Search the archive

AUTO EVALUATOR

WINDOW.AI

Window AI is a browser extension that lets you configure AI models in one place and use them on the web.

CAMEL

Camel explores the potential of building scalable techniques to facilitate autonomous cooperation among communicative agents and provide insight into their "cognitive" processes.

HOW TO USE CHATGPT IN YOUR PM WORK

AN EXAMPLE OF LLM PROMPTING ON PROGRAMMING

MICROSOFT REPORTEDLY WORKING ON AI CHIPS

JAPAN GOVERNMENT TO USE CHATGPT FOR FIRST TIME ON RED TAPE

Japan is using OpenAI’s ChatGPT to try and make its often opaque and complex government regulations easier to understand.

TURBOPILOT

TurboPilot is a self-hosted copilot clone which uses the library behind llama.cpp to run the 6 Billion Parameter Salesforce Codegen model in 4GiB of RAM.

TRUTH-SEEKING AI’

DATASET RELEASED TO REPRODUCE LLAMA TRAINING

The Llama models from Meta are powerful but unusable for many commercial applications due to licensing restrictions. However, because they’re so good, lots of folks have been building on top of them. It starts to get tricky when people go to deploy them, however, and often it is a complete impossibility. In the Llama paper they describe the training mix proportions which have now been reproduced as a 1.2 Trillion token dataset. Open models are coming soon.

AI READINESS REPORT

56%-72% of companies have accelerated their AI plans because of the recent generative AI progress. Many of these groups prefer to use open source technology, while OpenAI's APIs are the dominant closed models being used. The report outlines industries adopting AI technologies and good things to know to help cut through the hype.

WITH TEXT

Multimodal C4 (mmc4) is a publicly available dataset that augments the text-only c4 corpus with interleaved images, supporting in-context vision and language models like Flamingo. The dataset, created using a linear assignment algorithm with CLIP features, covers everyday topics and contains 103M documents, 585M images, and 43B English tokens, with 90% of images being topically relevant.

INSTRUCTION TUNING WITH GPT-4

Utilizing GPT-4-generated instruction-following data for LLaMA finetuning, researchers achieve superior zero-shot performance on new tasks in both English and Chinese, surpassing previous state-of-the-art models. This breakthrough is accompanied by comprehensive evaluation and reward model training, with the data and codebase made publicly available.

CEREBRAS-GPT: EFFICIENT, SCALABLE LANGUAGE MODELS WITH OPEN COMPUTE

Cerebras-GPT, a family of compute-optimal language models, demonstrates state-of-the-art training efficiency on pre-training and downstream objectives by leveraging advances in efficient pretraining, scaling, and Maximal Update Parameterization. As the first open and reproducible work comparing compute-optimal model scaling, pre-trained models and code are publicly released.

INPAINT ANYTHING

Inpaint Anything (IA) is a new mask-free image inpainting paradigm that combines various models to create a powerful and user-friendly solution for inpainting tasks. IA offers three main features: Remove Anything, Fill Anything, and Replace Anything, allowing users to edit images using clicks and text-based prompts seamlessly.

SOUNDINI: SOUND-GUIDED DIFFUSION FOR NATURAL VIDEO EDITING

MULTILINGUAL TEXT2IMAGE MODEL

A nice new text to image model that can handle multiple languages. The output is high quality as it was trained on LAION HighRes and an internal proprietary dataset of images. It doesn’t work with diffusers, but they have an online demo. It seems to have more parameters than stable diffusion but might be smaller than the new XL model in the works.

DATA SCIENTISTS CLONES HIS GROUP CHAT

DOLLY IS A “REALLY BIG DEAL”

HYPE GROWS OVER “AUTONOMOUS” AI AGENTS

AI FUNCTIONS

An easy-to-use implementation of AI functions using OpenAI's GPT-4 (or any other model version) to perform various tasks.

PRINCETON LAUNCHES CENTER FOR LANGUAGE AND INTELLIGENCE

Recent progress in large language models has been dominated by large industry players. However, many of the key ideas still come from academia. Princeton is hoping to continue contributions to language and intelligence with this new center. They’re hiring for research scientists, engineers, and postdoc scholars.

WORLD'S SIMPLEST TEXT-TO-IMAGE MODEL?

Laion just released the Paella image generation model. It uses a discrete, token based, latent space and a cross entropy denoising objective. One goal of this project is to make an image synthetics model that almost anyone can understand. Their sampling code is only 12 lines long!

