Basant Narayan Singh

I have posted about how ChatGPT makes mistakes on simple questions if they are phrased as puzzle. This is called template fixation where it may get the right answer, but logic is wrong. Link: https://lnkd.in/gfKEzRJ6 ChatGPT: https://lnkd.in/gbDVdQs8 #llms #genai #generatieveai #capgemini #llm #opensourceai #capgeminiindia #ai #artificialintelligence #software

15

Must attend workshop. For those who would like to learn more about fine-tuning LLM’s RAG: retrieval augmented generation is to do with grounding the LLM in what is relevant (vector databases) For example: a question like give me our company’s sales numbers in Argentina should not yield a lot of junk info by pulling out anything to do with the country, Argentina over the past 200 years

3

Recently read an interesting paper from Google Deepmind on some key techniques for Fine Tuning a Large Language model, should we just fine tune a model or give relevant examples. Does providing examples like few shot and few-shot in-context learning (ICL) hurt or improve performance. Some recommendations include: 1. LLM performance increases when going from few examples (1-5) to many (100's+) examples (many -shot ICL) 2. Propose using synthetically generated examples 3. The order of examples also influences many-shot performance 4. Reduction in Biases also seen with many shot ICL Link to the paper: https://lnkd.in/gx-ufU5f Good read. #LLM #GenAI #FineTuning #NLP #Analytics

5

Here is a conversation with the very passionate Gaurav Aggarwal on his experiences of building a deep-tech AI company from India - https://lnkd.in/gA6-Qjja I can relate to a lot of this(dearth of ambitious capital and people with very high risk appetite) as I tried to build a deep-tech company from India 5 years back.

36

India is behind the AI Innovation curve by miles. And this innovation gap will keep widening unless we see multifold investments in Compute Infrastructure and R&D. To catch up, we need to prioritize investments in compute infrastructure and research. As Large Language Models (LLMs) become increasingly commoditized, countries with robust compute infrastructure and R&D capabilities will continue to lead the way and Bharat will again left to be just the largest consumer. If you are interested in this space, let's connect. Image credit - Llama 3.1 70B Aviral Bhatnagar Sumit Agarwal Rakesh Dubey Meghdut RoyChowdhury #Innovation #compute #R&D #Llama #ai #chatgpt #openai #India

17

🚀 𝗧𝗵𝗲 𝗘𝘃𝗼𝗹𝘂𝘁𝗶𝗼𝗻 𝗼𝗳 𝗥𝗔𝗚 𝗧𝗲𝗰𝗵𝗻𝗶𝗾𝘂𝗲𝘀: 𝗘𝗻𝗵𝗮𝗻𝗰𝗶𝗻𝗴 𝗥𝗲𝘁𝗿𝗶𝗲𝘃𝗮𝗹 𝗮𝗻𝗱 𝗚𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝗼𝗻 🔎 In the domain of Retrieval-Augmented Generation (RAG), the standard practice often involves a singular retrieval step followed by generation. While effective, this method can be inefficient and insufficient for complex problems demanding multi-step reasoning. To address these challenges, various studies have optimized the retrieval process, leading to more sophisticated techniques. Let's explore the most popular ones: 🔄 𝗜𝘁𝗲𝗿𝗮𝘁𝗶𝘃𝗲 𝗥𝗲𝘁𝗿𝗶𝗲𝘃𝗮𝗹: This technique alternates between retrieval and generation, allowing for richer and more targeted context from the knowledge base at each step. 🔁 𝗥𝗲𝗰𝘂𝗿𝘀𝗶𝘃𝗲 𝗥𝗲𝘁𝗿𝗶𝗲𝘃𝗮𝗹: This method involves gradually refining the user query and breaking down the problem into sub-problems, continuously solving complex issues through repeated cycles of retrieval and generation. 🤖 𝗔𝗱𝗮𝗽𝘁𝗶𝘃𝗲 𝗥𝗲𝘁𝗿𝗶𝗲𝘃𝗮𝗹: This approach enables the RAG system to autonomously determine when external knowledge retrieval is necessary and when to stop retrieval and generation. It often uses special tokens generated by LLMs for control. These advanced techniques are especially relevant in production environments where the volume of documents necessitates a more strategic approach. Embracing these innovations can lead to significant improvements in handling complex queries and providing more accurate, context-rich responses. #AI #MachineLearning #RAG #GenAI

