Search the archive

CODE RELEASED FOR MOE PEFT

Cohere’s research arm For AI developed parameter efficient fine-tuning for Mixture-of-Experts models. It recently released the code for the community to build on.

USING CHATGPT FOR IMAGE CLASSIFICATION

This study shows that image classification can be improved by using large language models to organize classes into hierarchies, addressing biases in pretrained models like CLIP that confuse similar categories.

SIFTING THROUGH THE NOISE: HOW AI WILL IMPACT INFORMATION OVERLOAD

No doubt, you consume a lot of content on a weekly basis. Articles, blog posts, tweets, this newsletter. With AI, the rise of user-generated content is going to increase exponentially. The systems you build to defend your attention will be more important than ever.

AI’S ROLE IN RUNNING SOFTWARE COMPANIES

From helping you to write code, checking the quality of that code, and adding repeatability to your content generation, AI is going to impact how we think about leverage in the daily operations of our software companies. This article explores the common areas you should be thinking about adding AI to your operations.

TALK LLAMA

Talk with a LLaMA AI in terminal.

THE BEATLES JUST RELEASED THEIR FINAL SONG WITH THE HELP OF AI

Over five decades post their disbanding, The Beatles have released "Now and Then," their last song, with AI technology revitalizing John Lennon's voice.

TIM COOK ON GENERATIVE AI: “WE’RE INVESTING QUITE A BIT”

CEO Tim Cook acknowledged significant investments in AI without detailing how they plan to monetize generative AI during Apple's Q4 2023 earnings call.

JULIUS (PRODUCT)

Julius is an AI data scientist for analyzing data and files.

STABILITY AI TEASES RELEASE OF 3D GENERATION, SKY REPLACEMENT, AND MORE

3D asset generation in the .obj format is the next big innovation for creatives working with AI. Stability AI is working to make that a possibility. This article is light on details, but it is exciting to see what is around the corner.

MICROSOFT PUSHES THE BOUNDARIES OF SMALL AI MODELS WITH BIG BREAKTHROUGH

Microsoft Research has enhanced its smaller language model, Phi 1.5, with multimodal capabilities, enabling it to interpret images similarly to OpenAI's larger GPT-4 model but at a lower computational cost.

BIDEN’S AI ORDER MAY HAVE WIDE IMPACT FOR STARTUPS

Biden has gone all in on an executive order that, while lacking teeth, gives a good sense of where the administration is coming from when it comes to AI, and from that, we can glean potential impacts on the startup ecosystem.

NEW TECHNIQUES EFFICIENTLY ACCELERATE SPARSE TENSORS FOR AI MODELS

MIT and NVIDIA researchers developed techniques to accelerate sparse tensor processing, enhancing performance and energy efficiency in generative AI tasks.

LLMS LEARNING FROM THEIR MISTAKES

The study introduces 'Learning from Mistakes' (LeMa), a method that teaches large language models to solve math problems by learning from errors, similar to a human student improving through correcting their mistakes.

NEURAL NET HAS HUMAN-LIKE ABILITY TO GENERALIZE LANGUAGE

Researchers have developed a neural network that can generalize language similarly to humans, outperforming models like ChatGPT.

LANGCHAIN TEMPLATES

The easiest and fastest way to build a production-ready LLM application. These templates serve as a set of reference architectures for a wide variety of popular LLM use cases.

HOW-TO GUIDE TO GET STARTED WITH LLAMA

This guide provides information and resources to help you set up Llama including how to access the model, hosting, and how-to and integration guides

FINE TUNE YOUR OWN MISTRAL FOR $1 OF COMPUTE

Jupyter notebook example of how to fine-tune Mistral's 7B model on your own dataset.

SEGMENT ANYTHING AND METACLIP (HUGGING FACE SPACE)

Open vocabulary segmentation is a somewhat new and exciting topic in computer vision. This work combines powerful base models with strong language capabilities together to segment based on text input.

