Michael Pan’s Post

Have AI LLMs reached a local plateau as they apparently cannot get much better by just feeding them more data? “Recent studies have found that as AI models are fed more data and get larger, they don’t get broadly better but get better at specific tasks at the cost of their broader application.” “There are possible solutions to this, like optimizing how AIs are built to reduce the training data needed, running multiple AIs together, or implementing new computing architecture to make AI infrastructure far more efficient. However, all of these solutions are in their infancy and are years away from being deployable. What’s more, these solutions only kick this issue down the line, as all they do is make AI marginally more efficient with its energy and data, and they also don’t solve the issue of where these companies will get more fresh, high-quality data in the future.” https://lnkd.in/gwxeJawT

💎 The ability of processing specialized data is one of the touchstone criterion for building a truly intelligent AI. At INF AI, we pushed the boundaries of innovation with INF-34b by focusing on specialized, high-quality data. 🌍 The secret? A smarter pipeline. Our team develop a data retrieval pipeline that targets high-quality domain-specific data. Taking code, math, and educational data as examples, it starts from carefully identifying and refining domain-specific content—code repositories, math data, and wiki-style educational data - from professional platforms such as Common Crawl Foundation datasets and GitHub repositories and more. Through iterative rounds, we refine and expand this dataset amassing millions of samples. ✅ The result? This pipeline contributes substantial, quality data to the language model, enhancing its ability to understand technical, mathematical, and educational contexts in multiple languages. Ultimately, INF-34b is an AI that doesn’t just respond—it solves and explains, and adapts to the complexities of specialized fields. ⭐ AI isn’t just evolving—it’s specializing. And that’s where the real value lies. #AI #Innovation #MachineLearning #FutureOfTech

The biggest challenge today to develop AI agents is it's limited capacity to reason like humans. We at Ovonts are solving this problem through a reasoning engine powered by neuromorphic principles. We have been experimenting with a set of capabilities to enhance reasoning abilities of AI agents for real world use cases, the results have been interesting. We as humans do not simply generate a content, a solution or do a task in a single shot but take an iterative approach when we solve problems or execute tasks, AI agents should be able to do that to power true impactful human-computer interaction experience. We at Ovonts are making that happen one step at a time. Get a detailed understanding on reasoning capabilities of AI agents on our latest blog post. We would love to understand your perspective, use cases and challenges building the future of insight-led AI applications. #ai #aiadoption #aiapplications #reasoningai #generativeai

Our view at SiliconANGLE & theCUBE is that #LLMs alone are insufficient to power the future of #AgenticAI. More advanced AI models must complement LLMs to deliver the value of AI agents over task-oriented AI assistants. What is your view? We believe that differentiation will come "below the waterline" in the #AI model ecosystem composed of four specialized classes of models: DOMAIN INTELLIGENCE MODELS that understand what may happen within a specific domain based on historical events. #SLMs will encode predictive intelligence from domain-native datasets by fine-tuning domain-specific features, variables, and prompts. These domain models can also integrate relevant content, terminology, and policy on the fly to augment general-purpose #GenerativeAI services. SEMANTIC REASONING MODELS that understand relationships, workflows, and contextual knowledge. #KnowledgeGraphs (structured) or #GraphNeuralNetworks (statistical) will encode semantic relationships and meanings among entities in a dataset and leverage graph structures to improve data retrieval, generating more accurate, context-rich responses. CAUSAL REASONING MODELS that understand why things happen and why certain actions are better than others. #CausalAI models will encode the causal mechanisms within a system and infer cause-and-effect from newly retrieved real-time information. #CausalNeuralNetworks will be used to discover complex causal relationships in large, high-dimensional, or stochastic domains, such as financial markets or transportation. To learn more about causality in Agentic AI, watch this podcast https://lnkd.in/euUuwt-G or read this research https://lnkd.in/gu2MzYD4 LEARNING LOOP MODELS that build collective intelligence across the model ecosystem by orchestrating how models learn from each other in a shared environment, with each optimizing its own actions while learning from the specialized learnings of others (domain relationships, causality, etc.). They'll employ a mix of reinforcement learning methods to accomplish this. While we don't believe these holistic agentic AI model architectures will materialize overnight, we do see ample evidence that they will be developed progressively over time. Comment below to share your views and help us shape this PoV!!

