Sung Kim’s Post

🚀 Exciting News: Got It AI Achieves Lower than GPT-4 Hallucination Rates using fine-tuned Mixtral and our TruthChecker model! 🎉 In our latest blog post, we share the impressive results of our Mixtral-based RAG pipeline, which leverages Mixtral-8x7B models for both response generation and fact checking. 🌟 Key Highlight: Our fine-tuned Mixtral response generator model, combined with our fine-tuned Mixtral TruthChecker model to detect hallucinations, achieves a net 2.0% hallucination rate, better than GPT-4's hallucination rate of 2.1% on the same enterprise dataset using identical relevance and groundedness evaluation criteria. 📊 Results: - Base Mixtral-8x7B: 8% hallucination rate - Fine-tuned Mixtral response generator: 4.5% hallucination rate - Fine-tuned Mixtral response generator + Fine-tuned Mixtral TruthChecker: 2.0% hallucination rate (better than GPT-4’s 2.1%) Our findings demonstrate that open-source models, when fine-tuned for specific tasks and customer data, can compete with the best closed models available. 💪 Read the full blog post to learn more about our methodology and insights: https://lnkd.in/gDV7sqhH #GenerativeAI #LLMs #ArtificialIntelligence #MachineLearning #Mistral #Mixtral #GPT4 #Hallucination #GotItAI

LMSYS Org has introduced RouteLLM an open-source framework for cost-effective LLM routing. Queries can now be efficiently routed between various models based on the prompts, thus balancing performance and cost, with up to 85% cost reduction achieved while maintaining 95% of top model performance. The demo and code can be explored on GitHub. Blog : https://lnkd.in/g3aw7fXD Models and datasets: https://lnkd.in/gyD_d4ra Github: https://lnkd.in/gDRx9XmW Paper: https://lnkd.in/gaZvJYv4 #AI #GenAI #MachineLearning #OpenSource #RouteLLM #CostEfficiency #LLM

#AgenticAI is coming, are you ready? In my #opinion this is where the future is going and how we are going to increase quality and action while decreasing cost. it will be several years before we see it in the #smallbusiness arena, BUT, small businesses must work on data collection, cleaning and governance NOW to be able to leverage #Ai when it is more cost effective. #datamatters #doitnow #futurethoughts

Why Choose Extreme-ML? Zero-Code Platform: Harness the power of advanced AI without writing a single line of code. Lightning-Fast Performance: Process large datasets and generate insights at unprecedented speeds. End-to-End Solution: From data preparation to model deployment, Extreme-ML covers your entire workflow. Unparalleled Transparency: Understand every step of your analysis with our clear, interpretable outputs. Highly Versatile: Adapt to any data science challenge with our flexible, comprehensive toolset. Visit -- > https://lnkd.in/dsbCWbtb #MachineLearning #DataScience #AI #ModelDevelopment

We all know RAG unlocks a new level of information access for AI, letting it answer questions beyond its training data. But sometimes RAG can struggle with context.  Imagine asking, "Who is Ravi's best friend?" RAG might find 'Ravi'-related entities, but it may or may not point the best friend directly. That's where Knowledge Graphs come in. KGs map relationships between entities, giving AI a deeper understanding of the world. Think of it like this: the "Ravi" entity could have connections like "KNOWS," "LIKES," or "TALKED TO" with other entities in the KG. By combining RAG's data access with KG's relational embedding, AI can make smarter connections. AI can create the KG itself from text snippets similar to how RAG is made. This would make a good tool for tackling "multi-hop" questions, where the answer requires following multiple connections. So, if you're working with RAG, consider incorporating Knowledge Graphs. This project implements RAG with Knowledge Graph: https://lnkd.in/gY4duG8B #KnowledgeGraphs #RAG

📢 Exciting news for AI enthusiasts! The latest dimensionality reduction techniques such as UMAP now allow you to visualize RAG data and see the relationships between questions, answers, and sources. The complex, high-dimensional data is transformed into a clear, interactive two-dimensional map that can be used to debug and improve the performance of RAG models. Color-coded by their relevance to the question "Who built the Nürburgring?" - it's a great example of the power of this visualization technique. 😎 Check out the link below for the tutorial: https://lnkd.in/gnVnvTAZ   #AI #RAG #LinkLayer