WHAT AI MEANS FOR YOUR PRODUCT STRATEGY (1 HOUR PODCAST)

Paul Adam, a long-term CP at Intercom, discusses the ways AI is going to impact the world (and your tech stack). There’s quite a bit of fear-setting in this conversation, so if this doesn’t inspire you to look into how you should be using AI, I don’t know what will.

POLITICIANS COMMIT TO COLLABORATE ON AI SAFETY

Global stakeholders discussed collaborative efforts to address AI risks through the Bletchley Declaration at the AI Safety Summit in Bletchley Park, England, aiming for a human-centric, trustworthy AI deployment. U.S. Secretary of Commerce Gina Raimondo announced a new AI safety institute within the Department of Commerce, emphasizing global policy alignment on AI safety, while political leaders from various countries stressed inclusivity and responsibility in navigating AI advancements.

value with LLMs. Register now for AI Forward!

4 WAYS AI IS CHANGING MARKETING

Your ability to aggregate and understand customer data is improving, allowing you to have a better understanding of your customers, predict their future behavior more accurately, offer scaled customer support, and curate messaging that resonates more accurately.

NEW AWS SERVICE LETS CUSTOMERS RENT NVIDIA GPUS FOR QUICK AI PROJECTS

AWS introduced Amazon Elastic Compute Cloud (EC2) Capacity Blocks for ML to facilitate access to GPU resources for running large language models, allowing customers to reserve GPU instances for defined time periods.

KEYWORD GENERATOR (PRODUCT LAUNCH)

Use AI to generate keywords for SEO.

MAGICPOST (PRODUCT)

Use AI to craft better LinkedIn posts.

ISOMORPHIC TEASES THE NEXT VERSION OF ALPHAFOLD

AlphaFold is an amazing model for science. The new work by Isomorphic Labs (a spin off from Google) extends it beyond just proteins and improves overall performance.

ARTISTS LOSE FIRST ROUND OF COPYRIGHT CASE

A federal judge dismissed most claims in a landmark lawsuit where artists accused generative AI art generators of unauthorized image use, pinpointing deficiencies in copyright infringement accusations against Midjourney and DeviantArt.

OPENAI'S ILYA SUTSKEVER ON PREVENTING ROGUE SUPERINTELLIGENCE

OpenAI's co-founder and chief scientist Ilya Sutskever has shifted his focus from building the next model like GPT to figuring out how to avoid future artificial superintelligence from behaving undesirably. He believes the reality of AI surpassing human intelligence is imminent and anticipates technologies that will allow humans and AI to merge. Sutskever and his team at OpenAI are actively working on "superalignment", a set of procedures to control future AI technology.

IMPROVING THE RELIABILITY OF DIFFUSION MODELS

Researchers have found ways to use diffusion models to make long-term plans, but these plans aren't always realistic or safe. This study introduces a method to fix those unreliable plans, using a tool called the "restoration gap".

A DEEP DIVE INTO MEDICAL VISUAL QUESTION ANSWERING

The study assesses how well GPT-4 with Vision (GPT-4V) answers questions related to medical images from various sources, such as X-rays and CT scans.

A MASK PROPAGATION TECHNIQUE FOR VIDEO SEMANTIC SEGMENTATION

The study introduces MPVSS, a method for segmenting video content that reduces computational load by focusing on key frames and then predicting masks for other frames based on these key ones.

EFFICIENT VIDEO UNDERSTANDING

This project introduces TESTA, a method designed to speed up the process of understanding long videos by combining similar frames and patches. Using TESTA, researchers managed to greatly reduce the computational load and improve performance in matching paragraphs to videos and answering questions about long videos.

IMPROVING MULTI-MODAL LLMS

Researchers dived deep into the visual encoders used in Multi-modal Large Language Models (MLLMs) and discovered that certain features from CLIP and DINO models are especially effective for detailed visual tasks. They then introduced COMM, a strategy that combines the strengths of both models.

A NEW IMAGE DENOISING APPROACH

The study introduces PUCA, a new self-supervised denoising approach that maintains J-invariance, a critical aspect for effective denoising.

