AI Innovations: Unveiling the Latest Breakthroughs
Welcome to the July 2024 Edition of Bayes Bulletin!
Uncover the industry's latest breakthroughs, from innovative models to real-world applications. Stay informed and inspired as we navigate through the dynamic landscape of AI that is shaping the future of technology.
Subscribe for a journey into the future of intelligence!
Latest Models:
1. Introducing GPT-4o Mini: Open AI's Most Cost-Efficient Small Model
OpenAI introduces GPT-4o Mini, a cost-efficient small model designed for affordability and versatility. Scoring 82% on the MMLU benchmark and outperforming GPT-41 in chat preferences, it's priced at just 15 cents per million input tokens and 60 cents per million output tokens—over 60% cheaper than GPT-3.5 Turbo. Ideal for low-cost, low-latency tasks, it supports text and vision in the API with plans to include text, image, video, and audio inputs and outputs. With a 128K token context window and support for up to 16K output tokens per request, GPT-4o Mini excels in reasoning, math, and coding, outperforming small models like Gemini Flash and Claude Haiku, making it a powerful tool for diverse applications.
2. Introducing Codestral Mamba: A New Frontier in AI Architecture
Codestral Mamba, developed with Albert Gu and Tri Dao, is an advanced AI architecture featuring linear time inference and the ability to model infinite-length sequences, ensuring quick responses regardless of input size. This efficiency is ideal for code productivity tasks. Matching state-of-the-art transformer models, it excels as a local code assistant, tested with in-context retrieval up to 256k tokens. Deployable via the mistral-inference SDK and TensorRT-LLM, and soon locally through llama.cpp, it is free to use, modify, and distribute under the Apache 2.0 license. The 7.3 billion parameter model is available for testing on la Plateforme (codestral-mamba-2407), alongside the larger Codestral 22B, available under a commercial or community license.
Latest Frameworks:?
1.ScrapeGraphAI
ScrapeGraphAI, an open-source Python library designed to transform data scraping by integrating Large Language Models (LLMs) and modular graph-based pipelines to automate data extraction from diverse sources like websites and local files. Unlike traditional tools that rely on fixed patterns and manual configurations, Scrape Graph AI adapts to changes in website structures, minimizing the need for constant developer intervention. It supports a wide range of LLMs, including GPT and Gemini. It provides a highly customizable and low-maintenance solution. Key features include adaptability to website changes, support for various document formats, handling context window limits through chunking and compression, and troubleshooting with detailed logging, proxy rotation, and graphical pipeline visualization.
2. FastRAG by Intel:
fastRAG is a research framework designed for efficient and optimized retrieval augmented generative (RAG) pipelines, incorporating state-of-the-art LLMs and Information Retrieval. fastRAG empowers researchers and developers with a comprehensive toolset for advancing retrieval augmented generation. Now compatible with Haystack v2+, fastRAG's latest updates include Haystack 2.0 compatibility, Gaudi2 and ONNX runtime support, optimized embedding models, multi-modality and chat demos, and REPLUG text generation. Key features include optimized RAG pipelines for greater compute efficiency, Intel hardware optimization leveraging Intel extensions for PyTorch (IPEX), and full compatibility with Haystack and Hugging Face components.
Latest Research papers:
1.Lynx: An Open Source Hallucination Evaluation Model
LYNX, a new open-source LLM, tackles hallucinations in large language models (LLMs) – outputs that stray from factual accuracy or context. It excels at identifying these inconsistencies in Retrieval-Augmented Generation (RAG) systems. LYNX surpasses existing models on the novel HaluBench benchmark, a collection of real-world tasks from various domains like finance and medicine. This research offers a significant step forward in LLM evaluation, ensuring RAG systems produce reliable and trustworthy outputs. By detecting hallucinations, LYNX helps ensure these systems produce accurate and trustworthy outputs.
领英推荐
2.AutoRAG-HP: Automatic Online Hyper-Parameter Tuning for Retrieval-Augmented Generation? ? ?
This research paper proposes AutoRAG-HP, a framework that automatically tunes hyper-parameters in RAG systems. This framework simplifies the complex task of optimizing RAG performance by treating hyper-parameter selection as a multi-armed bandit problem(MAB) and introduces a novel two level Hierarchical MAB (Hier-MAB) method for efficient exploration of large search spaces. Experiments show good results, particularly for tasks of moderate complexity. The ultimate goal is for AutoRAG-HP to become a comprehensive AutoML framework for RAG, automatically optimizing various aspects and generating its own evaluation data.
AI Startup news:?
1.AI startup Hebbia raised $130M at a $700M valuation on $13 million of profitable revenue
Hebbia, an AI startup that leverages generative AI for searching and analyzing large documents, has raised $130 million in a Series B round, valuing the company at $700 million. The round was led by Andreessen Horowitz with contributions from Index Ventures, Google Ventures, and Peter Thiel. Founded by George Sivulka in 2020 while pursuing his PhD at Stanford, Hebbia boasts an ARR of $13 million and has achieved profitability, with revenue growing 15x in the last 18 months. The company’s primary product, Matrix, enables users to sift through extensive documents and respond to queries in a spreadsheet-like format, catering mainly to asset managers, investment banks, and financial institutions. Hebbia plans to use the new funds to expand its team, enhance its offerings to the financial sector, and enter new markets such as law firms and pharmaceutical companies.
2. Didero is using AI to solve supply chain management at mid-market?companies
Didero, a new startup focused on streamlining supply chain management for mid-market companies, has secured $7 million in seed funding and is now publicly launching its AI-powered tool. The product aims to simplify procurement tasks such as finding suppliers, negotiating contracts, managing orders, and handling payments. By leveraging AI models like OpenAI and Google Gemini, Didero addresses the resource constraints of smaller companies, offering automated solutions that were previously unattainable. The company uses both general and specialized AI models for tasks like data extraction from key procurement documents. Founded by Tim Spencer, Lorenz Pallhuber, and Tom Petit, Didero's seed round was led by First Round Capital with contributions from several notable investors.
AI Conferences-?
1.International Conference on Machine Learning (ICML) 2024
Renowned for its state-of-the-art research in artificial intelligence, statistics, data science, and machine learning, the International Conference on Machine Learning (ICML) has applications in robotics, machine vision, computational biology, and speech recognition. Exhibitors at this year's ICML include Google, Microsoft, Amazon, D. E. Shaw & Co., Jump Trading, and HSBC. There are also 2000+ teams, 100+ tech researchers, 1000+ collaborators, and 47 partner organizations, including Georgia Tech and the University of California.?In addition to the topics addressed, there are workshops and affinity events including the Women in Machine Learning Symposium, the LatinX in AI Research Workshop, LLMs, GNNs, DL, and Bayesian models. Neural network tutorials are also available.
Best Paper Awards at ICML 2024:
2.IEEE World Congress on Computational Intelligence (WCCI) 2024
The World Congress on Computational Intelligence (WCCI) is the largest event in computational intelligence, featuring three key IEEE CIS conferences: the International Joint Conference on Neural Networks (IJCNN), the IEEE International Conference on Fuzzy Systems (FUZZ-IEEE), and the IEEE Congress on Evolutionary Computation (IEEE CEC). WCCI 2024 will be held in Yokohama, Japan, a hub for academic and industrial collaboration in advanced technologies. The conference will focus on AI, physiology, and psychology, fostering partnerships with intelligence-related industries. Attendees will explore topics from neural networks at IJCNN 2024 to evolutionary computation at IEEE CEC 2024, and fuzzy systems at FUZZ-IEEE 2024, making it a key event for innovation in computational intelligence.
Stay tuned for more!