AI Roundup 08/24-> 09/07/2023

TII releases the world's most powerful open-source LLM, an open-source alternative to OpenAI's Code Interpreter, the first Arabic LLM, and a new scalable approach to RLHF using AI

Welcome to the weekly edition of AI Tidbits, where I curate the firehose of AI research papers and tools so you won’t have to. Advertise with us

Highlights

1. TII releases Falcon 180B - a 3.5 trillion tokens LLM outperforming Llama 2 and on par with Google's PaLM 2 (Company blog)

2. Meta publishes a comprehensive paper on how it trained and evaluated Code Llama (Meta AI)

3. A developer released Open Interpreter - an open-source alternative to OpenAI's Code Interpreter that runs locally (Twitter)

4. Ideogram launches a free text-to-image generator capable of rendering high-quality text in images (VentureBeat)

5. WizardCoder 34B - a fine-tuned version of StarCoder and Code Llama outperforming ChatGPT, Claude 2, and previous SOTA open-source language models on coding tasks (Twitter)

6. Refact releases Refact Code LLM - a 1.6B coding LLM outperforming similar-sized coding models across 20 programming languages (Company blog)

7. Meta open-sources DINOv2, along with new vision models under a commercially permissible license and a new benchmark for evaluating the fairness of computer vision models (Meta AI)

   

8. UAE open sources Jais and Jais-chat - the first Arabic language models with 13B parameters (Paper)

9. DeepMind introduces a new tool to detect synthetically-generated images (DeepMind)

10. OpenAI releases a guide for teachers using ChatGPT including suggested prompts and its limitations (OpenAI)

12. OpenAI announces ChatGPT Enterprise for companies seeking unlimited higher-speed GPT-4 access, a 32k context window, and enterprise-grade security and privacy (OpenAI)

13. Clinical Camel-70B - a new open-source medical LLM for clinical research (Twitter)

Research

1. Meta AI introduces Nougat - a visual transformer model for OCR, aiming to convert scientific PDFs, including mathematical expressions, into markup language

2. Meta AI unveils CM3Leon - a groundbreaking multimodal model setting new standards in text-to-image synthesis with optimal efficiency

3. Researchers propose MVDream - a diffusion model that generates consistent multi-view images from text, enhancing 3D generation's stability

4. Researchers present LLaSM - a pioneering multimodal model that seamlessly combines speech and language, paving the way for more natural human-AI interactions

5. UCSD researchers unveil BLIVA - a vision language model that excels in deciphering images with embedded text, outperforming the baseline by up to 17%

6. Microsoft and Zhejiang University present StableVideo - a text-driven video editing tool that utilizes diffusion models to ensure consistent edits in videos

   

7. Researchers propose a masked 3D diffusion model for video outpainting

8. Researchers introduce SLiMe - a groundbreaking technique that leverages large vision-language models for image segmentation, using just a single annotated sample for precise granularity

9. Cornell and Tel Aviv University introduce Doppelgangers - a dataset and a method that aid in identifying images that look alike but represent different 3D structures

   

10. Stanford unveils Sequential Dexterity - an RL-based system that connects multiple hand policies for complex, long-term tasks

11. DeepMind unveils RoboTAP - a method using dense tracking that enables robots to master complex tasks like stacking and path-following based on quick and simple demonstrations

12. Researchers present Active Neural Mapping - a new method to efficiently navigate and understand unfamiliar environments

13. Google and Stanford introduce PhysObjects - a substantial dataset to train vision-language models in understanding the physical attributes of objects, enhancing their application in robotic manipulation.

14. Researchers develop CLIPN - a novel technique for recognizing unknown object classes for computer vision models

15. fast.ai surfaces a pattern showing that LLMs can memorize inputs using a single example

16. Google shows that RL from AI Feedback (RLAIF) performs on par with Reinforcement Learning from Human Feedback (RLHF), offering a potential solution to the scalability limitations of RLHF

17. Researchers show that by eliminating Transformer's Feed Forward Network (FFN) component and consolidating a single encoder FFN, the revised Transformer architecture achieves both enhanced accuracy and reduced latency compared to its predecessor

Cool tools

1. TokenFlow - edit videos using text prompts

2. MS Image2Video - turn images into short videos

3. Voicemod's Text-to-Sing API - turn a melody + text into a short song

4. Voice Cloning - mimic any voice character using only 20 seconds of voice audio

Open-source

1. VALL-E X: An open-source implementation of Microsoft's text2speech model VALL-E

2. OpenCopilot - your own product's AI copilot

Last week’s AI Tidbits

AI Roundup 08/17-> 08/24/2023

Advertise to AI builders, researchers, and entrepreneurs by partnering with AI Tidbits

If you find AI Tidbits valuable, share it with a friend and consider showing your support.