Search the archive

THE MOBILE REVOLUTION VS. THE AI REVOLUTION

MICROSOFT CEO TESTIFIES ABOUT LIMITS OF AI AT GOOGLE ANTITRUST TRIAL

Microsoft CEO Satya Nadella testified in the U.S. antitrust case against Google, claiming Google's dominance in internet search is maintained through unfair tactics, including default agreements with Apple. Despite investments in AI and new features for Bing, Microsoft struggles to break Google's market stronghold. Nadella hinted that even with AI advancements, changing default habits might remain challenging.

INDEPENDENT COMPANY

DEEP DIVE: RUNNING SDXL IN JAX ON TPU V5

SDXL is a state of the art image generation model. It is powerful, but slow. Google and HuggingFace showcase just how fast it can be on the new TPU chips in Google's research language Jax.

REASONING

The recent work on registers for ViTs showed that adding a [reg] token fixes image attention map spikes. This work shows that adding a series of learnable [pause] tokens prepended to the sequence improves language model reasoning (decoder only, up to 1B) by as much as 18%. The intuition is that this adds a number of learnable and manipulatable vectors to store important global information to be used while generating the sequence. In general, adding tokens has scaled well past 13B parameters, but it's unclear how well this works at frontier model scale.

KNOWLEDGE EDITING IN LLMS

Think of LLMs like big, intricate puzzles with pieces of knowledge. This paper looks at what happens if we try to change some of those pieces - it could either mess up the whole puzzle or make it unclear.

MINIGPT-5: A NEW WAY TO GENERATE IMAGES AND TEXT TOGETHER

Imagine a model that not only understands and writes text but can also create matching pictures. This new model, MiniGPT-5, uses special techniques called "generative vokens" to do that.

DO LLMS WORK WELL TOGETHER LIKE HUMANS DO?

This paper studies how LLMs might work together in groups, just like people in a society. Researchers have made computer "societies" with different personalities and watched them team up on tasks, finding they sometimes act like humans - like following the majority or debating - and some teams work better than others.

DISTRIBUTED SHAMPOO OPTIMIZER IN PYTORCH

PROMPT ENGINEERING FOR CLAUDE'S LONG CONTEXT WINDOW

IS AI A PLATFORM SHIFT?

A platform shift alters the foundational layer of application development, influencing company dominance, architectures, and business models. Given its potential to change distribution and expand possibilities, AI is poised to become a significant platform shift, challenging existing structures and introducing new paradigms.

YOU CAN NOW USE THE DALL-E 3 AI IMAGE GENERATOR INSIDE BING CHAT

REKA LAUNCHES THEIR FLAGSHIP MULTIMODAL MODELS

Yasa-1 is an audio, video, and text model from the creators of the extremely popular T5 language model. It boasts impressive performance across a number of tasks. Yasa-1 is currently in private preview.

LINKEDIN GOES BIG ON NEW AI TOOLS

LinkedIn is launching a series of AI-based features to enhance its job hunting, marketing, and sales offerings. These include an improved version of its Recruiter talent sourcing platform with integrated AI, an AI-driven LinkedIn Learning coach, and a new AI tool for marketing campaigns.

Despite open-source AI democratization, over 60% of generative AI job postings from July 2022-2023 were in 10 major tech hubs. Interventions and expanded research programs are needed to diversify AI's geographic footprint.

GENERALIST ROBOT POLICY

A robot's policy is the guiding algorithm that determines how it acts. Typically these are robot and task specific. This makes it hard for robotics to adapt. This new work from Google shows a single (55B parameter) policy that can perform 160k+ different tasks.

MAKING MULTI-MODAL LLMS WORK FOR DIFFERENT TASKS

Multi-modal Large Language Models excel in many tasks but often struggle to pinpoint specific objects in images. Researchers have developed a way to improve this by turning object locations into text and training the model with special tasks.

EFFICIENT LARGE-SCALE 3D SCENE RECONSTRUCTION

Rebuilding large 3D scenes, especially when some data is missing, is crucial for self-driving cars. This new method, PC-NeRF, combines two modules to optimize the reconstruction at multiple levels, ensuring detailed 3D models even with incomplete data.

BETTER IMAGE RESTORATION WITH VISION-LANGUAGE MODELS

Vision-language models like CLIP work well for many tasks but struggle with image restoration. This project introduces DA-CLIP, a model that adjusts how CLIP works to handle these tasks better, leading to clearer image reconstructions.

ONNXSTREAM

A program that enables you to run Stable Diffusion 1.5 on a Raspberry Pi Zero 2.

LARGE WORLD MODELS FOR SELF DRIVING

Most self driving systems have been large combinations of rules-based, expert systems, and learned modules. This work by Wayve shows a 9B joint video, action, and text model that performs extremely well in diverse scenarios.

