AI Roundup 05/30 -> 06/06/2024

Mamba-2 sets speed records in language modeling, Stability’s new open audio model, Google's LLMs surpass adults in mental state reasoning, and an agent that completes tasks on phones autonomously

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 (6 entries)

• Language Models (6 entries)

• Vision (10 entries)

• Audio (1 entry)

• 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. ElevenLabs releases Sound Effects, turning text into rich sounds (Company blog)

2. Perplexity releases Pages - an AI-driven tool that generates customizable webpages from user prompts, aimed at simplifying report creation for educators and researchers (The Verge)

   

3. Mistral introduces model customization on its platform, enabling efficient fine-tuning of AI models to meet specific user needs with reduced costs and expertise

   

4. Princeton and CMU release Mamba-2, a significantly faster language model that maintains competitive performance in language modeling, and introduce the SSD framework, connecting SSMs and Transformers (Company blog)

   

5. Stability AI releases Stable Audio Open - an open-source model that generates high-quality audio samples from text descriptions (Stability AI)

6. OpenAI current and former employees propose enhanced whistleblower protections to address AI safety concerns, advocating for anonymous reporting channels and criticizing non-disparagement clauses (Blog)

   

7. Madrona publishes a comprehensive piece covering the current limitations in the AI agents space and discussing the different startups that are poised to address them (Madrona blog)

   

Language Models

1. Google examines higher-order theory of mind in LLMs, finding GPT-4 and Flan-PaLM achieve adult-level performance, with GPT-4 surpassing adults on 6th-order inferences (recursive reasoning about the mental states of others)

   

2. Researchers introduce Zyda - a comprehensive 1.3 trillion-token dataset that excels in language modeling by integrating and refining existing datasets, setting new performance standards

   

3. Hugging Face researchers release FineWeb-Edu, demonstrating the crucial role of data quality in outperforming unfiltered datasets in AI training

   

4. Databricks and MIT demonstrate that small language models (125M) can prune large datasets to significantly boost larger models' performance and efficiency, achieving significant improvements in downstream tasks

5. DeepMind presents an innovative approach to detect unreliable outputs in LLMs (à la hallucinations) using information-theoretic metrics and iterative prompting

6. The China Cyberspace Research Institute develops an LLM based on President Xi Jinping's philosophy, using a selective and authoritative knowledge base

Vision

1. Jina open sources Jina CLIP v1 - a new multimodal embedding model that significantly improves on OpenAI CLIP in text-only and image-to-image retrieval while maintaining state-of-the-art performance in cross-modal tasks

   

   Xenova's WebGPU demo using Jina's CLIP

2. Researchers open source Mobile-Agent-v2 - a multi-agent architecture that operates mobile devices, improving task completion compared to the previous version by over 30%

   

3. Stability AI announces it will open its 2B parameter Stable Diffusion 3 (SD3) Medium, enhancing photorealism and prompt responsiveness on consumer CPUs

   

4. UT Austin and Google develop RB-Modulation - a training-free method for diffusion models that ensures high fidelity to reference styles and text prompts, enabling seamless composition without external adapters or ControlNets

   

5. UT Austin and NVIDIA develop CamCo - a novel approach to enhance camera control and 3D consistency in image-to-video generation, achieving realistic object motion

   

6. Researchers present Video-MME - a new benchmark for multimodal LLMs in video analysis, showing Gemini 1.5 Pro's superior performance

   

7. Alibaba proposes I4VGen - a training-free video diffusion inference framework that enhances text-to-video generation by leveraging robust image techniques and a two-stage synthesis process

   

8. Researchers publish ZeroSmooth - a novel, training-free interpolation method for video diffusion models that boosts frame rates and ensures temporal consistency

   

9. Researchers propose ChronoDepth - a method for video depth estimation that leverages Stable Video Diffusion models and a procedural training strategy, achieving superior temporal consistency and spatial accuracy

   

10. Researchers introduce an innovative approach for diffusion models that uses a smaller, less-trained model for guidance, setting new benchmarks for image quality and maintaining variation

Audio

1. ByteDance introduces Seed-TTS - a family of TTS models that generate speech indistinguishable from human speech with superior control over attributes like emotion, achieving high performance in speaker similarity and naturalness

Open-source Packages

1. MusicGPT - generate music using natural language on your local machine

2. MusePose - turn an image of a human and a pose into an animated video

3. Ragapp - build deployable RAG-powered applications using a simple interface

4. Agenta - all-in-one LLM developer platform: prompt management, evaluation, human feedback, and deployment

Plus >70 more open-source packages for AI engineers

MusePose in action

Last week’s AI Tidbits roundup

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

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.