Today, I happily received 𝐑𝐀𝐆-𝐃𝐫𝐢𝐯𝐞𝐧 𝐆𝐞𝐧𝐞𝐫𝐚𝐭𝐢𝐯𝐞 𝐀𝐈 by Denis Rothman. This is the second book I’ve read from Denis this year, and just like the first, it’s packed with practical insights that bridge the gap between theory and real-world applications. I had the chance to read the early review version, which gave me a good look at the valuable content in this book. The clear explanations of complex AI concepts, combined with the hands-on approach, make this book both practical and informative. It’s a resource I found highly useful and worth recommending. The book covers retrieval-augmented generation (RAG) in a clear and practical way, even for complex topics. I especially appreciated how it provides step-by-step guidance on building RAG pipelines and reducing AI errors by connecting responses to real source documents. For more details, you can check out the book here: https://lnkd.in/e2--EYPC 𝐒𝐨𝐦𝐞 𝐤𝐞𝐲 𝐟𝐞𝐚𝐭𝐮𝐫𝐞𝐬: 𝐊𝐧𝐨𝐰𝐥𝐞𝐝𝐠𝐞 𝐠𝐫𝐚𝐩𝐡-𝐛𝐚𝐬𝐞𝐝 𝐑𝐀𝐆: Uses knowledge graphs for more accurate, traceable information retrieval. 𝐃𝐲𝐧𝐚𝐦𝐢𝐜 𝐑𝐀𝐆: Enhances real-time decision-making with tools like Chroma and Hugging Face Llama. 𝐈𝐧𝐝𝐞𝐱-𝐛𝐚𝐬𝐞𝐝 𝐑𝐀𝐆 𝐰𝐢𝐭𝐡 𝐋𝐥𝐚𝐦𝐚𝐈𝐧𝐝𝐞𝐱, 𝐃𝐞𝐞𝐩 𝐋𝐚𝐤𝐞, 𝐚𝐧𝐝 𝐎𝐩𝐞𝐧𝐀𝐈: Builds RAG pipelines using indexing techniques for efficient data retrieval. 𝐌𝐮𝐥𝐭𝐢𝐦𝐨𝐝𝐚𝐥 𝐑𝐀𝐆 𝐟𝐨𝐫 𝐝𝐫𝐨𝐧𝐞 𝐭𝐞𝐜𝐡: Combines text and visual data for dynamic AI applications. 𝐌𝐢𝐧𝐢𝐦𝐢𝐳𝐞 𝐀𝐈 𝐡𝐚𝐥𝐥𝐮𝐜𝐢𝐧𝐚𝐭𝐢𝐨𝐧𝐬: Reduces errors by linking responses to real source documents. A big thank you to Anamika Singh and Packt for providing me the book, and to Denis Rothman for making advanced AI topics so easy to understand and apply. It’s been a rewarding read, and I highly recommend it to anyone from AI enthusiasts to professionals looking to deepen their understanding of this field.   #AI #RAG #RetrievalAugmentedGeneration #GenerativeAI #MachineLearning #PacktPublishing

Polling data can be difficult to understand in isolation. Context is needed to dive deeper into the story behind the data. Today, RealClearPolitics employs ContextLens on its RealClearPolling site — a new generative AI tool that generates informational graphics based on readers' needs. "This is the next step forward in how AI can be used in publishing to enhance reader experience, instead of solely focusing on AI as only an efficiency tool," says Code and Theory's Dan Gardner. "ContextLens offers a glimpse into the future of AI and design, where context, anticipation and visual interpretation come together to create a fluid user experience." Created by Code and Theory, the ContextLens is currently in beta and available to other publishers. Learn more: https://lnkd.in/gQFjMQhN

🚀 New Highly optimized 61B and 50B LLMs from Nyun AI! 🚀 At Nyun AI, our mission is to create the most optimized AI models to help our clients speed up deployments and save costs. Today, we're thrilled to introduce our latest innovation: the nyun-c2 class of models! These models deliver high performance with significantly fewer parameters, revolutionizing the efficiency of large language models (LLMs), all possible via our compression platform - nyunkompress. ✨ Introducing nyun-c2-llama3-61B and nyun-c2-llama3-50B ✨ 🔹 nyun-c2-llama3-61B: With 13% fewer parameters, this model outperforms Llama3-70B by 1.5%. 🔹 nyun-c2-llama3-50B: This variant uses 29% fewer parameters than Llama3 with only a 2% performance drop. 🔹 Performance Benchmark: Our nyun-c2-llama3-61B model, with 23% fewer parameters, surpasses MBZUAI-K21-65B by an impressive 5.8%. We are confident that this release will significantly impact the speed and cost of LLM deployments. We are eager to connect with clients and partners interested in designing and deploying powerful open-source LLMs in-house. Huggingface Links - nyun-c2-llama3-61B - https://lnkd.in/grQY3g_X nyun-c2-llama3-50B - https://lnkd.in/gr-pr6cc 📧 Reach out to us at connect@nyunai.com. Notes: 1. Our models and pipeline have undergone extensive testing, and the datasets used for benchmarking are the only ones we've tested on. – no cherry-picking! 2. We do not fine-tune over Llama3-70B in any stage of compression, which is a significant achievement as we induce no external knowledge in the model. #AI #MachineLearning #LLM #TechInnovation #NyunAI #AIoptimization #LLMdeployment #AIefficiency #OpenSourceAI #AIresearch #TechNews

