONTEXT LEARNING DIFFERENTLY

As model size increases large language models seem to exhibit some amazing emergent behavior. This paper explores the phenomenon of “in-context learning”. They find that smaller models rely on their semantic priors and ignore examples that are wrong while larger models learn from the new examples.

MODELS

Researchers from Microsoft introduce Visual ChatGPT, a system that combines ChatGPT language model with different visual models, enabling users to interact with ChatGPT using both language and images. The system allows users to provide feedback and ask for corrected results. It injects visual model information into ChatGPT and considers models with multiple inputs/outputs and those that require visual feedback.

X-AVATAR: EXPRESSIVE HUMAN AVATARS

X-Avatar is a new avatar model that captures the full expressiveness of digital humans. It can be learned from 3D scans or RGB-D data and models bodies, hands, facial expressions, and appearance holistically. The method uses a learned forward skinning module and novel sampling strategies. The paper also contributes a new dataset called X-Humans. The proposed method outperforms strong baselines on the animation task.

VISUAL PERCEPTION WITH PRE-TRAINED DIFFUSION MODELS

VPD (Visual Perception with Pre-trained Diffusion Models) is a framework that leverages the high-level and low-level knowledge of a pre-trained text-to-image diffusion model to downstream visual perception tasks.

DIFFUSION MODELS FOR ROBOTICS

If you diffuse a motion path, robotic control becomes significantly more effective! The diffusion process improves the ability of robot manipulators to accomplish challenging tasks.

EMERGENT ABILITIES OF LARGE LANGUAGE MODELS

THE AI FUTURE OF FRONTEND DEVELOPMENT

LEETCODE PROBLEM

The creator ran ChatGPT against every single challenge and recorded the results. The author explores everything he’s learned in exploring AI and ChatGPT to solve this specific niche of problems.

LLAMA-DL

Llama-dl is a high-speed download of Facebook's LLaMA model.

TSAI

Tsai is an open-source deep learning package.

SALESFORCE $250MM GENERATIVE AI FUND

STABILITY AI ACQUIRES CLIPDROP

STATE OF COMPETITIVE ML 2022

PALM-E

The term “positive transfer” is somewhat of a holy grail for large scale models. It means that when you train on two modalities, you get better performance than you would have training on just one or the other. This paper from Google shows that training on robot controls, video, and text shows positive transfer between domains. This is hypothesized to be a function of model scale, and this model is an impressive 562 billion parameters.

PRISMER: A VISION-LANGUAGE MODEL WITH MULTI-MODAL EXPERTS

The article discusses the high computational and data requirements of large pre-trained models used in vision-language learning, which require massive amounts of image-text data to be trained. The proposed alternative approach involves using separate sub-networks or "experts" for specific tasks, allowing for the use of domain-specific data and architectures that would not be feasible with a single large network. The proposed model, Prismer, is a visually conditioned autoregressive text generation model that utilizes powerful vision-only and language-only models along with modality-specific vision experts for open-ended vision-language reasoning tasks. The expert models are pre-trained and frozen, and connected through lightweight trainable components, comprising only 20% of total network parameters, leading to improved training efficiency and scalability.

DAVINCI FUNCTIONS

There are common things you want to do with large language models such as text-divinci-003 such as write a list, answer a question, or verify a statement. This repo provides some simple tools for doing those tasks. It is a cool step towards building an abstraction layer on top of these models.

REPRESENTATION FROM 2D SUPERVISION

The paper introduces "Nerflets," a local neural radiance field that represents a 3D scene efficiently and with structure awareness. Nerflets are composed of multiple neural radiance fields that maintain their own spatial position, orientation, and extent to contribute to panoptic, density, and radiance reconstructions. By using only photometric and inferred panoptic image supervision, Nerflets allow the extraction of panoptic and photometric renderings from arbitrary views, and enable tasks rare for NeRFs, such as 3D panoptic segmentation and interactive editing. The experiments show that Nerflets fit and approximate the scene more efficiently than traditional global NeRFs.

SGD IN SQL

Looking to reduce MLOps complexity? Just write everything in your database! This fun little post takes this idea to the limit and writes gradient calculations in SQL.