81

2 Comments

"The AI Imperative: Essential Advances in Quality Engineering" 🔥 by Nagabhushan Ramappa   Writers often go through a cycle of drafting and discarding, refining their thoughts until they crystallize into words on the page. What if LLMs mimic this behavior ? Nag had some interesting content which i can relate to (not just) this ! 𝐖𝐡𝐚𝐭 𝐄𝐧𝐭𝐞𝐫𝐩𝐫𝐢𝐬𝐞 𝐄𝐱𝐩𝐞𝐜𝐭𝐬 𝐟𝐫𝐨𝐦 𝐏𝐫𝐨𝐥𝐢𝐟𝐞𝐫𝐚𝐭𝐢𝐨𝐧 𝐨𝐟 𝐓𝐨𝐨𝐥𝐬 Enterprises expect a surge in specialized SaaS productivity tools. They foresee a need for an automated, end-to-end product reliability center, ensuring that from development completion to end-user delivery, every step is autonomous and seamless.    𝑮𝒓𝒐𝒘𝒊𝒏𝒈 𝑺𝒐𝒇𝒕𝒘𝒂𝒓𝒆 𝑪𝒐𝒎𝒑𝒍𝒆𝒙𝒊𝒕𝒚  Post-COVID, companies have accelerated software development, introducing more features in one year than in the previous twelve. Managing this complexity is crucial, as errors/failures can lead to escalations.   𝘛𝘳𝘢𝘥𝘪𝘵𝘪𝘰𝘯𝘢𝘭 𝘘𝘈 𝘢𝘯𝘥 𝘮𝘢𝘯𝘶𝘢𝘭 𝘵𝘦𝘴𝘵𝘪𝘯𝘨 𝘢𝘳𝘦 𝘪𝘯𝘤𝘳𝘦𝘢𝘴𝘪𝘯𝘨𝘭𝘺 𝘪𝘯𝘢𝘥𝘦𝘲𝘶𝘢𝘵𝘦 𝘧𝘰𝘳 𝘵𝘩𝘪𝘴 𝘨𝘳𝘰𝘸𝘪𝘯𝘨 𝘤𝘩𝘢𝘭𝘭𝘦𝘯𝘨𝘦.   𝑹𝒐𝒍𝒆 𝒐𝒇 𝑮𝒆𝒏𝒆𝒓𝒂𝒕𝒊𝒗𝒆 𝑨𝑰 Generative AI can automate testing ( or any other tasks IMHO) and streamline workflows, enhancing efficiency and reducing errors in complex software development. This is where MAGS help us in stitching and unifying multiple workflows.   𝑴𝑨𝑮𝑺 - 𝑴𝒖𝒍𝒕𝒊-𝑨𝒈𝒆𝒏𝒕𝒊𝒄 𝑮𝒆𝒏𝒆𝒓𝒂𝒕𝒊𝒗𝒆 𝑺𝒚𝒔𝒕𝒆𝒎  MAGS utilizes multiple agents, each specialized in a specific task, working independently and sequentially. Let take an example of using GenAI In essay writing.   A non-agentic workflow produces an essay in a single effort - no backspace. Whereas, an agentic workflow iteratively refines the essay, with various agents conducting web research, evaluating content, and making improvements behind the scenes, ensuring a better final product.   Applying MAGS to complex software testing workflows, we can leveraging specialized agents to automate testing processes, ensuring thorough coverage and reliability in software development cycles.   𝗔𝗜 𝗶𝘀 𝗳𝗹𝗲𝘅𝗶𝗯𝗹𝗲 - 𝗶𝘁 𝗰𝗵𝗮𝗻𝗴𝗲𝘀 𝗮𝘀 𝗼𝘂𝗿 𝘄𝗼𝗿𝗸 𝗼𝗯𝗷𝗲𝗰𝘁𝗶𝘃𝗲𝘀 𝗰𝗵𝗮𝗻𝗴𝗲𝘀 Distributed intelligence systems are advancing and evolving from deterministic workflow tasks and freestyle chat to goal-based objectives. They are maturing from embedded AI and workflow-based approaches to case flow and self-organizing workflows.  As Nag explained these concepts, Amit Kumar Das balanced off with a Demo on the subject swiftly. I was amazed looking at the logs and trying to infer what was going on behind the scenes.   Nag - I loved the way the session went and how the demo were organized. Thanks for bringing this (new) awareness on AI stuff once again. Very informative session ! P.S. Have a look into the comments , for some questions Nag had for us ( i am yet to figure tis out even after a yr ;)) #TribeQonfReporter #TheTestTribe #TribeQonf #AI #saiprasadsabeson

