Search the archive

A NEW APPROACH FOR AUDIO-VISUAL SEGMENTATION

This paper introduces the Explicit Conditional Multimodal Variational Auto-Encoder (ECMVAE) for better identification of sound sources in videos.

IMPROVING TRAFFIC PREDICTIONS WITH GRAPH MIXFORMER

Researchers have developed the Spatio-Temporal Graph Mixformer (STGM) network, a new model for predicting traffic patterns.

BUILDING A MORE POWERFUL 3D MODEL

Researchers have introduced a new way to train 3D models, making them more efficient and versatile. Using a unique approach that blends 3D and 2D information, it shows significant improvements in tasks ranging from creating realistic images to recognizing 3D objects.

SUDOLANG

LLMS FOR DUMMIES

LLMs are all the rage these days, but do you really know how they work? This article provides a high-level overview of the plumbing behind LLMs and then looks at four frameworks for LLM applications that might emerge in the startup ecosystem.

OPENAI’S FINANCIAL ASCENT

THE AI RESEARCH JOB MARKET CRAP SHOW

CHARACTER.AI INTRODUCES GROUP CHATS

AI SAFETY GUARDRAILS ARE EASILY THWARTED

STATE OF AI 2023

A great summary of the language modeling industry. There are good discussions here on safety, capabilities, chips, and more. The predictions around AI filmmaking, capital costs, and AI-fueled IPOs are interesting.

OPENAI UNVEILS MAJOR UPDATES TO BOOST AI DEVELOPMENT EFFICIENCY

GOOGLE’S AI-POWERED SEARCH EXPERIENCE CAN NOW GENERATE IMAGES

SPEED UP COLD BOOT TIMES FOR LLAMA 2

TEXT EMBEDDINGS REVEAL ALMOST AS MUCH AS TEXT

Using iterative re-embedding and treating embedding inversion as a conditional generation problem allows attackers to successfully recreate 92% of the embedded texts just with access to the dense vectors.

ENHANCED SEGMENTATION WITH SAM-OCTA

Researchers have improved the segmentation of specific targets in optical coherence tomography angiography (OCTA) images with a method called SAM-OCTA.

SCALECRAFTER: HIGH-RESOLUTION IMAGE GENERATION

Researchers studied how to produce high-quality images at large sizes using pre-trained diffusion models trained on smaller images. They discovered issues with repeated patterns in the bigger images and pinpointed the problem to the limited perception of the model's structures.

IMPROVING TEXT TO IMAGE GENERATION WITH LLMS

Researchers are improving how we communicate with text-to-image models, like Midjourney and Stable Diffusion, which currently require complex prompts for generating images. They've introduced "interactive text to image", allowing people to naturally interact with Large Language Models for better image creation and related tasks.

SOFTBANK

HUGGING FACE MISTRAL MODEL

OPENAI IS TOO CHEAP TO BEAT

OPENAI HAS QUIETLY CHANGED ITS ‘CORE VALUES’

MICROSOFT AUTOGEN

Framework to enable agent system deployment.

BIG TECH STRUGGLES TO TURN AI HYPE INTO PROFITS

Tech giants are investing in costly generative AI tools, leading to pricier software upgrades. Firms are exploring ways to balance innovation with profitability while anticipating future cost reductions.

CLEANLAB RAISES $30M TO HELP BUILD TRUSTWORTHY LLMS

Starting as an open sourced package at MIT, CleanLab has now grown into an enterprise-grade no-code tool for curating high quality data for industry ML pipelines.

DEEP DIVE INTO MULTIMODAL FRONTIER MODELS

An excellent dive into the current research activity around multimodal language models.

PAPER FOR MISTRAL 7B MODEL

The Mistral 7B model was released via a tweet. It beats many models twice its size. This paper explores some of the technical contributions and details. No discussion of the pre-training dataset, but good insights into the architecture.

IMPROVING MEMORY IN CONTINUAL TEXT CLASSIFICATION

This study explores the challenges faced in "continual learning", where a model tries to keep learning new things without forgetting old ones, especially in categorizing texts. It introduces a new method called InfoCL, which uses special techniques to help the model remember better and reduce mistakes.

MAKING AI-GENERATED SENTENCES SOUND NATURAL

Researchers found that the sequence in which ideas or 'concepts' are presented to models that generate sentences affects how good the sentences turn out to be. Among various programs tested, BART-large did the best job when given ideas in a certain order, but surprisingly, bigger isn't always better.

