Lingua Custodia’s Post

Data Analysis is undoubtably one of the hottest use cases for Generative AI. But it’s also one of the hardest to get right if you want accurate results that you can actually trust. Here’s what most AI tools (including ChatGPT) don’t give you for data analysis: 1. Clean data Garbage in = garbage out. If you don’t provide tables with a well defined schema and clearly defined column names, your LLM will struggle to identify correct columns to use in queries. 2. Context Unlike a human, an LLM can’t always infer context directly from a column name. Passing additional context and metadata is therefore crucial for an LLM to be able to know what columns to use and how to use them to answer your question. 3. Agents Agents became a bit of a buzzword in the AI world, but in Natural Language to SQL they play an important role in improving the accuracy of your queries. Agents help to create a feedback loop that can iteratively improve your generated query based on the results. ------------ Follow for more tips on how you can use AI for data analysis! #ai #llm #dataanalytics

"Data is the fuel and lifeline of any #AI system" Today I am sharing an abstract yet simple example of an #AI system’s lifecycle. Developers and deployers of #artificialintelligence (AI) systems have a variety of distinct responsibilities for addressing risks throughout the lifecycle of the system. These responsibilities range from problem definition to #data collection, labelling, cleaning, #modeltraining and ‘fine-tuning’, through to testing and deployment as shown in this Figure. #datascience #artificialintelligence  #aimodels #datadeployment #adaption source: https://shorturl.at/DLcOq

“Shall we do RAG or fine-tuning?” I've been asked this question very often. One does not exclude the other. There is discussion whether Retrieval-Augmented Generation (RAG) or fine-tuning yield better results when deploying a solution based on Large Language Models (LLMs) The RAG pattern combines retrieval of companies'  information with generation of answers based on that information, through an LLM. This technique is a relatively low hanging fruit, and allows for up to date information. Fine-tuning means further training the model and adjust the LLM's weights to specific domains and tasks, enhancing performance on taking challenges. In both cases, and specially so in fine-tuning, the data pre-processing is an important step that shall not be overlooked. But, these techniques are NOT mutually exclusive and complement each other very well. For instance, a fine-tuned customer support chatbot can utilize RAG to incorporate the latest customer data, delivering contextual and personalized responses. We don't need to choose one over the other. If you need help on this topic, talk to us. #ai #largelanguagemodels #finetuning #rag

LLM vs VLM Comparison? Nopes, The AI Duo You Didn't Know You Needed! Not sure if you should be using LLMs or VLMs (or both)? 👀 Let’s break it down so you can unlock their full potential! LLM (Large Language Model): This AI powerhouse excels at understanding and generating text. From writing reports to answering questions, LLMs are your go-to for anything language-related. 📚🧠 VLM (Vision-Language Model): Here’s where things get cooler. VLMs combine both visual and textual data—perfect for tasks like image recognition, generating captions, or even interacting with visual content. Think of it as AI that can see and speak! 👀🖼️ What’s the difference? LLM = Master of Text, understands language like a pro. VLM = Combines the power of images and text for more versatile applications. Imagine the possibilities when you combine both! 😎 Curious how to make these AI tools work for your project? Scenario: A customer sends an image of a damaged product to a company's chatbot. Solution: The VLM can analyze the image to identify the product and the damage. The LLM can then generate a response, such as suggesting a replacement or offering a refund. That's not all, many things can be done😉 Image source: Snorkel AI ♻️ Repost if you think your fellow AI enthusiasts need to know this! #ArtificialIntelligence #AIApplications #GenAI

RAG: Powering Next-Gen AI 🚀 Have you ever wondered how business personalize LLMs for their own use case? Enter RAG - Retrieval-Augmented Generation. 🔍 What is RAG? RAG combines the power of large language models with external knowledge retrieval. It's like giving AI a constantly updated knowledge base to reference. 💡 How it works: 1. Query understanding: The LLM analyzes the user's input. 2. Relevant info retrieval: It searches its knowledge base for relevant data. 3. Context-based response generation: The LLM crafts a response using both its training and the retrieved context. This process allows AI to leverage vast amounts of current data without constant retraining. 🌎 Real-world applications: - Customer support chatbots: RAG enables bots to access product manuals, FAQs, and recent updates, providing accurate and timely support. - Personalized content recommendation: By retrieving user preferences and behavior data, RAG can suggest highly relevant content across platforms. - Intelligent search engines: RAG enhances search results by understanding context and retrieving the most up-to-date and relevant information. RAG is revolutionizing AI's ability to provide accurate, current, and contextually relevant information. It's not just about having knowledge - it's about knowing how to use it. #AI #MachineLearning #GenerativeAI #RAG #LLM #SoftwareEngineering