IS DATA STILL A MOAT?

Generative AI is reshaping the traditional value placed on vast data collections. As large language models can be fine-tuned with minimal data and even generate synthetic datasets, the uniqueness and importance of proprietary data may be waning.

AI IS GETTING BETTER AT HURRICANE FORECASTING

Hurricane Lee's path was predicted by AI-based models from Nvidia, Huawei, and DeepMind. These AI models, which analyze patterns from historical atmospheric data, have shown comparable or sometimes better performance in storm tracking than conventional models, delivering rapid results in minutes instead of hours, and might play a significant role in official forecasts in the coming years.

THE NATIONAL SECURITY AGENCY IS STARTING AN AI SECURITY CENTER

The National Security Agency is starting an artificial intelligence security center in order to keep pace with the rapid integration of AI into U.S. defense and intelligence systems.

NETWORK

Cloudflare unveils "Workers AI," a serverless AI inference platform on its global GPU network. Workers AI features popular open-source models and collaborates with Hugging Face.

HUMANE REVEALS FIRST AI DEVICE - THE AI PIN

Humane Inc unveiled the "Humane Ai Pin", a screenless wearable designed for AI integration, at Paris Fashion Week in collaboration with Coperni. The device offers AI-powered optical recognition, laser-projected displays, and prioritizes user privacy without requiring smartphone pairing.

REWIND PENDANT

Rewind Pendant is a wearable that captures what you say and hear in the real world and then transcribes, encrypts, and stores it entirely locally on your phone.

KOSMOS-2.5

The UniLM group at Microsoft has done some great work in the past few years around natural language. With the suite of Kosmos models, they’ve recently moved into images. This specific instantiation is for reading text-intensive documents from an image and generating the text or markdown for that document. It is similar to the recent Meta work for academic OCR.

EFFICIENT VIDEO MODEL TRAINING

Training top-notch video models usually need huge resources, often beyond what academia can access. Researchers have found a way to train these models using just one machine with eight standard GPUs in a day.

TEACHING IMAGE-RECOGNIZING MODEL THROUGH CONVERSATION

Researchers have developed a new way to improve vision-language models, which are models that understand both text and images, using just natural language prompts.

STREAMINGLLM

Language models are limited by their context length. The context length is usually limited by compute hardware and clever algorithmic updates. This is an algorithmic update that streams the tokens through the attention mechanism to allow theoretically infinite context window size. Usually, these claims fall over at scale, but this one seems robust as it can work on existing pre-trained models without fine-tuning. Will it make the forgotten middle issue worse though?

IMPROVING OPTICAL FLOW WITH GAUSSIAN ATTENTION

Optical flow helps figure out how things move in images. This study introduces new techniques using Gaussian Attention to focus on finer details and match them better, resulting in a model named GAFlow.

OPEN-SOURCE EVALUATION SUITE FOR LARGE MODELS

OpenCompass is a free tool designed to test large models quickly and efficiently.

8 WAYS SAAS COMPANIES CAN USE AI TO GET AHEAD

This blog post features several AI SaaS use cases for various departments, highlighting ways the technology can be used to gain a competitive edge.

AI, HARDWARE, AND VIRTUAL REALITY

AI, combined with hardware, is going to enable groundbreaking forms of virtual and augmented reality. Here’s a breakdown of the latest news in each of those domains.

HOW MUCH CAN ARTISTS MAKE FROM GENERATIVE AI?

Despite some companies proposing "creators' funds," the specifics of how and how much artists will be paid remain unclear, with many details still under development or undisclosed.

META QUIETLY UNVEILS LLAMA 2 LONG

SHOPIFY AI FOR PRODUCT PHOTO CREATION

AI AND JOBS

VISION TRANSFORMERS NEED REGISTERS

One of the coolest and simplest vision papers in recent weeks. Vision Transformers use "useless" pixel values as places to store global information. This makes attention maps uninterpretable. However, if you add a simple [reg] token to the vocabulary, the model will use it and won't store information in pixel values.

TRANSFORMER-VQ: EFFICIENT LINEAR-TIME ATTENTION

This study presents Transformer-VQ, a new design of transformer that processes attention faster thanks to unique vector-based keys and caching.

IDENTIFYING REAL VS FAKE FACES

This study uses visual and language-based tools to improve how systems recognize real vs. fake faces.

LOCAL LLMS MADE EASY

With all the new models coming out, it can be somewhat challenging to find a way to pull and run them. It's also hard to remix and remake them easily. This neat little tool, Ollama, makes it easy to test the latest models.

CREATING VIDEOS FROM SOUND

SAPIENTML

SapientML is an AutoML technology that can learn from a corpus of existing datasets and their human-written pipelines and efficiently generate a high-quality pipeline for a predictive task on a new dataset.

SNOWFLAKE, PINECONE, AND AI21 LABS