Ganesh Jagadeesan’s Post

In the fast-paced world of AI and LLMs, finding a profile that is both insightful and brimming with hands-on experience can feel like hitting a jackpot. For those navigating between the business and tech worlds, seeking to understand what's achievable, I strongly suggest taking a look at Hanane D.'s profile and her most recent posts. Of course, I enjoy reading Andrew Ng thoughts, but who are some other visionaries in the field that you find worth following? Growth comes with exploration and learning from others! #AI #LLMs #TechLeaders

The ability to make sense of charts and tables within a complex financial report remains a key challenge which seems to be on track for resolution. this could drastically improve the performance of RAG pipelines in the coming months.

📣 DON'T MISS: Webinar "AI in production" 📌22.11.2024 @Bayern Innovativ GmbH Join us to learn how 𝗹𝗮𝗿𝗴𝗲 𝗹𝗮𝗻𝗴𝘂𝗮𝗴𝗲 𝗺𝗼𝗱𝗲𝗹𝘀 and 𝘃𝗲𝗰𝘁𝗼𝗿 𝗱𝗮𝘁𝗮𝗯𝗮𝘀𝗲𝘀 can make your production processes more efficient and secure. Gain valuable insights into the integration of these technologies, 𝗚𝗗𝗣𝗥-𝗰𝗼𝗺𝗽𝗹𝗶𝗮𝗻𝘁 𝗱𝗮𝘁𝗮 𝗽𝗿𝗼𝗰𝗲𝘀𝘀𝗶𝗻𝗴 and protection against unauthorized access. Discover how AI applications such as 𝗥𝗔𝗚 𝘀𝘆𝘀𝘁𝗲𝗺𝘀 contribute to the optimization of data quality and which measures facilitate the integration into existing systems.   📊 What you'll learn: o  Applications in production data analysis and quality assurance o  Strategies for efficiency gains and predictive maintenance o  Enhancing internal communication   Speaker: Johannes be Isa (INOYAD GmbH): Quality assurance through computer vision: increasing efficiency in production with AI-supported systems Wolfgang Graßhof (WOGRA AG): Unleash the potential of LLMs and vector databases for your corporate production!   Free registration ➡️ https://lnkd.in/d8S46t-F transform.by   #production #ai #computervision #qualityassurance

Retrieval-Augmented Generation (RAG) combines search and generative AI for accurate, context-aware responses but comes with limitations. Data preparation is challenging, as messy sources like PDFs and Word documents introduce noise, affecting retrieval accuracy. The retrieval process operates as a "black box," and splitting long documents can break context. Generative models are constrained by limited context windows, often leading to incomplete answers for complex queries. Additionally, insufficient data context can cause hallucinations, resulting in fabricated information. RAG is excellent for extracting relevant fragments but struggles with exhaustive queries. Optimizing data and retrieval processes is essential to maximize its effectiveness. 了解RAG的限制後，對LLM為什麼胡言亂語，胡亂分段、斷章取義有更深的了解。也知道該如何和LLM互動 #betterPrompt

Anthropic has introduced PDF analysis in the Claude 3.5 Sonnet model. Here’s what you need to know: ☑ Versatile Capabilities: → Claude can analyze any standard PDF—this includes extracting data from charts, tables, and text. Perfect for tasks like: 1️⃣Analyzing financial reports 2️⃣Extracting information from legal documents 3️⃣Providing translation support ☑ Easy Integration: → You can send any PDF directly through the API. Just make sure your files are under 32MB and have fewer than 100 pages! ☑ Image-Based Cost: → Note that each PDF page is treated as an image and may use 1,500 to 3,000 tokens depending on the content density. No extra fees—just standard input pricing. This new functionality not only streamlines document processing but also enhances how we interact with complex data in PDFs. What are your thoughts on integrating AI with document analysis?

𝗠𝗼𝘀𝘁 𝗯𝘂𝘀𝗶𝗻𝗲𝘀𝘀𝗲𝘀 𝗱𝗼𝗻’𝘁 𝗿𝗲𝗮𝗹𝗶𝘇𝗲 𝘁𝗵𝗲𝘆’𝗿𝗲 𝗹𝗲𝗮𝘃𝗶𝗻𝗴 𝗺𝗮𝘀𝘀𝗶𝘃𝗲 𝗽𝗼𝘁𝗲𝗻𝘁𝗶𝗮𝗹 𝗼𝗻 𝘁𝗵𝗲 𝘁𝗮𝗯𝗹𝗲 𝗱𝘂𝗲 𝘁𝗼 𝗶𝗻𝗲𝗳𝗳𝗶𝗰𝗶𝗲𝗻𝘁 𝗔𝗜 𝘄𝗼𝗿𝗸𝗳𝗹𝗼𝘄𝘀. DXwand creates multiple experiments to optimize accuracy, utilizing different chunk sizes and various document types, all identified with unique metadata. Our evaluation system includes a 𝗥𝗲𝘁𝗿𝗶𝗲𝘃𝗮𝗹 𝗦𝗰𝗼𝗿𝗲 to measure how well the system finds correct answers and a 𝗥𝗲𝘀𝗽𝗼𝗻𝘀𝗲 𝗦𝗰𝗼𝗿𝗲 to assess chunk accuracy. This approach ensures you achieve the best results. 𝗗𝗼𝗻’𝘁 𝗹𝗲𝘁 𝘀𝗹𝗼𝘄 𝗔𝗜 𝗽𝗿𝗼𝗴𝗿𝗲𝘀𝘀 𝗵𝗼𝗹𝗱 𝘆𝗼𝘂 𝗯𝗮𝗰𝗸. With DXwand, 𝘆𝗼𝘂𝗿 𝘀𝗼𝗹𝘂𝘁𝗶𝗼𝗻𝘀 𝗰𝗮𝗻 𝗯𝗲 𝗱𝗲𝗽𝗹𝗼𝘆𝗲𝗱 𝟵𝟬% 𝗳𝗮𝘀𝘁𝗲𝗿 𝗮𝘁 𝟵𝟱% 𝗹𝗲𝘀𝘀 𝗰𝗼𝘀𝘁, 𝘄𝗵𝗶𝗹𝗲 𝗺𝗮𝗶𝗻𝘁𝗮𝗶𝗻𝗶𝗻𝗴 𝟯𝟬% 𝗵𝗶𝗴𝗵𝗲𝗿 𝗮𝗰𝗰𝘂𝗿𝗮𝗰𝘆 than the industry average. Book a demo through the link below and discover how DXwand can transform your AI development from a crawl to a sprint. https://lnkd.in/gBmYkdFS

