Sid Bhattacharya’s Post

It's hard to understand how "predicting the next word" is the mechanism behind the miraculous capabilities of Large Language Models (LLM) like GPT-4, Gemini and Claude to interact at length in sophisticated ways indistinguishable from how humans write and converse. This also makes it difficult to accept AI inputs to medical decision making. I summarise interesting research that illuminates some of the inner workings of LLMs, and the brain 🧠, while invoking the insights of an ancient Greek philosopher. If "explainability" increases, will acceptability in clinical medicine also? https://lnkd.in/dPhYmFvr

Just a moment...: Recent AI models like GPT-3 are linguistically adept but often lack accuracy, posing challenges in critical fields. To improve this, researchers are exploring external memory stores for better fact-based grounding. Soman et al.'s 2023 study suggests methods for optimizing knowledge retrieval and integration in domain-specific models. - Artificial Intelligence topics! #ai #artificialintelligence #intelligenzaartificiale

𝐁𝐨𝐧𝐢𝐭𝐨 𝐋𝐋𝐌 - 𝐈𝐧𝐬𝐭𝐫𝐮𝐜𝐭𝐢𝐨𝐧 𝐓𝐮𝐧𝐢𝐧𝐠 𝐃𝐚𝐭𝐚𝐬𝐞𝐭𝐬 𝐆𝐞𝐧𝐞𝐫𝐚𝐭𝐢𝐨𝐧 Bonito is an open-source model developed to convert unlabeled text into instruction-tuning datasets. This model is trained on a novel dataset comprising 1.65 million examples, developed by remixing existing instruction-tuning datasets into meta-templates, which then produce training examples from unannotated text. Bonito is used to generate synthetic data across tasks like yes-no question answering, extractive question answering, and natural language inference—for adapting language models to specialized domains. Bonito model significantly enhances the performance of both pretrained and instruction-tuned models, showing a notable improvement of 22.1 F1 points over the baseline. Bonito paper - https://lnkd.in/g4Aea5M3 #llms #instructiontuning #generativeai #ai #nlproc #deeplearning

🚀 Exposing Fundamental Weaknesses in State-of-the-Art LLMs! 🤖✨ A recent study, titled "Alice in Wonderland: Simple Tasks Showing Complete Reasoning Breakdown in State-Of-the-Art Large Language Models," reveals significant deficiencies in the reasoning capabilities of current Large Language Models (LLMs), despite their high performance on standardized benchmarks. 🔍 Key Findings: 1️⃣ Representational Collapse: The study demonstrates that even the most advanced LLMs struggle with basic reasoning tasks that humans solve effortlessly using common sense. This collapse is particularly evident in tasks such as copying and counting. 2️⃣ Overconfidence and Confabulations: LLMs frequently provide incorrect answers with strong overconfidence, generating plausible-sounding but nonsensical explanations to justify their wrong solutions. 🧠 The AIW Problem: The researchers introduced the "Alice in Wonderland" (AIW) problem: "Alice has N brothers and she also has M sisters. How many sisters does Alice's brother have?" While the correct answer (M+1) is straightforward for humans, most LLMs, including GPT-3.5/4, Claude, Gemini, LLaMA, Mistral, and others, failed to answer correctly, often providing absurd responses. 📈 Performance Highlights: - AIW+ Variation: A harder version of the AIW problem, called AIW+, caused an even stronger performance collapse across all tested models, including GPT-4 and Claude 3 Opus. - Standard Benchmarks' Shortcomings: The study highlights a stark discrepancy between LLMs' high scores on standardized benchmarks (e.g., MMLU, ARC, Hellaswag) and their poor performance on the AIW problem, suggesting that current benchmarks do not adequately reflect models' true reasoning capabilities. 💡 Why It Matters: - Improving Benchmarks: The study emphasizes the need for the ML community to develop new reasoning benchmarks that can accurately detect such deficits and guide the improvement of LLMs' reasoning skills. - Transparency and Reproducibility: It calls for fully open and reproducible training pipelines, including dataset composition, to enable proper analysis and progress in enhancing LLMs' reasoning capabilities. Stay tuned for more updates on this crucial AI development! 🚀 📊 PAPER: https://lnkd.in/eyg8wm8B #AI #LLM #MachineLearning #TechInnovation #DeepLearning #Reasoning #DataScience #Benchmarking

Enhancing Efficiency in AI: The Power of Prompt Compression in Language Models Exciting insights into the world of large language models (LLMs) and the importance of prompt compression for enhanced efficiency and cost-effectiveness! Here is Microsoft framework called Llmlingua 1. Why Focus on Prompt Compression? • Long prompts in LLMs lead to higher costs and inefficiency. Learning to compress these prompts is a game changer in the AI world. 2. Two Innovative Techniques: • Coarse-Grained Compression: Summarizing content and using perplexity for evaluating summarization quality. • Fine-Grained Compression: Streamlining text by removing unnecessary tokens while keeping the core meaning intact. 3. Tools for Optimization: • Discover how tools like LLMLingua and Llama index make prompt compression more effective and intuitive. 4. Measuring Success: • Perplexity becomes a key metric in determining the quality of our summarization efforts. 5. Practical Benefits: • Understand how reducing prompt length can significantly cut down on the costs associated with using LLMs. Paper: https://lnkd.in/gs4UEfM6 Code: https://lnkd.in/gEmCAFZE This knowledge is invaluable for anyone in the field of AI and machine learning, offering practical ways to enhance the use of language models. Thoughts? Let’s discuss how these techniques can reshape our approach to AI! #AI #MachineLearning #TechInnovation #DataScience #Emids

