AI Roundup 11/02-> 11/09/2023

OpenAI’s packed DevDay, xAI's new language model and prompting tool, Runway's text-to-video gets an upgrade, and a new emotional prompt boosts LLMs’ accuracy

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.

---

Highlights

1. OpenAI announces a host of product releases in its inaugural dev conference: GPT-4 Turbo with a 128k-context window, a new text-to-speech model, and an autonomous agents API
   More in our coverage earlier this week:

OpenAI DevDay - a pivotal moment for AI

2. xAI announces Grok-1 - a language model outperforming GPT-3.5 and Llama 2 having real-time direct access to tweets (Company blog)

   

3. 01.ai openly releases Yi-34B - the first Chinese model to top Hugging Face's LLM Leaderboard supporting a 200k context window (Company blog)

   

4. Samsung unveils Gauss, a new generative AI model capable of tasks like image generation and document summarization, and is set to be incorporated into smartphones and other devices (CNET)

5. DeepSeek open sources DeepSeek Coder - a 1.3B and 33B coding models, substantially outperforming previous models (Project website)

   

6. Amazon is set to unveil Olympus - a new AI model, aiming to rival industry leaders in conversational AI (Analytics India Magazine)

7. xAI introduces PromptIDE, an advanced development environment for refining AI prompts, complete with a Python SDK for sophisticated prompt design and a full suite of analytics for insight into model responses (Company blog)

8. Stability AI is set to simplify 3D content creation with Stable 3D, an automated process that generates textured 3D objects from image or text prompts (Stability AI)

9. Runway ML has updated its Gen-2 video generator to improve video quality and consistency, and to support higher-resolution video generation from images (VentureBeat)

10. Deci AI releases YOLO-NAS Pose - a SOTA pose estimation model, outperforming previous models with 38% lower latency (Company blog)

Research

1. Researchers find that incorporating emotional cues into prompts such as "This is very important to my career" enhances LLMs' performance, with experiments showing an 8% improvement in rule-based tasks and a 115% boost in creativity benchmarks (Paper)

2. Researchers release LLaVA-Interactive - pioneering efficient multimodal AI interaction by combining visual conversation, image segmentation, and generation without extra training (Paper)

3. Researchers propose I2VGen-XL, a new approach to video creation that uses images and text to generate high-quality, accurate, and clear videos (Paper)

2. Researchers introduce mPLUG-Owl2, a versatile multi-modal LLM that leverages modality collaboration to enhance performance in both text and multi-modal tasks (Paper)

3. Meta and CUHK create a system that captures and models real spaces for virtual reality with high precision, using a custom camera rig and advanced image processing for real-time, high-resolution rendering (Paper's Github repo)

   

4. AllenNLP releases MADLAD-400 - a 7.2 trillion tokens multilingual (>400 languages) filtered dataset (Hugging Face)

5. Researchers introduce GLaMM - the first model to combine language responses with object segmentation, offering flexible interactions with images and text for detailed visual-language understanding (Paper's Github repo)

7. Researchers propose OtterHD-8B, a high-resolution multimodal model that can understand and interpret high-quality images with great detail, demonstrating superior performance (Paper)

8. UC Berkeley and Google researchers present Idempotent Generative Networks (IGNs), a new AI method that generates matching images from different inputs in one step (Paper)

9. Researchers propose EmerNeRF, a new technology that teaches itself to understand and improve images of moving scenes, greatly enhancing image clarity and accuracy in estimating space occupancy by ~37% (Paper's Github repo)

10. Google AI develops COSP and USP methods for self-prompting in LLMs, allowing AI to improve its own question-answering abilities and perform complex reasoning tasks autonomously, broadening its application spectrum. (Google AI)

11. Researchers from Stanford and UC Berkeley propose Rule-following Language Evaluation Scenarios (RuLES), a programmatic framework for measuring and improving the rule-following ability of Large Language Models (LLMs) (Paper)

12. Google DeepMind is set to improve robot capabilities with RoboVQA, a new data collection approach that uses real-world interactions (Paper)

13. DeepMind proposes a framework for classifying the progress of Artificial General Intelligence (AGI) by introducing levels of performance, generality, and autonomy, akin to the levels used for autonomous driving (Paper)

14. Researchers introduce S-LoRA, a system that improves the efficiency of using multiple AI models for different tasks, boosting processing speed by four times and supporting thousands of models on a single GPU (Paper)

15. Meta proposes FLAP, a self-guided approach that improves how computers understand the relationship between sound and text, setting new records for correctly identifying sounds based on descriptions (Paper)

16. CMU and Tsinghua develop RoboGen - automating the expansion of robotic abilities through a self-generating cycle, proposing and learning new tasks without extensive human guidance (Github repo)

17. Georgia Tech, Berkeley, and Microsoft introduce PASTA - a technique enabling LLMs to recognize and prioritize text with user-specified emphasis, improving instruction-following accuracy by 22% (Paper)

18. NousResearch publishes YaRN - a new compute-efficient method enabling longer context windows and powering the release of two new variants of Llama 2 and Mistral 128k (Github repo)

Announcements

1. Figma introduces the new FigJam AI to automate mundane design tasks (Company blog)

2. YouTube is testing AI-powered features that organize large comment sections into themes and provide a conversational tool for deeper content understanding (Google AI)

Cool tools

1. ProPainter - video inpainting made easy

Open-source

1. OpenGPTs - an open-source implementation of OpenAI's GPTs from LangChain

2. Whisper large v3 - a new open-source version of OpenAI's speech2text model

3. realtime-bakllava - llama.cpp with BakLLaVA model describes what it sees

Last week’s AI Tidbits

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

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.