37

2 Comments

Here is my column on the #TiEcon2024 conference in #indicanews:

59

4 Comments

In the span of just a couple of decades, there has been a remarkable evolution in software development. Gone are the days when displaying a simple window on the screen required a bunch of lines of code. Today, as Satya Nadella said, with GitHub Copilot Workspace developers can now convert their ideas into code into software instantly in natural language. This empowers developers and increases their productivity. It allows them to focus on solving actual problems to build real solutions. Kudos to Microsoft! Microsoft Genai ChatGPT https://lnkd.in/g7WbazHy

3

Totally agree with you, Anthony. The name of the game is to get the right context as tight as possible to the question or task on hand. Graphs are one good way to do this, but you should use all the tools in your belt. With our solution, we use graphs, tables, key value pairs, geospatial information, regular search, vectors, and so on. When you combine all these things, the probability of getting the right context as much higher.

5

It was an incredible experience, taking part in the nasscom insights show sharing our vision of why we are building Hyperleap AI, and how businesses can benefit big time in terms of reduced costs, and faster time to market with our Generative AI platform. Thank you so much Ms. Namita Jain and team for hosting me - it was an honor.🙏 Thank you Mr. Madhumay for your constant support and guidance, and for various strategic initiatives that are helping startups get the visibility on a global stage. 🙏 If you're building Generative AI into your applications or workflows, it should be simple, but it isn't. This is either because your integration is good as a PoC but cannot scale to production or because you've combined a myriad of different tools that must all work in unison so you can deliver. Hyperleap AI solves it all, with an end-to-end platform that works seamlessly so you can focus on delivering use cases. With our product, businesses can light up Generative AI use cases in a matter of days. 🚀 #HyperleapAI #GenAI #Platform #PlaygroundToProduction #GenerativeAI #Innovation T-Hub MATH - ML & AI Technology Hub nasscom nasscom ai nasscom startups India DST Startup India

23

It's always good to reconnect with old friends of BHyve! With Kunjal Kamdar it's special because he wrote about us when we were literally a set of ideas in a PowerPoint! Cut to today, his views confirm so many of our learnings over these last 3 years, our understanding of the market's needs, and the journey we chart ahead. 🚀 #BuildingBHyve #GenerativeAI #Knowledgemanagement #reimaginewhatspossible

