AI Roundup 10/26-> 11/02/2023

OpenAI supercharges ChatGPT and eliminates a swarm of startups, the world's smallest multimodal model, and a tiny coding language model outperforming GPT-3.5

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.

---

⭐️ Heads-up: A special AI Tidbits edition covering OpenAI’s anticipated DevDay will be sent on Monday.

Highlights

1. OpenAI's ChatGPT now supports all of its modes in one conversation: Browsing, Advanced Data Analysis, and DALL-E (VentureBeat)

   

2. NousResearch releases Obsidian - the world's smallest multimodal model (3B) that is capable of running on consumer devices (Twitter)

   

3. US President Biden issues an executive order targeting Foundation Models companies with new AI safety and security standards, requiring companies to share safety test results before public deployment (TechCrunch)

   

4. Google DeepMind introduces the next generation of AlphaFold, extending molecular predictions beyond proteins and significantly improving biomolecule structure predictions (DeepMind)

   

5. Hugging Face releases Zephyr - a series of Mistral-based chat models with comparable performance to Anthropic's Claude 2 on AlpacaEval (Hugging Face)

   

6. Boston Dynamics leverages ChatGPT and additional AI tools to transform Spot, their robot, into an engaging talking tour guide, showcasing AI's potential to enhance robotic interactions and user experiences (Company blog)

   

Research

1. Researchers at Microsoft unveil CodeFusion - a model for diverse and accurate code generation across languages outperforming in top-3 and top-5 accuracies with a fraction of the size (75M parameters)

2. Hugging Face unveils Distil-Whisper - a compact and 6x faster version of OpenAI's Whisper with fewer hallucinations

   

3. Phind releases a new 16k context model that beats GPT-4 at coding at GPT-3.5-like speed

   

4. Perplexity releases two new language models, pplx-7b-chat and pplx-70b-chat, that substantially outperform Llama 2 according to human evaluators

   

5. Researchers propose JudgeLM - a fine-tuned LLM to efficiently and effectively evaluate other LLMs in open-ended benchmarks

   

6. Nvidia and UT Austin develop MimicGen - a system that synthesizes large-scale robot learning datasets from a small number of human demonstrations (50K across 18 tasks from only 200 human demonstrations)

   

7. Researchers introduce Learning from Mistakes (LeMa) - a method that fine-tunes LLMs using error-correction pairs from GPT-4, enhancing their performance in solving mathematical reasoning tasks

   

8. Researchers propose ControlLLM - a novel framework enabling LLMs to utilize multimodal tools for solving complex real-world tasks, addressing challenges in tool invocation and scheduling

   

9. Researchers propose a novel denoising network for diffusion models, reducing parameters by 75% and Floating Point Operations by 30%, accelerating inference by up to 70%

10. Microsoft presents MM-VID - a system leveraging GPT-4V and specialized tools for enhanced video understanding, transforming long-form videos into textual scripts for advanced analysis

    

11. Researchers propose SEINE - a model generating long videos with smooth transitions between scenes using textual descriptions

    

12. Google and Stanford propose a 3D-aware diffusion model, ZeroNVS, for single-image novel view synthesis in 'in-the-wild' scenes

    

13. Together AI open sources RedPajama-Data-v2 - a 30 trillion tokens dataset for LLM training

14. UC San Diego assesses GPT-4 in a public online Turing Test, achieving a 41% pass rate, outperforming ELIZA and GPT-3.5

    

15. Waymo releases Waymax – a lightweight simulator developed for autonomous driving research

    

16. NYU, ETH Zurich, and Boston University showcase that LLMs can discern truth from falsehood by modeling a 'truthful persona'

    

17. Researchers critically evaluate the performance of GPT-4V in medical applications, specifically in answering visual questions, concluding against its use in real-world diagnostics

Cool tools

1. Misgif - put your face in your favorite GIFs

2. Wonder3D - turn images into 3D assets using Cross-Domain Diffusion

   

   Misgif

Open-source

1. XAgent - AI agents designed for intricate task-solving, outperforming AutoGPT and GPT-4 across tasks and benchmarks

2. JudgeLM - an open-sourced LLM judge for evaluating LLM-generated answers

XAgent

Last week’s AI Tidbits

AI Roundup 10/19-> 10/26/2023

Reach AI builders, researchers, and entrepreneurs by advertising in AI Tidbits

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