HACKERS ARE WEAPONIZING AI TO IMPROVE PHISHING

AI is poised to dramatically improve the success of phishing attacks, which is already used in 90% of cyberattacks.

SEISMIC WAVES OF GEN Z BEHAVIOR

This article discusses the influence of Gen Z on various industries and how their behavior is causing seismic waves in the market. The author emphasizes the importance of understanding Gen Z's habits, preferences, and values, as this generation drives significant changes in consumer behavior. By examining the social, technological, and economic aspects of Gen Z's lives, businesses can better anticipate their needs and adapt their offerings to cater to this influential demographic.

DEATH OF A FLYWHEEL

Death of a Flywheel discusses the decline of business flywheels, which represent self-sustaining momentum that drives company growth, due to factors including market saturation, evolving customer preferences, competitive threats, regulatory changes, and misaligned incentives. The article emphasizes the importance of recognizing and addressing these threats to ensure the continued health of a company's flywheel.

LABELBOX IS HIRING A FOUNDATION MODELS LEAD ENGINEER (SAN FRANCISCO, $170K-$215K + EQUITY)

Labelbox is a data-centric AI platform for building intelligent applications. As the Foundation Models Lead, you will lead research and development for applying foundation models to real world problems for customers like Walmart and Adobe. Learn more [17].

ANDREW NG SAYS BIG TECH IS LYING ABOUT AI RISKS TO SHUT DOWN COMPETITION

AI expert and Google Brain co-founder Andrew Ng suggests that Big Tech companies are amplifying fears about AI risks to stifle competition and prompt strict regulations that could hinder the open-source community.

ALIBABA LAUNCHES UPGRADED AI MODEL

Alibaba has unveiled an enhanced AI model, Tongyi Qianwen 2.0, that aims to compete with global tech behemoths like Amazon and Microsoft, particularly in generative AI applications similar to ChatGPT.

PHOT.AI (PRODUCT LAUNCH)

Generative AI tools for photo editing and design.

GPT-3.5-TURBO IS 20B PARAMETER

Researchers wrote that gpt-3.5-turbo (the model behind ChatGPT) is a 20B parameter model in a paper by a Microsoft research group that studies diffusion models for code. The original model was said to be 175B parameters, so there could have been retraining or distillation.

ANNOUNCING THE OPEN EMPATHIC PROJECT

The group behind Stable Diffusion has launched the Open Empathic project to foster emotion-detecting capabilities in open-source AI systems for more authentic AI-human interactions.

BIDEN RELEASES AI EXECUTIVE ORDER

President Joe Biden signed an executive order outlining eight goals to regulate generative AI, focusing on safety, privacy, equity, consumer protection, innovation, and responsible government use, with agencies like the NIST tasked to develop AI standards and large AI model developers required to share safety test results.

USING EMBEDDINGS IN COMPUTER VISION

Embeddings are machine-readable lists of numbers that represent some information about a piece of data. They are increasingly obtained by running the data through a neural network and looking at the activations of the model. This post pushes that idea forward with CLIP on a computer vision task and explores how to use dimensionality reduction to reason about the learned embeddings.

REPRODUCING OPENAI'S FIRST RLHF REPOSITORY

The HuggingFace team reproduces a 2019 repository from OpenAI that introduced some of the first RLHF ideas.

AMD TRAINING ROUND 2

A post from MosaicML on using AMD GPUs for training large language models.

PERSONALIZED IMAGE CLUSTERING

This paper introduces Image Clustering Conditioned on Text Criteria (IC⁠TC), a method that leverages vision-language models to sort images based on user-given text descriptions.

AZURE OPENAI LOGGER

This project aims to create a simple and easy to deploy solution to add observability to your Azure OpenAI instance. The approach adds an API Management instance as a proxy for your existing Azure OpenAI service, and funnels logs/requests/responses to an Application Insights instance. It also saves a prebuilt query to a workbook for easy access to logs.