14

1 Comment

Claude 3.5 Sonnet has now shown up on the LMSYS Chatbot Arena Leaderboard (https://lnkd.in/gDjMe_8f), coming in at number 2, slightly behind GPT-4o. 1. GPT-4o-2024-05-13: 1287 Elo 2. Claude 3.5 Sonnet: 1272 3. Gemini-Advanced-0514: 1267 (statistical tie with Claude 3.5 due to confidence interval) While there are more and more ways to benchmark LLMs, I still think the Chatbot Arena is tops as it is based upon a large number of head-to-head comparisons with human judges. Anyone can contribute, the battle arena allows you to prompt and then judge the output anonymously. Here's some of my takeaways here: - The tight range of Elos at the top means that in many prompts there's not consistent winner. For example, in GPT-4o vs Gemini, GPT won 51% of the time vs. 49% for Gemini -- when there was a winner at all. - Even rank 12 Llama 3 70b (Elo 1207), won 33% of the time vs. GPT-4o. - This means while at the very top there's not a lot of difference, if you are using the LLM extensively then you are giving yourself a lot of "runs" at being able to see differences, making the more than a few points of difference obviously quickly enough. In chess, where Elo ratings come from, you aren't going to play 500 games against someone, but you may well make 500 prompts against your LLM. This can make Elos a little counter-intuitive... as in chess a 1290 player and a 1270 player wouldn't be considered a significant difference, but in LLMs it is. The relatively low Elos of today's LLMs also don't mean anything about how smart they will get. Just because Magnus Carlsen has a mind boggling 2830 chess Elo that doesn't say anything about growth potential for LLMs, it just means he has been incredibly dominant against those who themselves have been incredibly dominant.

5

MIT researchers have developed a new approach called Natural Language Embedded Programs (NLEPs) that enables large language models like GPT-4 to solve complex tasks with greater accuracy and transparency. With NLEPs, the model generates a step-by-step program in Python code to solve a user's query, and then outputs the solution as natural language. This method achieves higher accuracy on a wide range of reasoning tasks and offers improved efficiency. The benefits of NLEPs include: Improved data privacy: local execution reduces the need for sensitive user data transmission Better performance from smaller models: no need for retraining for specific tasks According to Hongyin Luo, co-lead author of the paper, "There is no magic here. We do not have a more expensive or fancy language model. All we do is use program generation instead of natural language generation, and we can make it perform significantly better." This technique has the potential to improve AI development and decision-making processes. . . #NLEPs #genai #llm #ai #ml https://lnkd.in/gFZ82KBr