Ready to elevate engineering with #EngineeringIntelligence? 🚀 Engineering Intelligence enhances your tools by integrating and connecting data throughout the product lifecycle, serving as an intelligent layer rather than a replacement. SPREAD AI’s Engineering Intelligence Graph merges different #AI technologies—such as Graphs, Large Language Models (#LLMs), and advanced AI algorithms—into a unified data model representing the complete mechatronic system. Discover more by watching our video ⤵️

I've published a new article on LangChain Chains - your secret weapon for creating powerful, multi-step AI applications. Here's what we'll cover: ✅ Simple Chains: Learn how to interact with LLMs (like GPT-4) for tasks like text generation and summarization. ✅ Sequentional Chains: Combine multiple LLMs to create sophisticated workflows for question-answering, data analysis, and more. ✅ Custom Chains: Unleash your creativity and build your chains for unique applications. ✅ Transform Chains: for modifying or processing the output of one chain before passing it to another chain. ✅ Chain: dynamically selects which chain to use based on its input. To learn more, click the link below to access the full tutorial (with code examples!) published in AI Advance Publication. #langchain #generativeai #llms

Choosing the right inference strategy is crucial when integrating machine learning into applications. Three must-know inference strategies are discussed below: - 𝐎𝐟𝐟𝐥𝐢𝐧𝐞 𝐁𝐚𝐭𝐜𝐡 𝐏𝐫𝐨𝐜𝐞𝐬𝐬𝐢𝐧𝐠: Runs models continuously for all entities, using tools like Airflow, Dagster, or cronjobs. It's resource-intensive but ensures comprehensive prediction. It is ideal for scenarios where every entity's prediction adds value. - 𝐈𝐧𝐟𝐞𝐫𝐞𝐧𝐜𝐞 𝐛𝐲 𝐑𝐞𝐪𝐮𝐞𝐬𝐭: Triggers model inference only when needed, suitable for user interactions that can accommodate on-the-fly predictions. It's efficient for fast models or when user experience allows upfront triggering. - 𝐒𝐭𝐫𝐞𝐚𝐦𝐢𝐧𝐠 𝐏𝐫𝐨𝐜𝐞𝐬𝐬𝐢𝐧𝐠: For high-throughput applications, this event-driven approach uses technologies like Kafka to trigger inferences, perfect for real-time data processing. Each strategy has its place. Batch processing suits comprehensive analysis, request-based inference for user-driven scenarios, and streaming for real-time applications. Understanding these can optimize your AI integration, making your systems more efficient. #AI #MachineLearning #InferenceStrategies #Efficiency #DataProcessing #RealTimeApplications

🚀 Dive into the future of document querying with LLM models using Kernel Memory! 🧠 This new tech allows seamless integration of various document formats into your AI projects, enhancing information retrieval and data interaction like never before. Perfect for devs looking to leverage AI in a more meaningful way! Check out how to boost your projects with this innovative framework. #AI #TechInnovation #KernelMemory 🌐✨

In an era where advancements in artificial intelligence are reshaping the boundaries of technology and we see something new every day, the release of Gemini 1.5 Pro marks another significant milestone. This new model is another leap forward, setting a new standard for multimodal understanding across an unprecedented scale. Gemini 1.5 Pro extends language model context lengths by over an order of magnitude, enabling processing of upto a million of tokens of context. Its ability to engage in in-context learning from entire long documents and achieving near-perfect recall on synthetic retrieval tasks, feats that will transform how the model interacts with complex datasets and long documents. I do like the competition among the various models :) Quoting from Google's Blog. https://lnkd.in/eesDKhC4 "Through a series of machine learning innovations, we’ve increased 1.5 Pro’s context window capacity far beyond the original 32,000 tokens for Gemini 1.0. We can now run up to 1 million tokens in production." "Gemini 1.5 Pro can reason across 100,000 lines of code giving helpful solutions, modifications and explanations." "Gemini 1.5 Pro maintains high levels of performance even as its context window increases. In the Needle In A Haystack (NIAH) evaluation, where a small piece of text containing a particular fact or statement is purposely placed within a long block of text, 1.5 Pro found the embedded text 99% of the time, in blocks of data as long as 1 million tokens."

