Farhad Kasad’s Post

Enhancing your gen AI use case with Vertex AI embeddings and task types #ai

💡 Did you know that Vertex AI's latest "Task types" embedding models can significantly improve #RAG search accuracy? ➡️ Problem: Traditional semantic similarity search often fails to deliver relevant answers in RAG. Why? Because questions and their answers aren't always semantically similar! 🤔 Question: "Why is the sky blue?" Answer: "The scattering of sunlight ..." ➡️ Solution: Vertex AI's new embedding models let you specify "task types" (like QUESTION_ANSWERING, CODE_RETRIEVAL_QUERY, etc.,) to capture the true relationship between queries and answers. ➡️ Results: More accurate search! This blog by Parashar S. and Kaz Sato explains these concepts in-depth and how to get started using them. #VertexAI #GenAI #RAG #GoogleCloud

Enhancing your gen AI use case with Vertex AI embeddings and task types Learn how to improve the search quality of your Retrieval Augmented Generation (RAG) system with new task type embeddings in Vertex AI. Optimize your Gen AI applications with this quick and efficient solution. Read mode on following blog post!

Microsoft AI Introduces LazyGraphRAG: A New AI Approach to Graph-Enabled RAG that Needs No Prior Summarization of Source Data Key takeaways: Cost Efficiency: LazyGraphRAG reduces indexing costs by over 99.9% compared to full GraphRAG, making advanced retrieval accessible to resource-limited users. Scalability: It balances quality and cost dynamically using the relevance test budget, ensuring suitability for diverse use cases. Performance Superiority: The system outperformed eight competing methods across all evaluation metrics, demonstrating state-of-the-art local and global query-handling capabilities. Adaptability: Its lightweight indexing and deferred computation make it ideal for streaming data and one-off queries. Open Source Contribution: Its integration into the GraphRAG library promotes accessibility and community-driven enhancements. More: https://lnkd.in/dfkdfy_P

Throwing it back to the Document Understanding AI journey in 2019! 🚀 Revisiting the breakthroughs and innovations that paved the way for smarter document handling. Unleashing the power of AI to understand documents like never before. Let's reminisce and celebrate the strides we've made! 📄✨ #AIInnovation #DocumentUnderstanding #TechFlashback https://lnkd.in/f6x49Eh

OpenAI’s GPT-4? Please. Claude's Opus Is Where It's A(I)t. TLDR: Opus is the largest model from Anthropic’s newest family of Claude 3 models. It now ranks at the top of the LMSYS Chatbot Arena Leaderboard, a crowdsourced open platform for evaluating AI models. Here are some things you can try with Claude: • Summarize Files like PDFs, research papers, essays, and more. • Understand Images: Upload an image, like a complicated diagram, and ask Claude to explain it to you. •Understand Complex Calculations: You can upload images of a calculation and ask it to explain what the equations are for and how to use them. • Suggest Design Improvements: Upload an image of a website — or even your living room — and ask how you can improve the design. • Transcribe Handwritten Notes: Upload pictures of handwritten notes and ask Claude to transcribe them for you. NOTE: Claude 3 Sonnet is free but the top-end Opus model requires you to subscribe to Claude Pro. PS Amazon added $2.75B to its investment in Anthropic, bringing its total investment in the artificial intelligence startup to $4B. The news marks the biggest investment Amazon has made in another company since its founding and represents its latest effort to compete with Microsoft and Google on AI.

📢 New GPT-4o-2024-08-06 API with #JSON Structured Outputs for precise and efficient AI outputs. 📯 Three Key Highlights of GPT-4o-2024-08-06 1) New Feature: JSON Structured Outputs: It ensures that outputs are well-structured and consistent. It also provides significant cost savings, with up to 50% savings on input costs and up to 33% savings on output costs. 2)  Improved Cost Efficiency: Input Costs: Reduced by up to 50% ($2.50 per 1M input tokens) Output Costs: Reduced by up to 33% ($10.00 per 1M output tokens) 3) Available globally. #AI #AzureOpenAI #AOAI #Llms more info:

I seem to have AI on my mind. Maybe it's the work I'm doing for Outlier.ai. I read an article in Semafor which discussed how the big players were lining up in this space. I've included an extract and the link below. I remembered this data visualization from Information is Beautiful (which has been recently updated) so I thought I'd share that too. AI is another evolution for technology and it's going to make our world spin even faster, changing it in ways which affect everybody. We're further on than we think with the RLHF stage of the game heading hurriedly towards deep learning and I find it fascinating! "At a high level, here’s how the five tech giants see the AI world: Microsoft believes its operating system and Office suite of tools will become supercharged with AI Copilots. At the same time, its cloud business will thrive in part because of its exclusive arrangement with OpenAI, the leading foundation model provider. Google thinks it can keep selling search ads the same way it always has and use its dominant position to keep the lead in the generative AI race. And it also hopes to benefit by offering AI services in the cloud. Meta’s bet is that the value of AI is not in the models themselves, but somewhere else in the ecosystem that AI will enable. It has decided to use its trove of data and infrastructure capabilities to create powerful foundation models and make them free, undercutting competitors and instantly making it a central player in the space. Amazon is looking to the large and powerful AI models other companies invent that will require massive compute, helping the e-commerce giant make a fortune by charging businesses to train and run AI models on its Amazon Web Services platform. And Apple, hindered in its AI capabilities because of its consumer privacy approach, is hoping that it can somehow catch up in the AI game without compromising its values. Perhaps it can shrink these powerful AI models down so they can run efficiently on smartphones, or outsource it to Google. In the long run, the real risk is that the technology is too capable. Who needs a search engine when powerful AI agents can tell you everything you need to know? And it doesn’t really matter which device or operating system I’m using when it’s all being filtered through a virtual assistant. If these models eventually run locally, the AI future may be decentralized, diminishing the role of the cloud. This is all a very long way off, but if you’ve read The Coming Wave, Suleyman’s book on AI, you know there are drastic, disruptive changes on the horizon." https://lnkd.in/ekTycuVW https://lnkd.in/exgYq9gQ