The world of Large Language Models (LLMs) is moving at lightning speed! It's almost embarrassing to highlight a review paper from February, but trust me, it's still a gem worth your time. 📚✨ Last year the leaps in LLM progress were gigantic, but lately the progress seems to be more nuanced and steadier, so this comprehensive review of the different LLM families (with GPT, LLaMa and PaLM as their representative models) is still worth a read. The authors of the paper (Shervin Minaee and his colleagues) put them in a bit of fascinating historical context. 🕰️✨ From the humble beginnings of statistical language models that used n-gram models to estimate word probabilities, to the early neural language models that mapped words to low-dimensional vectors, to our current state-of-the-art pre-trained models that have revolutionised the field. 🌐🔍 The paper does an excellent job of reviewing how the three LLM families are built, used, and extended for real-world applications. For my AI-savvy followers, it also delves into the datasets and benchmarks used to evaluate LLM performance. 📊🧠 It concludes with another section aimed at everyone, exploring the current challenges that need to be addressed and future research directions. One of the key takeaways, of course, is the ongoing challenge of improving the efficiency of LLMs. (Perhaps we are not starting to see smaller, more specialised models addressing specific use cases. And, of course, the authors highlight the exciting work being done to reduce hallucinations through advanced augmentation techniques - a topic I'll be exploring in a future post! 🤖💡 At just 36 pages, this paper is a quick read as it is packed with insightful figures and tables. Check out this amazing tree of current LLM skills I borrowed from the paper! 🌳📈 So if you've got some spare time and want to stay ahead in the AI game, give it a read: https://lnkd.in/e9-VjVXA Happy reading! 📖✨ #AI #MachineLearning #LLMs #GPT #LLaMa #PaLM #Innovation #FutureOfAI #TechTrends

A Guide to Mastering Large Language Models - Large language models (LLMs) have exploded in popularity over the last few years, revolutionizing natural language processing and AI. From chatbots to search engines to creative writing aids, LLMs are powering cutting-edge applications across industries. However, building useful LLM-based products requires specialized skills and knowledge. This guide will provide you with a comprehensive yet accessible overview of the key concepts, architectural patterns, and practical skills needed to effectively leverage the huge potential of LLMs. What are Large Language Models and Why are They Important? LLMs are a class of deep learning models that are pretrained on massive text corpora, allowing […] - https://lnkd.in/et9JHxgF

The GNN-RAG model merges the language understanding skills of LLMs with the reasoning capabilities of GNNs in a RAG format. This innovative approach enhances vanilla LLMs on KGQA, even outperforming GPT-4 with a 7B tuned LLM while retrieving essential multi-hop information. The integration of tree and graph-based methods with LLMs is a promising trend in advancing AI reasoning abilities. #AI #MachineLearning #Research #GNNRAG https://lnkd.in/ewqRtHjG

⭐ Today's Newsletter Andrej Karpathy LLM Paper Reading List for LLM Mastery ➡ Link: https://lnkd.in/g8XEKR9i Andrej Karpathy is known for his significant contributions and leadership at the intersection of machine learning and language, therefore Karpathy’s paper reading list serves as a compass for those seeking to unravel the intricacies of LLMs. This blog outlines the papers recommended by Karpathy. It provides insightful commentary and context, offering readers a deeper understanding of the groundbreaking concepts and methodologies that have been pivotal in shaping the landscape of Language Models. From the foundational principles outlined in “Attention is All You Need” to the latest advancements showcased in “Sparks of Artificial General Intelligence: Early experiments with GPT-4,” this journey through the reading list promises to be a riveting exploration of the frontiers of AI research. As the authors navigate through each section of the blog, they will uncover the key insights and advancements encapsulated within the recommended papers, shedding light on the evolution of LLMs and their role in the broader field of artificial intelligence. Join us in this intellectual voyage as they decode the wisdom curated by Andrej Karpathy, providing both novice and seasoned AI enthusiasts with a roadmap to LLM mastery.

Let's get down to work! Transcending the Image-Text Boundary is now a pre-requisite for any Generative AI specialist. Multimodal Generative AI is expanding. The attached free, open-source notebook(GitHub link in the comment section) illustrates this through Stable Diffusion. Understanding how Stable Diffusion works has become an essential AI skill along with LLMs. Let me break this down for you: 1. LLMs have attained superhuman functional levels. 2. There is a spillover of Generative AI into computer vision. 3. In Transformers for Natural Language Processing and Computer Vision, 3rd Edition(available on Amazon and link in the comment section), I explain the Stable Diffusion code behind the attached notebook in detail.  4. The detailed explanation of the code in the chapter will help you understand the code you can run in the attached free open-source notebook. Bottom line: The real-life project managers of the Generative AI market are moving away from discovery and into multimodal implementations. Please ask me any questions you wish in the comment section. #generativeai #computervision #ai