𝐓𝐡𝐞 𝐍𝐞𝐱𝐭 𝐅𝐫𝐨𝐧𝐭𝐢𝐞𝐫 𝐢𝐧 #𝐋𝐋𝐌𝐬 - 𝐈𝐧𝐭𝐫𝐨𝐝𝐮𝐜𝐢𝐧𝐠 𝐨𝐩𝐭𝐢𝐦𝐚𝐥-𝐩𝐚𝐫𝐚𝐦𝐞𝐭𝐞𝐫-𝐜𝐨𝐦𝐩𝐫𝐞𝐬𝐬𝐢𝐨𝐧 𝐟𝐨𝐫 𝐋𝐥𝐚𝐦𝐚 𝟐 𝐌𝐨𝐝𝐞𝐥𝐬 By Rodrigo Hernandez, Multiverse Computing We move towards Small Language Models (SLMs). This revolutionary approach is not just a concept but a reality, thanks to the brilliant minds at Multiverse Computing via #CompactifAI. By integrating cutting-edge 𝐓𝐞𝐧𝐬𝐨𝐫 𝐍𝐞𝐭𝐰𝐨𝐫𝐤𝐬 𝐦𝐞𝐭𝐡𝐨𝐝𝐨𝐥𝐨𝐠𝐢𝐞𝐬, they have achieved what many thought was near impossible: compressing Llama 2 models without noticeably compromising on performance. 𝐃𝐨𝐰𝐧𝐬𝐜𝐚𝐥𝐞 𝐋𝐥𝐚𝐦𝐚 𝟐 𝐭𝐨 𝟓𝐁/𝟐𝐁 𝐩𝐚𝐫𝐚𝐦𝐞𝐭𝐞𝐫𝐬: these optimized models redefine efficiency. Metrics like MMLU, trivia QA, and more, remain unaffected, proving that less can indeed be more. Why does this matter? ✅ 𝐑𝐞𝐝𝐮𝐜𝐞𝐝 𝐆𝐏𝐔 𝐦𝐞𝐦𝐨𝐫𝐲 c𝐨𝐧𝐬𝐮𝐦𝐩𝐭𝐢𝐨𝐧: Leaner models mean broader accessibility. ✅ 𝐈𝐧𝐟𝐞𝐫𝐞𝐧𝐜𝐞 𝐞𝐟𝐟𝐢𝐜𝐢𝐞𝐧𝐜𝐲 𝐨𝐧 𝐬𝐦𝐚𝐥𝐥𝐞𝐫 𝐆𝐏𝐔𝐬 / 𝐂𝐏𝐔𝐬: Making advanced computing power not a prerequisite, enabling edge / on-device AI. ✅ 𝐁𝐚𝐥𝐚𝐧𝐜𝐢𝐧𝐠 𝐜𝐨𝐧𝐬𝐮𝐦𝐩𝐭𝐢𝐨𝐧 𝐜𝐨𝐬𝐭𝐬 𝐚𝐧𝐝 𝐬𝐩𝐞𝐞𝐝: Unlike strictly quantized models that do not affect inference performance, these optimized models enhance operational efficiency. In a world where #sustainability, speed and efficiency are paramount, this breakthrough stands as a testament to innovation. It's not just about making LLMs faster or smaller; it's about reimagining their potential, making them accessible to a wider range of devices, and, ultimately, democratizing AI. As we stand on the brink of this new era, the implications for both development and application are immense. This isn't just an upgrade. It's the ongoing revolution. Your thoughts? How do you see parameter-optimized LLMs/SLMs models impacting your field? #MultiverseComputing #LLMs #AI #SLMs #TensorNetworks #ParameterEngineering #MachineLearning #ParameterCompression