People once wondered if AI could learn and adapt to the battlefield at the speed of relevance. What if it encounters something it’s never seen? Can adversaries trick the perception system? Doesn’t it require thousands of images and videos from relevant environments to perform well? Today's answers: - Chances are, it will classify correctly. - Not for long—tricks won’t last. - No, an image or two is enough. A never-before-seen target can be correctly classified by reasoning about the sum of its parts. And if you manage to confuse the AI, high-performance is just a few training images and a software update away. Thirteen minutes. That’s the time it takes us to retrain our AI on a never-before-seen target and achieve high accuracy, precision, and recall. What used to take months of data collection and annotation—or was deemed infeasible—can now be accomplished in minutes.

𝗠𝗼𝘀𝘁 𝗯𝘂𝘀𝗶𝗻𝗲𝘀𝘀𝗲𝘀 𝗱𝗼𝗻’𝘁 𝗿𝗲𝗮𝗹𝗶𝘇𝗲 𝘁𝗵𝗲𝘆’𝗿𝗲 𝗹𝗲𝗮𝘃𝗶𝗻𝗴 𝗺𝗮𝘀𝘀𝗶𝘃𝗲 𝗽𝗼𝘁𝗲𝗻𝘁𝗶𝗮𝗹 𝗼𝗻 𝘁𝗵𝗲 𝘁𝗮𝗯𝗹𝗲 𝗱𝘂𝗲 𝘁𝗼 𝗶𝗻𝗲𝗳𝗳𝗶𝗰𝗶𝗲𝗻𝘁 𝗔𝗜 𝘄𝗼𝗿𝗸𝗳𝗹𝗼𝘄𝘀. DXwand creates multiple experiments to optimize accuracy, utilizing different chunk sizes and various document types, all identified with unique metadata. Our evaluation system includes a 𝗥𝗲𝘁𝗿𝗶𝗲𝘃𝗮𝗹 𝗦𝗰𝗼𝗿𝗲 to measure how well the system finds correct answers and a 𝗥𝗲𝘀𝗽𝗼𝗻𝘀𝗲 𝗦𝗰𝗼𝗿𝗲 to assess chunk accuracy. This approach ensures you achieve the best results. 𝗗𝗼𝗻’𝘁 𝗹𝗲𝘁 𝘀𝗹𝗼𝘄 𝗔𝗜 𝗽𝗿𝗼𝗴𝗿𝗲𝘀𝘀 𝗵𝗼𝗹𝗱 𝘆𝗼𝘂 𝗯𝗮𝗰𝗸. With DXwand, 𝘆𝗼𝘂𝗿 𝘀𝗼𝗹𝘂𝘁𝗶𝗼𝗻𝘀 𝗰𝗮𝗻 𝗯𝗲 𝗱𝗲𝗽𝗹𝗼𝘆𝗲𝗱 𝟵𝟬% 𝗳𝗮𝘀𝘁𝗲𝗿 𝗮𝘁 𝟵𝟱% 𝗹𝗲𝘀𝘀 𝗰𝗼𝘀𝘁, 𝘄𝗵𝗶𝗹𝗲 𝗺𝗮𝗶𝗻𝘁𝗮𝗶𝗻𝗶𝗻𝗴 𝟯𝟬% 𝗵𝗶𝗴𝗵𝗲𝗿 𝗮𝗰𝗰𝘂𝗿𝗮𝗰𝘆 than the industry average. Book a demo through the link below and discover how DXwand can transform your AI development from a crawl to a sprint. https://lnkd.in/gBmYkdFS

AI: Ask the Expert! ⁉️ QUESTION: "Implementing AI for competitive document development comes with its own set of challenges. How can organizations overcome obstacles such as data quality and AI bias?" ANSWER: "Organizations need to be mindful of data quality even more than before. In the big data lakes of old, with their powerful BI tools connected, humans were always in the loop to analyze, interpret and make decisions. When AI is independent, data quality becomes paramount as the human intuition is no longer part of the equation." Don't miss John Ingve Eielsen at the Xait Forum for his thought-provoking and informative Keynote, Cracking the Code of Trust: Securing Your Advantage in the Age of AI. ➡️ Learn how to maximize AI's benefits while minimizing the risks. Secure your spot! https://hubs.la/Q02YVPft0

AI is coming for everything you do that requires intelligence. o3 (which is not on this graph) recently passed the ARC AGI challenge, and I'm sure it will also pass Simple Bench when it's tested on it. That, and this graph (credit to Ethan Mollick, and Dave Shapiro for reposting it) show it's reasoning and generalized intelligence capabilities are growing outside of its training distribution. Prior to a month or two ago, there were many in the AI research community that thought reasoning was going to be a much more difficult and time-consuming problem to solve - it was the bottleneck that us knowledge workers and small business owners were hoping would slow it down. That hope has been shattered. What are you doing to prepare?