GOOGLE RUSHING TO ADD AI TO SEARCH ENGINE

SCALING, EMERGENCE, AND REASONING IN LLMS

SO LONG HYPERPARAMETERS

What if we didn’t have to tune the pile of hyperparameters we do to make deep learning work? This is a pretty technical paper that presents Automatic Gradient Descent which is a new optimizer that takes into account network architectures and eliminates a set of previously required hyperparameters. They work through all the math thoroughly, and the resulting optimizer seems to be fairly simple. They don’t have it working yet for all network architectures, but that feels like only a matter of time.

DINOV2: LEARNING ROBUST VISUAL FEATURES WITHOUT SUPERVISION

Recent breakthroughs in natural language processing have paved the way for foundation models in computer vision, producing all-purpose visual features through self-supervised methods and large, diverse datasets. By utilizing an automatic pipeline for curated data and training a 1-billion-parameter ViT model, researchers have created smaller models that surpass existing all-purpose features in most benchmarks at image and pixel levels.

OPENAI CONSISTENCY MODELS CODE

We wrote recently about the new type of genitive model from OpenAI that is faster than diffusion and more stable than a GAN. They just released code and model checkpoints on a set of small experimental data sets. This model is a cool next step in generative image modeling, and now we can play around with it!

TRANSFORMER

Combining spectral and multi-headed attention layers in the novel SpectFormer architecture improves transformer performance for image recognition tasks. This approach achieves state-of-the-art accuracy on ImageNet-1K and consistently strong results in transfer learning, object detection, and instance segmentation tasks across various datasets.

COST TO TRAIN ANTHROPIC'S NEXT MODEL

Anthropic recently announced plans to spend a billion dollars on training a massive language model. They announced this plan before their strategic partnership with Google, which may offer cheaper compute via their TPU pods. In any case it seems like one of the largest expenses of this endeavor is actually AI researcher salaries, followed closely by compute costs.

91% OF ML MODELS DEGRADE OVER TIME

EUROPEAN PARLIAMENT PREPARES TOUGH AI MEASURES

The European parliament is preparing tough new measures over the use of artificial intelligence, including forcing chatbot makers to reveal if they use copyrighted material.

GPT5 IS NOT IN TRAINING YET

GENERATING CODE WITH TESTS

WOLVERINE

Wolverine uses GPT-4 to edit your scripts and tell you what went wrong.

BUILDING LLM APPLICATIONS

AWS BEDROCK

DREAMPOSE: FASHION IMAGE-TO-VIDEO SYNTHESIS VIA STABLE DIFFUSION

DATASET FOR ANIMATED KID’S DRAWINGS

VICUNA: AN OPEN-SOURCE CHATBOT IMPRESSING GPT-4 WITH 90% CHATGPT QUALITY

This paper introduces Vicuna-13B, an open-source chatbot trained by fine-tuning LLaMA on user-shared conversations. Vicuna-13B outperforms other models like LLaMA and Stanford Alpaca in over 90% of cases while achieving a quality of over 90% compared to OpenAI ChatGPT and Google Bard. The cost of training Vicuna-13B is around $300, and the code and demo are publicly available for non-commercial use.

VIDEO EDITING WITH VIDSTYLEODE

VidStyleODE combines StyleGAN and Neural-ODEs for disentangled video representation, enabling applications like appearance manipulation, motion manipulation, image animation, and video interpolation/extrapolation.

CONTINUAL DIFFUSION

ADAPTIVE HUMAN MATTING FOR DYNAMIC VIDEOS

AdaM, an Adaptive Matting framework for Dynamic Videos, tackles the limitations of trimap-free methods by simultaneously differentiating foregrounds from backgrounds and capturing alpha matte details in diverse, unstructured videos. Utilizing an interconnected encoder-decoder network and a transformer network, AdaM improves matting realism and temporal coherence, achieving best-in-class generalizability for complex real-world videos.

CHROMA’S HACKER IN RESIDENCE

On the tails of a recent fundraise, the vector database startup Chroma is looking to hire a hacker for 3 months to build cool AI applications that require long-term memory. Specifically, they hope this hacker can be mostly self driven and build something powerful! I expect to see more of these sorts of positions pop up, given the ease of MVP development on top of language models.

HOW AI COULD DISRUPT VIDEO GAMING

THE MOUNTING HUMAN AND ENVIRONMENTAL COSTS OF AI

META OPEN SOURCES ANIMATED DRAWINGS

EUROPEAN PRIVACY WATCHDOG CREATES CHATGPT TASK FORCE