AI Roundup 05/23 -> 05/30/2024

Mistral’s first coding LLM, an open-source model for finance, a few AI agent packages automating coding tests and essay writing, and new details about OpenAI’s ousting of Sam Altman

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

---

Overview

• ✨ Highlights (4 entries)

• Language Models (12 entries)

• Vision (7 entries)

• Audio (3 entries)

• Industry announcements (2 entries)

• Open-source Packages (4 entries)

Recent Deep Dives

Top 8 leaderboards to choose the right AI model for your task

12 techniques to reduce your LLM API bill and launch blazingly fast products

Harnessing research-backed prompting techniques for enhanced LLM performance

✨ Highlights

1. Mistral releases Codestral - a 32k context window coding model supporting over 80 programming languages and setting a new standard in performance and latency (Mistral)

   

2. Researchers release MiniCPM-Llama3-V 2.5 - an 8B multimodal model rivaling proprietary models in overall performance (e.g. GPT-4V, Gemini Pro), supporting multimodal conversation in over 30 languages and enabling on-device deployment thanks to its small size (Project website)

   

3. Meta releases a detailed guide to Vision-Language Models (VLMs), explaining their mechanics, training, and evaluation and exploring the challenges of extending VLMs to videos (Paper)

   

4. Helen Toner, the former OpenAI board member, sheds light on the reasons behind ousting Sam Altman last November (The Verge)

Language Models

1. Researchers unveil AutoCoder - the first coding model to outperform GPT-4 Turbo and GPT-4o on the Human Eval benchmark, achieving a pass@1 score of 90.9% and featuring an enhanced code interpreter with the capability to install external packages

   

2. AI4Finance Foundation unveils FinRobot - an open-source AI agent platform featuring financially specialized AI agents powered by LLMs

   

3. LLM360 open sources K2-65B - a commercially permissive LLM that outperforms Llama 3 70B on coding and text-generating tasks

   

4. Cohere open sources Aya 23 - a new family of multilingual LLMs (8B, 35B) that supports 23 languages and outperforms existing models in various linguistic tasks

   

5. Researchers unveil Zamba - a 7B SSM-transformer hybrid model, achieving competitive performance with leading models at similar scales, featuring a unique architecture combining a Mamba backbone and a shared attention module for faster inference and reduced memory usage

   

6. Researchers unveil HippoRAG - a novel retrieval framework that enhances LLMs' ability to integrate new information efficiently, outperforming state-of-the-art RAG methods by up to 20% while being 10x-30x faster and up to 13x cheaper

   

7. Hugging Face releases a new leaderboard for multilingual European LLMs

   

8. A new study across six countries showcases the adoption rate of LLM-powered chatbots, revealing that despite the hype, not many use ChatGPT regularly

   

9. Stanford shows that AI systems for legal research, such as LexisNexis and Thomson Reuters, provide incorrect information for one in six queries despite reducing errors compared to general AI models like GPT-4

   

10. AI2 proposes a method to internalize chain-of-thought steps in language models, enabling significant improvements in reasoning accuracy, including 99% accuracy on complex multiplications and over 50% on challenging benchmarks without intermediate outputs

    

11. Researchers address transformer's poor performance on arithmetic tasks by introducing position embeddings for each digit, achieving state-of-the-art accuracy on complex arithmetic problems (99% accuracy) up to 100 digits long

    

12. By creating a synthetic dataset from mathematical competition problems and fine-tuning the DeepSeekMath 7B model, researchers achieved groundbreaking accuracy in formal theorem proving, outperforming GPT-4 and successfully solving complex benchmarks

    

Vision

1. UW–Madison and Microsoft introduce Matryoshka Multimodal Models - a novel model that improves efficiency and performance in multimodal tasks by allowing adjustable visual granularity, significantly reducing token usage without sacrificing accuracy

   

2. Researchers present I2VEdit - a novel framework that extends image editing tools to videos, ensuring high-quality, temporally consistent outputs by propagating edits throughout the video using a pre-trained image-to-video model

   

3. Peking University presents ReVideo, enabling precise video editing by specifying both content and motion and achieving promising performance in various accurate video editing applications

   

4. Researchers introduce FaceAdapter - an efficient adapter designed for high-precision and high-fidelity face editing with pre-trained diffusion models, achieving comparable or superior performance to fully fine-tuned models

   

5. UPenn presents KnoBo - an innovative approach that enhances medical imaging models by integrating explicit medical knowledge from PubMed, achieving a 32.4% average improvement and demonstrating superior generalization and prediction performance

   

6. Researchers present Generative Camera Dolly - a framework that generates novel viewpoints from monocular videos by simulating camera movements, enhancing dynamic scene understanding, robotics perception, and interactive 3D experiences for virtual reality

   

7. Researchers present RoboCasa - a comprehensive simulation framework that employs generative AI to create detailed environments and extensive training tasks for robots, featuring over 2,500 3D assets and leveraging automated data generation techniques

   

Audio

1. Apple introduces Denoising LM - a scaled error correction model trained with synthetic data, achieving state-of-the-art speech recognition performance on Librispeech without external audio data

   

2. Researchers unveil an intelligent hearable system that allows selective focus on a target speaker in noisy environments, achieving a ~7 dB signal quality improvement without performance degradation

   

3. Researchers introduce Instruct-MusicGen - a groundbreaking text-to-music editing method that finetunes a pretrained MusicGen model with minimal additional parameters and training, achieving superior performance

Industry announcements

1. MavenAGI launches out of stealth with an OpenAI-powered automated customer support agent, reducing operational costs by 80%

2. Canva introduces a range of AI-powered features: Magic Design, Magic Grab, and Enhance Voice, leveraging AI to simplify and enhance the creation of images, videos, and text

Maven’s AI autonomously resolves 93% of customer inquiries

Open-source Packages

1. Storm - LLM-powered knowledge curation system that researches a topic and generates a full-length report with citations

2. Cover Agent - automatically generate qualified tests to enhance existing test suites

3. LlamaFS - a self-organizing file system with llama 3

4. Mistral-finetune - a light-weight codebase that enables memory-efficient and performant finetuning of Mistral's models

Plus >70 more open-source packages for AI engineers

Cover Agent automates test generation

Last week’s AI Tidbits roundup

AI Roundup 05/16 -> 05/23/2024

Reach 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.