SCALE AI: WHY DATA WILL POWER THE AI REVOLUTION

CONSIDERATIONS FOR AI-NATIVE STARTUPS

DON’T FEAR AN AI-INDUCED JOBS APOCALYPSE JUST YET

This article makes the case that because the west suffers from too little automation, not too much, it is unlikely that AI causes mass unemployment.

VCS SKEPTICAL OVER AI

CATEGORY THEORY MACHINE LEARNING

STABLEDIFFUSION-PLUGIN

GOOGLE’S UNIVERSAL SPEECH MODEL

AI IS BEING USED TO DETECT BREAST CANCER

MICROSOFT LAUNCHES COPILOT

Microsoft, having brought artificial intelligence to its battle with Google over search, is now turning to the latest AI technology to catch up with rivals in the corporate applications market such as Oracle, Salesforce, and SAP. Microsoft’s new AI “copilot” will help answer customer calls, summarize sales meetings, and generate marketing pitches.

HUMAN MOTION DIFFUSION AS A GENERATIVE PRIOR

The paper proposes a method to address the low availability of data in motion generation tasks by using a pre-trained diffusion-based model as a generative prior. The method includes a zero-shot setting that generates long animations with controlled transitions, a few-shot setting that generates two-person interaction, and a fine-tuning setting that enables fine-grained control and editing. The proposed method outperforms state-of-the-art approaches in quality scores and interaction levels in a user study.

IN-CONTEXT INSTRUCTION LEARNING

The paper presents a new approach to instruction learning called In-Context Instruction Learning (ICIL) that significantly improves the zero-shot task generalization performance of Large Language Models (LLMs). While previous approaches to instruction learning have been predominantly fine-tuning based, ICIL uses a single fixed prompt to evaluate all tasks, resulting in a 9.3% improvement in performance for the most powerful instruction-fine-tuned baseline. ICIL is shown to be complementary to instruction-based fine-tuning, making it a promising approach for improving zero-shot task generalization for LLMs.

STYLIZE YOUR FACE IN ONLY ONE-SHOT

The paper presents StyO, a new model for one-shot face stylization that combines content and style attributes from input images using disentanglement and recombination strategies. The model uses latent diffusion models and consists of two modules: Identifier Disentanglement Learner (IDL) and Fine-grained Content Controller (FCC). Evaluation results show that StyO outperforms current state-of-the-art methods in producing high-quality stylized face images.

FASTER MODEL INIT ON GPU

Loading language models to GPUs can be painfully slow, potentially minutes long. This is because they tend to load all the weights onto the CPU before transferring them to the GPU. Instead of this process, you can write an initialization context manager which puts the weights directly on the GPU - leading to pretty dramatic savings.

AN OPEN-SOURCE FRAMEWORK FOR IN-CONTEXT LEARNING

The authors introduce OpenICL, an open-source toolkit that facilitates In-Context Learning (ICL) and Large Language Model (LLM) evaluation. ICL is a new paradigm for LLM evaluation that adapts pre-trained models to unseen tasks without parameter updates. OpenICL provides a flexible architecture that users can easily combine with different components to meet their requirements. The toolkit also includes various state-of-the-art retrieval and inference methods, making it an efficient and robust tool for LLM evaluation on a wide range of Natural Language Processing (NLP) tasks.

WORDS AS IMAGE - SEMANTIC TYPOGRAPHY

Semantic Typography is where letters in a word are shaped in such a way to reflect the meaning of the word. So the A in Paris might look like the Eiffel tower. This is typically a very labor intensive process. However, using a differentiable renderer and a strong text-to-image prior, it is now feasible to automate some of this process. A delightful piece of work with fun visual examples included.

THE WALUIGI EFFECT

RECREATING FLAMINGO - RETROSPECTIVE

THOUSANDS SCAMMED BY AI VOICES

A DEEP DIVE INTO THE AI OBJECTIVE INSTITUTE

Inspired by digital-privacy pioneer Peter Eckersley, the AI Objective Institute is working toward guiding AI toward “human flourishing”

LLAMA INT8

LLaMA int8 is a fork of the LLaMA code that runs LLaMA-13B comfortably within 24 GiB of RAM.

BRAVE SEARCH LAUNCHES AN AI-POWERED SUMMARIZATION FEATURE

100 MOST CITED AI PAPERS IN 2022