Search the archive

CODE FOR BAYESIAN FLOW NETWORKS

Bayesian Flow Networks are an exciting new architecture and training algorithm. This is a simple, unofficial, reproduction of the paper. The goal is to build a codebase that can scale to GPT-2 sized models.

TXTAI

txtai is an all-in-one embeddings database for semantic search, LLM orchestration, and language model workflows.

ANTI-HYPE LLM READING LIST

It's hard to know what is real in the current AI kerfuffle. This list contains foundational papers, interesting open questions, and a guide for gaining deeper insights into the space.

YOU PROBABLY DON’T NEED TO FINE TUNE AN LLM

Fine-tuning is often not necessary for most LLM applications. It may be better to use few-shot prompting or retrieval-augmented generation (RAG) instead. Few-shot prompting involves providing the LLM with examples of the desired output, while RAG involves querying a vector database for information that the LLM was not trained on.

GOOGLE’S PROJECT IDX

An AI-enabled browser-based development environment for building full-stack web and multiplatform apps.

AIH

Talk with AI chatbots directly in your terminal.

FORMER GOOGLE BRAIN AND STABILITY HEAD OF RESEARCH LAUNCHES AI COMPANY

Tokyo-based Sakana is a new AI research lab founded to discover novel foundation models based on nature. It’s not yet clear if they have raised funding, but the team and website are exciting.

META’S AI AGENTS LEARN TO MOVE BY COPYING TODDLERS

Meta AI has developed a platform called MyoSuite that can simulate the human body's musculoskeletal system and train AI agents to control simulated arms and legs with human-level dexterity and agility.

TOYOTA, PONY.AI PLAN TO MASS PRODUCE ROBOTAXIS IN CHINA

MAKING 3D PICTURES BETTER: A NEW WAY TO SPOT KEY POINTS

Keypoint detection is like finding specific dots in pictures to create a 3D image. This research, called DeDoDe, finds these dots in a new, better way, ensuring they match up in different views and improving 3D image creation.

IMPROVING EYE IMAGE ANALYSIS: A NEW WAY TO CREATE RETINAL IMAGES

Experts use detailed pictures of the inside of the eye when studying eye diseases, but creating these images is hard and time-consuming. This paper presents a new method (DDPM) and a dataset (ReTree) that makes creating and analyzing these images easier and faster than before.

A SMARTER WAY TO BUILD CONVOLUTION KERNELS

KernelWarehouse offers a new approach to dynamic convolution in computer vision by breaking down and reassembling kernels in a more efficient way. This method allows for more flexibility and power, using fewer parameters by cleverly sharing and mixing predefined parts among layers.

RIFT

Rift is open-source infrastructure for AI-native development environments, acting as an AI-native language server for your personal AI software engineer.

DEEPEVAL

DeepEval provides a Pythonic way to run offline evaluations on your LLM pipelines so you can launch comfortably into production.

THE RACE TO CAPTURE VALUE: CLOUD LESSONS FOR THE AI ERA

FINTECH IN AI: FIVE LESSONS FOR FOUNDERS

POTENTIAL NYT LAWSUIT COULD FORCE OPENAI TO WIPE CHATGPT

LANGCHAIN IN THE REAL WORLD

OPENAI ACQUIRES DIGITAL PRODUCT COMPANY GLOBAL ILLUMINATION

OpenAI acquired a team that has worked on a variety of products over the years including early functionality at Facebook, Youtube, and Riot. The team will help improve ChatGPT. This comes off the heels of several key product leaders leaving OpenAI.

NEW CONDA PACKAGE MANAGER PIXI

Dependency management is one of the hardest parts of working in Python. Conda aims to solve that issue, but the underlying package manager in Conda hasn't seen much love. Pixi is a new manager that is, of course, written in Rust.

GOOGLE DEVELOPING AI LIFE COACH

DeepMind, Google’s artificial intelligence research lab, has been testing new technology that could effectively turn generative AI into a personal life coach.

ENHANCING CONTINUAL LEARNING THROUGH RICH REPRESENTATIONS

A new approach named L3DMC boosts the performance of continual learning by preserving prior knowledge and enabling better adaptation to new concepts. It accomplishes this by combining the benefits of different fixed-curvature spaces and operating in a higher-dimensional space.

OPEN CHALLENGES IN LLM RESEARCH

10 open challenges in LLM research by the amazing Chip Huyen. The first two, hallucinations and context length, get discussed a lot today. But the most interesting may be multimodality, architecture improvements, and alternatives to GPUs. There are lots of smart folks working on these challenges, so we expect to see several of these improved upon in the near term.

USING LLMS TO ENHANCE PROBLEM-SOLVING

This research introduces Solo Performance Prompting (SPP), a method that uses multiple personas in LLMs to simulate cognitive synergy, a collaborative process that enhances problem-solving. By using several finely-tuned personas, SPP enables LLMs to tackle tasks that require deep domain knowledge and complex reasoning.

IMPROVING HOW COMPUTERS SEE AND UNDERSTAND IMAGES

Sometimes computers see the same object multiple times in one picture. This project developed a method that helps the computer recognize and focus on a single complete version of the object.

