AI Roundup 07/25 -> 08/01/2024

OpenAI's SearchGPT and ChatGPT's Advanced Voice Mode launches, Google's open Gemma 2B outperforming GPT-3.5, DeepMind’s AI silver medalist in math, and Meta’s new Segment Anything 2 vision model

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.

📩 Published a new breakthrough paper? Just released an open-source package? Submit it here to ensure we don’t miss it and that it gets featured in next week’s post.

---

Overview

• ✨ Highlights (6 entries)

• Language Models (7 entries)

• Robotics (4 entries)

• Autonomous Agents (3 entries)

• Vision (8 entries)

• Audio (2 entries)

• Open-source Packages (2 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. OpenAI showcases SearchGPT - a Google Search competitor combining real-time web information with conversational abilities for precise and fast answers (OpenAI)

   

2. OpenAI begins alpha testing Advanced Voice Mode - ChatGPT's new real-time and natural conversational AI demoed in OpenAI's April Spring Updates event (X)

   

3. Google releases Gemma-2-2B - a commercially permissive compact 2 billion parameter model that outperforms models 20x its size, including Mixtral 8x7B, GPT 3.5, and Llama-2 70B

4. Meta releases Segment Anything Model (SAM) 2 - a real-time promptable object segmentation model for images and videos, achieving state-of-the-art performance and broad applicability without custom adaptation (Meta)

5. Meta presents AI Studio - a platform that allows creators to easily build and customize AI characters to interact with followers, leveraging Llama 3.1 models (Meta)

6. DeepMind presents AlphaProof and AlphaGeometry 2 - AI models that solve advanced math problems, matching the performance of silver medalists in the International Mathematical Olympiad (DeepMind)

Language Models

1. Galileo release the LLM Hallucination Index - evaluating 22 leading models across 3 RAG tasks, providing insights on cost, verbosity, and performance to help teams choose the optimal LLM for their use case

   

2. Alibaba openly releases SeaLLMs v3 - an LLM designed for Southeast Asian languages, achieving state-of-the-art performance in various tasks with reduced training costs

   

3. Researchers release SaulLM-54B and SaulLM-141B - open large language models tailored for the legal sector, achieving state-of-the-art performance through extensive domain adaptation and legal-specific protocols

   

4. OpenAI develops Rule Based Rewards (RBR) - a new AI feedback alternative to RLHF that uses minimal human data to enhance LLM safety behaviors, achieving a superior F1 score of 97%

   

5. Cohere releases Rerank 3 Nimble - a foundation model three times faster than its predecessor, optimizing enterprise search and RAG systems with high accuracy

   

6. Researchers introduce LAMBDA - a code-free multi-agent data analysis system leveraging advanced models to enhance data analysis in complex applications through natural language interactions

   

7. Meta AI develops Discrete Flow Matching - a new paradigm supporting diverse probability paths that significantly enhances high-dimensional discrete data generation, achieving substantial improvements in generative perplexity and benchmark performance

   

Robotics

1. Researchers develop Theia - a comprehensive vision foundation model that advances robot learning by combining multiple vision tasks, demonstrating superior performance and efficiency

   

2. Nvidia releases an update on its generalistic robotics project, Project GR00T, introducing a new data pipeline that multiplies robot demonstration data through immersive teleoperation and generative simulation

   

3. Berkeley introduces Berkeley Humanoid - a low-cost humanoid robot platform optimized for learning-based control with robust, agile locomotion across varied terrains

   

4. ICL and DeepMind introduce Diffusion Augmented Agents (DAAG) - a framework combining LLMs and diffusion models to enhance reinforcement learning sample efficiency and transfer learning, demonstrating significant improvements in simulated robotics tasks

   

Autonomous Agents

1. Google presents Oscar - an open-source project that utilizes AI agents to streamline the maintenance of open-source software by handling tedious tasks

2. Researchers develop AppWorld Engine and AppWorld Benchmark to advance interactive coding agents, showcasing GPT-4o's superior performance on complex tasks with up to 49% success, demonstrating the benchmark's rigorous challenges and potential for progress

3. Apple publishes the Massive Multitask Agent Understanding (MMAU) benchmark to comprehensively evaluate LLMs across five domains and essential capabilities, providing deep insights and enhanced interpretability

Vision

1. DeepMind presents Mixture of Nested Experts (MoNE) - a framework that reduces inference time compute by over 2x for Vision Transformer models while maintaining equivalent performance

2. Nvidia, Berkeley, and MIT introduce Wolf - a mixture-of-experts video captioning framework that leverages Vision Language Models to enhance video understanding, outperforming GPT-4V and achieving superior performance over state-of-the-art methods

   

3. Researchers present Visual Riddles - a benchmark testing vision and language models on complex visual scenarios requiring commonsense and world knowledge, revealing a significant gap between human (82%) and model (40%) performance

   

4. Alibaba develops Tora - a novel Diffusion Transformer framework that achieves high motion fidelity and realistic video generation with controllable dynamics

   

5. Researchers unveil EchoMimic - an innovative approach for portrait image animation that leverages both audio and facial landmarks to produce more stable and natural results compared to conventional methods

   

6. Researchers develop a method combining StyleGAN2 and guided super-resolution to produce photorealistic 3D avatars with studio-quality lighting and complete textures from casual phone captures

   

7. Researchers develop BetterDepth - a novel method that significantly improves zero-shot monocular depth estimation by refining pre-trained model outputs through a diffusion-based approach, enhancing both geometric accuracy and detail capture

   

8. UIUC and Purdue University present ORG - a novel approach for 3D object reconstruction from single images that accurately captures the relationship between the object, ground, and camera, enhancing shadow generation and pose manipulation

   

Audio

1. Stanford releases DiVA Llama 3 - an open distilled voice assistant that outperforms existing speech language models in QA, emotion recognition, and translation

Open-source Packages

1. MindSearch - an LLM-based agent framework for web search (like SearchGPT or Perplexity)

2. OpenPlexity-Pages - an open-source alternative to Perplexity Pages (content creation engine powered by AI)

Plus >70 more open-source packages for AI engineers

MindSearch leverages LLM-powered agent for web search and summarization

Last week’s AI Tidbits roundup

AI Roundup 07/18 -> 07/25/2024

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.