What is Multi token Prediction in LLMs? 📖 Imagine reading a book, but instead of reading one word at a time, you glance at a phrase or even a whole sentence. Your brain processes multiple words simultaneously, allowing you to understand the meaning more quickly and efficiently. 🤖 Multi-token prediction works similarly for AI language models. Traditionally, these models predict the next word based on the previous words, like reading a book one word at a time. With multi-token prediction, the model looks ahead and predicts multiple words at once, just like skimming a phrase or sentence ⚡ By processing larger chunks of text in one go, the AI can: > Generate text up to 3 times faster > Better understand and maintain context over longer distances > Improve overall performance, especially for large models It's like giving the AI a bigger piece of the puzzle to solve at once, enabling it to see the bigger picture and make more accurate predictions. This innovative approach has the potential to revolutionize various applications of language models, such as more efficient code completion, enhanced writing assistance, and improved language translation. What potential applications do you see for multi-token prediction in AI? Share your thoughts below! #MachineLearning #DataScience #AI #AIWhisperers #artificialintelligence #AIExplained #AIdevelopment #AIstrategy #AIInteractions #AILiteracy #EffectiveAI #LLM #chatgpt #DataScience #AIModels #DigitalTransformation For more cool AI stuff, check us out

🌟 Day 14 of the '10 Days of Tinkering With AI Agents' Challenge 🤖 Why is Agentic AI poised to be the next big thing? The answers are in the attached image! 🌐 As highlighted by research from DeepLearning.AI and a blog by Andrew Ng, the performance leap is significant when comparing AI models(for generating Code) using the HumanEval coding benchmark: 🚀 GPT-3.5 (zero-shot): 48.1% accuracy 🚀 GPT-4 (zero-shot): 67.0% accuracy 🔥 GPT-3.5 with an iterative agent workflow: Up to 95.1% accuracy! As I’ve been saying: LLM + Iterative Agentic Flow = More Value 💡 Andrew Ng also shared these foundational elements for an Agentic Framework: 𝗥𝗲𝗳𝗹𝗲𝗰𝘁𝗶𝗼𝗻: The LLM examines its own work to find ways to improve it. 🔍 𝗧𝗼𝗼𝗹 𝗨𝘀𝗲: The LLM utilizes tools like web search, code execution, and more to gather information, take action, or process data. 🛠️ 𝗣𝗹𝗮𝗻𝗻𝗶𝗻𝗴: The LLM devises and executes a multi-step plan to achieve its goals (e.g., writing an outline, conducting research, drafting). 📋 𝗠𝘂𝗹𝘁𝗶-𝗔𝗴𝗲𝗻𝘁 𝗖𝗼𝗹𝗹𝗮𝗯𝗼𝗿𝗮𝘁𝗶𝗼𝗻: Multiple AI agents collaborate, splitting tasks, discussing, and debating ideas to find better solutions than a single agent could. 🤝 𝙐𝙥𝙙𝙖𝙩𝙚 𝙤𝙣 𝙢𝙮 𝙇𝙚𝙖𝙙 𝙂𝙚𝙣𝙚𝙧𝙖𝙩𝙞𝙤𝙣 𝘼𝙜𝙚𝙣𝙩: It’s still a work in progress, and I’m seeking some help with it. So far, I’ve managed to achieve the below with Agents and Tasks: Generate Search Queries 🔍 Get the Search Results 📄 Scrape Data from Websites 🕸️ I’ll give it another go tomorrow! 💪 #AI #GenerativeAI #MachineLearning #DeepLearning #LLMs #AgenticAI #ArtificialIntelligence #FutureOfAI #AgenticFlow #mwmusings