2

2 Comments

In today's world of AI, it's not just about getting the right knowledge but also about giving clear, factual responses. 🤖✨ This is where Retrieval-Augmented Generation (RAG) shines, making AI responses more accurate and up-to-date by combining the power of large language models with external data sources. 📚🔍 Why is evaluating RAG systems important? 🚀 𝗣𝗲𝗿𝗳𝗼𝗿𝗺𝗮𝗻𝗰𝗲 𝗢𝗽𝘁𝗶𝗺𝗶𝘇𝗮𝘁𝗶𝗼𝗻: Find the system's strengths and weaknesses in search and generation to make it better. ✅ 𝗥𝗲𝗹𝗶𝗮𝗯𝗶𝗹𝗶𝘁𝘆 𝗔𝘀𝘀𝗲𝘀𝘀𝗺𝗲𝗻𝘁: Ensure the AI provides correct and reliable information. 😊 𝗨𝘀𝗲𝗿 𝗘𝘅𝗽𝗲𝗿𝗶𝗲𝗻𝗰𝗲: Better metrics lead to happier, smoother encounters. 📊 𝗖𝗼𝗺𝗽𝗮𝗿𝗶𝘀𝗼𝗻: Compare your system to others or previous versions. ⚖️ 𝗘𝘁𝗵𝗶𝗰𝗮𝗹 𝗖𝗼𝗻𝘀𝗶𝗱𝗲𝗿𝗮𝘁𝗶𝗼𝗻𝘀: Identify and fix biases or false information. K͟͟e͟͟y͟͟ ͟͟M͟͟e͟͟t͟͟r͟͟i͟͟c͟͟s͟͟ ͟͟f͟͟o͟͟r͟͟ ͟͟E͟͟v͟͟a͟͟l͟͟u͟͟a͟͟t͟͟i͟͟o͟͟n͟͟:͟͟ 𝗥𝗲𝘁𝗿𝗶𝗲𝘃𝗲𝗿-𝗙𝗼𝗰𝘂𝘀𝗲𝗱: 🔍 Retrieval Accuracy: How successfully does the system find the correct data? 📌 𝗥𝗲𝗹𝗲𝘃𝗮𝗻𝗰𝗲: Does the query-relevant information appear? 🌐 𝗦𝗰𝗮𝗹𝗮𝗯𝗶𝗹𝗶𝘁𝘆: Can it handle several requests on diverse topics? 𝗚𝗲𝗻𝗲𝗿𝗮𝘁𝗼𝗿-𝗙𝗼𝗰𝘂𝘀𝗲𝗱: 📝 𝗚𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝗼𝗻 𝗤𝘂𝗮𝗹𝗶𝘁𝘆: Are the answers easy to understand? 🧠 𝗙𝗮𝗰𝘁𝘂𝗮𝗹 𝗖𝗼𝗿𝗿𝗲𝗰𝘁𝗻𝗲𝘀𝘀: Is the data being created correct? 📖 𝗞𝗻𝗼𝘄𝗹𝗲𝗱𝗴𝗲 𝗼𝗳 𝗖𝗼𝗻𝘁𝗲𝘅𝘁: Does it make good use of and understand the context? 𝗜𝗻𝗰𝗹𝘂𝗱𝗶𝗻𝗴 𝗕𝗼𝘁𝗵: 🔗 𝗙𝗮𝗶𝘁𝗵𝗳𝘂𝗹𝗻𝗲𝘀𝘀: Does the created material match the obtained data? 📋 𝗥𝗲𝗹𝗲𝘃𝗮𝗻𝗰𝗲: Does the output answer the question? This is a high-level overview, and we do have a lot of specialized frameworks and techniques to evaluate these systems. #genai #RAG #learning

13

1 Comment

Story Time: The Bahubali Analogy for Generative AI! Many of you reached out intrigued by the snippets I've shared about using the epic tale of Bahubali to explain the dynamics of Generative AI. Your curiosity has inspired me to weave this story in full color for all of us here on LinkedIn. Imagine a setting where each scene of Bahubali translates into a unique application of Generative AI: - In an intellectual meeting, just as Bahubali displays the virtues of leadership and moral judgement, a Generative AI model articulates these ideals through thoughtful essays or comprehensive blogs, unraveling complex social issues with the same depth and wisdom. - At a lively social gathering among friends, we marvel at the thrilling, physics-defying stunts of Bahubali. This excitement is akin to the way Generative AI crafts stories or scripts that defy the mundane, injecting a spark of creativity and wonder akin to Bahubali’s battlefield feats. - In a quiet corner speaking with a love interest, the tender moments between Bahubali and Devasena unfold. Here, Generative AI mirrors this romance by composing poignant poems or heartfelt letters, capturing the essence of human emotions with surprising delicacy and precision. Each context, like a scene from our beloved movie, showcases the adaptive prowess of AI, echoing the narrative’s shift from valor to romance, from awe to reflection. Just as Bahubali adapts to the needs of his kingdom, so does Generative AI adapt to the varied demands of its users—whether it's drafting an academic piece, scripting a fantasy, or penning a love note. The magic of AI, much like the magic of Bahubali, lies in its versatility and the seamless transition from one form to another, enchanting us with its capacity to not just imagine but to recreate and inspire. Thank you all for your enthusiastic responses. #GenerativeAI #Storytelling #Bahubali #AIinAction #CreativeTechnology

15

Thought of a use case of *Personalized Blogging with #AIAgents* https://lnkd.in/dgtSu5GE https://lnkd.in/d2gCdBkw Let me know your thoughts from a technologist's perspective as to what functionality you can think of showcasing here as well as end-users about what more features you would like to see here. Would continue posting my updates on https://lnkd.in/d9Dd6DAM

7