OPEN LLMETRY

OpenLLMetry is a set of extensions built on top of OpenTelemetry that gives you complete observability over your LLM application.

Researchers have developed a new technique that combines sound information and language data to correct mistakes in automatic speech recognition systems.

CHATGPT HISTORY EXPORT TO MD

A CASE FOR GPT-4’S CAPACITY TO REASON

Although GPT-4 lacks consciousness and long-term memory, there is good evidence that it is able to reason and comprehend concepts.

UNIVERSAL, INTERACTIVE, REAL WORLD SIMULATOR

What if your simulator was generative AI? It would work to zero shot transfer robotic policies! This work from DeepMind and colleagues tackles the challenge that RL usually can't generalize past the training environment. If we just make everything the training environment then there is no need for extrapolating.

$500 MILLION INVESTMENT IN AI STARTUP ANTHROPIC

Prosecutors want to prevent former FTX cofounder Sam Bankman-Fried’s attorneys from mentioning his $500M investment in AI startup Anthropic during trial, deeming it irrelevant. Meanwhile, Anthropic, seen as a competitor to OpenAI, is seeking funding that could propel its valuation to $20B-$30B, after recently securing a $1.25B commitment from Amazon.

GOOGLE INSIDERS ARE SLAMMING BARD FOR NOT BEING WORTH IT

GROWTH IS SLOWING

ADOBE BEEFS UP PHOTOSHOP WITH AI TOOLS

Adobe has released the official web version of Photoshop, equipped with AI tools powered by Firefly.

MICROSOFT ANNOUNCES NEW AI TOOLS TO ASSIST DOCTORS

Microsoft introduced new AI tools in Microsoft Fabric and Azure AI to help healthcare organizations consolidate and interpret vast medical data.

US WARNS EU’S LANDMARK AI POLICY WILL ONLY BENEFIT BIG TECH

The US warns the EU that its proposed AI regulations could favor well-resourced companies and hurt smaller ones. Washington's analysis criticizes the EU's focus on AI model development rather than the risk of model use, suggesting it could deter investment and AI R&D in the EU. The debate continues, aiming for alignment and shared AI objectives.

LEVERAGE LLMS FOR ENHANCED VIDEO UNDERSTANDING

Researchers have introduced FAVOR, a new approach that allows large language models to understand videos by finely blending audio and visual details at the frame level.

BOOSTING MATHEMATICAL REASONING OF LLMS

Researchers investigate the impact of data augmentation techniques on improving the math reasoning capabilities of large language models. They create a new dataset, AugGSM8K, by enhancing queries from an existing dataset and developing a model named MuggleMath.

HUMAN TO ROBOT MOTION RETARGETING

Researchers have been able to transfer human motion onto robots by using a deep learning encoder. This is usually done with specialized hardware that has to be calibrated to the robot. Combining this with video-based motion capture could be a cool future project!

IMPROVING VISUAL GENERATION WITH A NEW VIDEO TOKENIZATION

This study introduces MAGVIT-v2, a video tokenizer that efficiently translates image and video inputs into tokens for Large Language Models. With MAGVIT-v2, LLMs surpass the performance of diffusion models in visual generation tasks.

SPEEDING UP DIFFUSION MODELS FOR TEXT GENERATION

Researchers introduce a new technique for diffusion models used in text generation. By using a soft absorbing state, they improved how the model handled discrete text changes, resulting in faster learning and quicker sampling.

TESTING LLMS IN THE COMPETITIVE WORLD OF AUCTIONS

Researchers have created AucArena, a simulation platform designed to test Large Language Models in auction settings, which are dynamic and require strategic thinking. Initial tests show that, with the right prompts, these models can perform impressively in auctions, showing skills like budgeting and long-term planning.

AI HYPE IS BUILT ON FLAWED TEST SCORES

A growing number of researchers want to overhaul the way LLMs are assessed, calling for more rigorous and exhaustive evaluation, as they believe the practice of scoring machines on human tests is misguided.

8 WAYS SAAS COMPANIES CAN USE AI TO GET AHEAD

From task automation to writing code, AI is adding leverage to how we analyze customer data and build products. The key takeaway here is that AI is enabling people to do a lot more, and if you’re not using it, you’re likely going to get outpaced by your competition.