USE GPT MODEL TO ENHANCE COMMIT MESSAGES

Commit messages are vital in software development, but many are lacking in detail. This paper introduces a shift from simply generating commit messages to completing them using past commit history. The paper introduces a new dataset, CommitChronicle, and shows that using history can improve the quality of messages. It highlights the fact that the GPT-3.5-turbo model, while not always outperforming, has the potential for generating longer, detailed messages.

ALIGNING TINY STORIES MODELS

Reinforcement learning is probably the best tool for aligning large language models today (and even vision models). However, it is challenging when you need a bunch of human data to train a reward model. What if you can just “use prompting”? This is another great example of a trend of RLCF - C for computational.

THE MATHEMATICS OF TRAINING LLMS

An extremely deep dive into the viral Transformers Math 101 article and high-performance distributed training for Transformers-based architectures.

DOES AI JUST SUCK?

AI is extremely hyped, but there are multiple areas where it still needs a lot of work.

AI AND THE MEMORY WALL

The rise of LLaMA.cpp was surprising to many. Why can an ordinary MacBook run state-of-the-art models only a bit slower than a mighty A100? The answer is that when you're only doing a single batch, one token at a time, memory bandwidth is the bottleneck and the gap between the M2's memory bandwidth and the A100's is not that large.

GOOGLE UPGRADES SEARCH GENERATION EXPERIENCE

Google is rolling out a few new updates to its Search Generative Experience, the company’s AI-powered Search. These updates include automatic term definitions, programming assistance, and AI-generated summaries.

HUGGING FACE

IBM and Hugging Face are releasing IBM's geospatial foundation model, crafted using NASA satellite data, to the public. Built through a collaboration with NASA, this model aims to democratize AI's role in climate science and Earth studies, addressing the challenges of handling large environmental datasets. The model has already shown a 15% improvement in satellite data analysis and can be adapted for various tasks like tracking deforestation and monitoring greenhouse gasses.

BAYESIAN FLOW NETWORKS

Unifying the modeling of discrete or continuous data is a challenge. More so when attempting to maintain a probabilistic likelihood interpretation of the output. This new paper showcases a new class of generative models with a pile of interesting properties such as no forward pass, gradient-based sample guidance for discrete data, and competitive language modeling likelihood.

INSTRUCTION TUNING LARGE CODE MODELS

Most code models are instruction tuned on GPT-4 distillation data. This new work shows how to avoid that, introduces an extended human eval dataset, and gets 46.7 on the original human eval. While this is impressive, it showcases the extreme gap between closed and open code models.

ENHANCED VIRTUAL TRY-ON WITH DIFFUSION MODELS

Virtual try-on tasks, where clothes from one image are transferred to another, are challenging, especially at higher resolutions. Many methods use Generative Adversarial Networks. The solution presented in this paper, DCI-VTON, uses a diffusion model combined with an exemplar-based inpainting and a warping module. This combination ensures that clothing details are preserved and produces realistic virtual try-on images.

ADVANCED FOOD IMAGE SEGMENTATION

While the Segment Anything Model (SAM) excels in segmentation tasks across various datasets, it misses specific class details for masks. FoodSAM overcomes this. It combines SAM's strengths with new features to segment food images more effectively, even identifying individual food items and other objects in scenes.

OPENDREAM

OpenDream brings layering, non-destructive editing, portability, and easy-to-write extensions to your Stable Diffusion workflow.

A GUIDE TO DESIGNING AND SHIPPING AI DEVELOPER TOOLS

This article outlines ten key design principles for effective AI developer tools, drawing from GitHub’s experience designing, launching, and extending Copilot. They encompass leveraging natural language, addressing developer pain points, gathering meaningful feedback, acknowledging biases, and prioritizing customization, accessibility, performance, and continuous improvement in AI models.

CAN CHATGPT BECOME A CONTENT MODERATOR?

ChatGPT creator OpenAI has been using its most advanced large language model to enforce the company’s content policies.

HUMAN CREATIVITY AMONGST AI DOMINANCE

More marketers and content producers are turning to AI for efficiency. Over time, this will lead to a level of genericism in the market. The antidote will be authentic human production.

KNERON TO RELEASE AI CHIP THIS YEAR

Kneron, an AI edge company, announced that it will ship its new neural processing unit chips by the end of the year. The KL730 chips are specifically designed for machine learning and AI applications and they are said to be 3 to 4 times more energy efficient than previous chips.

TEACH A COMPUTER TO GHOTI

Quick read on why AI is on track to disrupt. A little cheeky too.

AMAZON IS ROLLING OUT AN AI FEATURE THAT SUMMARIZES PRODUCT REVIEWS

Amazon is rolling out a new generative AI feature that is designed to help shoppers determine at a glance what other customers said about a product before they spend time reading through individual reviews.

REGULATORS AGREE TO EXPANSION OF DRIVERLESS CAR SERVICES IN SAN FRANCISCO

Autonomous Cruise and Waymo taxis have been buzzing around the city for some time, and now they can offer round-the-clock rides for pay.

THE AI RULES THAT US POLICYMAKERS ARE CONSIDERING, EXPLAINED

The US government is getting serious about regulating AI, with proposals focused on new rules, oversight institutions, funding for research, and expanding the expert workforce. However, the area remains slippery as the technology rapidly evolves. Key players like OpenAI and Microsoft have endorsed new government licensing and auditing for advanced AI models, but the specifics remain contested.