𝗧𝗵𝗲 𝗝𝗮𝗴𝗴𝗲𝗱 𝗙𝗿𝗼𝗻𝘁𝗶𝗲𝗿: 𝘄𝗵𝗮𝘁 𝗱𝗼𝗲𝘀 𝗶𝘁 𝗺𝗲𝗮𝗻 𝗳𝗼𝗿 𝘆𝗼𝘂𝗿 𝗔𝗜 𝗮𝗽𝗽𝗹𝗶𝗰𝗮𝘁𝗶𝗼𝗻𝘀? Last year, Ethan Mollick and colleagues proposed the metaphor of the Jagged Frontier to explain that “some tasks are easily done by AI, while others, though seemingly similar in difficulty level, are outside the current capability of AI”. How does the Jagged Frontier change over time? And, what are the implications for your organization? 1. 𝗘𝘅𝗽𝗮𝗻𝗱𝗶𝗻𝗴 𝗙𝗿𝗼𝗻𝘁𝗶𝗲𝗿: while the edge of capabilities remains jagged, the boundary of what is possible keeps expanding. Generative AI models keep on improving and capabilities expanding. 2. 𝗡𝗼𝗻-𝗕𝗮𝗰𝗸𝘄𝗮𝗿𝗱 𝗖𝗼𝗺𝗽𝗮𝘁𝗶𝗯𝗶𝗹𝗶𝘁𝘆: The jaggedness keep moving due to evolving alignment of models. What was possible in the past, is not necessarily possible on the future. "Silent releases" make this issue worse. 3. 𝗥𝘂𝗴𝗴𝗲𝗱 𝗣𝗿𝗼𝗴𝗿𝗲𝘀𝘀: The jaggedness remain the same but deepens further over time. That is, models become better on tasks they already did well (e.g., text understanding) but remain surprisingly bad at other tasks (e.g., counting). 𝗚𝗿𝗲𝗮𝘁 𝗼𝗯𝘀𝗲𝗿𝘃𝗮𝘁𝗶𝗼𝗻, 𝗕𝗲𝗻, 𝗯𝘂𝘁 𝘄𝗵𝗮𝘁 𝗱𝗼𝗲𝘀 𝗶𝘁 𝗺𝗲𝗮𝗻 𝗳𝗼𝗿 𝗺𝗲? When you are developing AI-powered applications, you want to future proof those. Here are my three guidelines of how to do so. ↳ 𝗙𝗹𝗲𝘅𝗶𝗯𝗹𝗲 𝗔𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲𝘀: Implement modular and interoperable system. Ensure that you can swap models to follow the rapid advancements in technology. ↳ 𝗟𝗮𝗿𝗴𝗲-𝗦𝗰𝗮𝗹𝗲 𝗧𝗲𝘀𝘁𝗶𝗻𝗴: Your prompt will break at some point. Automate modification and testing of prompts to quickly adapt to new models when you need to. ↳ 𝗦𝗽𝗲𝗰𝗶𝗮𝗹𝗶𝘇𝗲𝗱 𝗦𝗺𝗮𝗹𝗹 𝗠𝗼𝗱𝗲𝗹𝘀: Develop and refine your own task-specific, small-scale AI models for high performance in specialized areas. Employing your own smaller models makes you less depended on idiosyncratic changes in frontier models. The Jagged Frontier remains hard to navigate. How are you dealing with these challenges in your AI projects? Share your stories and strategies in the comments below! #AI #LLM #jaggedfrontier #prompting ___ Enjoyed this post? Like 👍, comment 💭, or re-post ♻️ to share with others. Questions about how to do this? DM me!