Ever wondered which AI is best? John Breeden II explores a cool tool called the LMSYS Org Chatbot Arena that lets you compare AIs head-to-head. Check out the full article to see how the author pitted two AIs against each other for tasks like writing code and epic poems. This is a great resource for anyone working with AI. #government #AI #technology #productivity

Check this out! I'm currently putting the Mistral 8x22B model through some testing and came across this interesting behavior with the base model (non-instruct, think of it like an autocomplete). It has created a Stack Overflow question and answer. Fascinatingly, the model has included a warning about posting AI-generated answers to Stack Overflow. This suggests Mistral AI is actively scraping the website for recent data, including the warning that was recently placed on the site about AI-generated content. Ironically, while this is a strong, coherent answer in terms of content and structure, the model seems to inherently recognize patterns in its own output that resemble AI-generated content. Because Stack Overflow now includes warnings on such content, the next tokens the model predicts are the warning itself - suggesting an impressive level of pattern recognition and contextual awareness. It's almost as if the model is "self-aware" enough to recognize its own "signature" in the generated text and predict the appropriate cautionary notice that would likely be appended to it on the site. This raises fascinating questions about just how attuned language models are becoming to the nature and structure of their own outputs. Do you think this level of pattern recognition and self-referential prediction is a significant milestone in AI's ability to understand and emulate human-like online content moderation? Can models be considered "self-aware" in a meaningful sense if they're able to flag their own generated text as likely to be labeled as AI-generated by human moderators? I'm really curious to hear others' thoughts and reactions! Let me know what stood out to you and where you think this technology is headed. #ai #ml #agi

𝗔𝗜 𝗜𝗻𝗳𝗲𝗿𝗲𝗻𝗰𝗲: 𝗡𝗮𝘃𝗶𝗴𝗮𝘁𝗶𝗻𝗴 𝗖𝗼𝗻𝘁𝗲𝘅𝘁, 𝗦𝗽𝗲𝗲𝗱, 𝗮𝗻𝗱 𝗞𝗻𝗼𝘄𝗹𝗲𝗱𝗴𝗲 𝗜𝗻𝘁𝗲𝗴𝗿𝗮𝘁𝗶𝗼𝗻 Last week, I explained some foundational concepts of GenAI. This week, I continue with inference. Inference is the process of using a trained model to generate outputs based on new input data. While training large language models (LLMs) is computationally intensive, inference is the magic that happens when a model is being put in use. Three related concepts that influence the inference performance. 𝗖𝗼𝗻𝘁𝗲𝘅𝘁 𝗪𝗶𝗻𝗱𝗼𝘄 The context window is the maximum amount of input data (usually measured in tokens, roughly equal to the number of words) that a model can process in a single inference step. A larger context window allows the model to consider more information when generating outputs but often comes at the cost of slower inference speeds. 𝗟𝗮𝘁𝗲𝗻𝗰𝘆 Latency is the time it takes for a model to generate an output given an input. Low latency is critical for user experience and system responsiveness, especially when hosting solutions for many users. Factors that impact latency include model size, context window, hardware resources, and network conditions. 𝗥𝗲𝘁𝗿𝗶𝗲𝘃𝗮𝗹-𝗔𝘂𝗴𝗺𝗲𝗻𝘁𝗲𝗱 𝗚𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝗼𝗻 (𝗥𝗔𝗚) RAG combines LLMs with external knowledge retrieval to expand the model's context beyond its fixed context window. RAG systems use vector databases to store and retrieve relevant information based on the input query, allowing them to handle tasks that require access to large amounts of external knowledge. However, RAG systems typically have higher latency compared to pure in-context learning approaches due to the additional retrieval step. 𝗖𝗼𝗻𝗰𝗹𝘂𝘀𝗶𝗼𝗻 Understanding the interplay between context window, latency, and retrieval-augmented generation is essential for making informed decisions when deploying AI solutions. By aligning these factors with your business needs and user expectations, you can ensure that your AI system delivers the right balance of performance, responsiveness, and scalability. #AI #GenAI #LLM ___ Enjoyed this post? Like 👍, comment 💭, or re-post ♻️ to share with others.