🌟 Introducing LLM-as-a-Judge in AWS Bedrock 🌟 During re:invent, AWS Bedrock has launched a new feature called 𝐋𝐋𝐌-𝐚𝐬-𝐚-𝐉𝐮𝐝𝐠𝐞 along with new RAG evaluation, designed to streamline the evaluation of AI models. This innovative tool allows enterprises to perform automated evaluations of their generative AI applications, ensuring high-quality outputs with minimal human intervention. 𝐖𝐢𝐭𝐡 𝐋𝐋𝐌-𝐚𝐬-𝐚-𝐉𝐮𝐝𝐠𝐞, 𝐲𝐨𝐮 𝐜𝐚𝐧: 1)𝐀𝐮𝐭𝐨𝐦𝐚𝐭𝐞 𝐐𝐮𝐚𝐥𝐢𝐭𝐲 𝐀𝐬𝐬𝐞𝐬𝐬𝐦𝐞𝐧𝐭𝐬: Evaluate model outputs using large language models (LLMs) to measure correctness, helpfulness, and adherence to responsible AI principles. 2)𝐒𝐚𝐯𝐞 𝐓𝐢𝐦𝐞 𝐚𝐧𝐝 𝐂𝐨𝐬𝐭: Reduce the need for extensive human evaluations by leveraging automated assessments. 3)𝐎𝐩𝐭𝐢𝐦𝐢𝐳𝐞 𝐀𝐈 𝐀𝐩𝐩𝐥𝐢𝐜𝐚𝐭𝐢𝐨𝐧𝐬: Quickly identify and address issues, shortening feedback loops and accelerating improvements. 𝐄𝐱𝐚𝐦𝐩𝐥𝐞𝐬 𝐨𝐟 𝐇𝐨𝐰 𝐋𝐋𝐌-𝐚𝐬-𝐚-𝐉𝐮𝐝𝐠𝐞 𝐖𝐨𝐫𝐤𝐬: 𝐑𝐞𝐟𝐞𝐫𝐞𝐧𝐜𝐞-𝐁𝐚𝐬𝐞𝐝 𝐄𝐯𝐚𝐥𝐮𝐚𝐭𝐢𝐨𝐧: Compare new responses against a reference or "ground truth" to ensure consistency. For instance, if you have a set of approved responses for customer support queries, LLM-as-a-Judge can automatically compare new responses to these references to check for accuracy. 𝐎𝐩𝐞𝐧-𝐄𝐧𝐝𝐞𝐝 𝐄𝐯𝐚𝐥𝐮𝐚𝐭𝐢𝐨𝐧: Evaluate responses based on custom criteria when there's no reference available. For example, if you're developing a chatbot, LLM-as-a-Judge can assess the relevance and coherence of its responses without needing predefined answers. This feature is currently in preview, but it's already showing great promise in enhancing the efficiency and reliability of AI development. Can't wait to see how it transforms the AI landscape! https://lnkd.in/g8ngnjkB #AWS #AI #MachineLearning #Innovation #AWSreInvent #Bedrock

🚀 Amazon's Olympus AI: A Game-Changer in the AI Landscape Amazon's Olympus AI is set to revolutionize the tech world with its massive 2 trillion parameter model! 🤖 Here's why it's a big deal: • Large-Scale AI model : 2x bigger than GPT-4 • Multi-modal capabilities: Processes text, images, and videos • AWS integration: Seamless cloud solutions for enterprises • Natural language search: Find specific scenes in videos effortlessly Potential industry impacts: 📦 E-commerce: Enhanced product discovery 🎬 Media: Automated content analysis 🏭 Industrial: Streamlined equipment inspections This strategic move positions Amazon as a frontrunner in AI, challenging OpenAI and Google head-on. What are your thoughts on this AI race? Will Olympus redefine industry standards? ------------------------------------------------------------------- Please follow PONIAK for more such topics. Please follow our handle on X - https://x.com/PoniakTimes ------------------------------------------------------------------- https://lnkd.in/gpypYCPv Share your insights in the comments! 👇 #AmazonAI #OlympusAI #TechInnovation #AIRevolution #AWS #Poniak