Gareth Isaac’s Post

At KYD we have been working with clients and partners to gain the realistic benefits of this AI #llm knowledge explosion. Naturally certain companies believe AI does everything 🤠 – being practitioners in AI we understand what it can be used for, and what will be a number of years away (if at all)! This is where we help our clients and partners. At KYD we have been working with the foundations underpinning RAG🚦 (Retrieval-Augmented Generation) for generative models and reorientating this to apply to several Solidatus use cases. “In fact, almost any business can turn its technical or policy manuals, videos or logs into resources called knowledge bases that can enhance LLMs. These sources can enable use cases such as customer or field support, employee training and developer productivity.” (NVIDIA Blog explaining RAG) One of the unsung foundational hero's of 🚦 RAG are “Embeddings”. If you have used ChatGPT then you have probably unknowingly used Embeddings! Simply put Embeddings are vectors (think of a 3D shape - but in reality they can have 1000+ dimensions rather than 3 🤯) and based on the 'shape' they can be compared to other shapes to find the closest match. That is a super simplified description of a Embedding. So given this how are we leveraging embeddings in Solidatus? We at KYD Analytics have developed matching code based on the foundations of Embeddings that performs linkage between Objects and Models in your Solidatus ‘universe’. The KYD integrations use cutting edge data driven approaches to unlock common meanings in your Solidatus models and then provides mapping candidates. Internally we have used this in the “KYD Policy Solution” built on top of Solidatus. Here we find the common aspects and meaning of multiple policies and establish where there is commonality and provide links. We are using other AI techniques in the product as well. To know more about embedding AI technology in Solidatus contact us at ✉ info@kyd.ai or 💻kyd.ai/solidatus.

More than a year after the launch of #ChatGPT, companies are still facing the same question when they first considered the technology: How do they actually go about putting it into business use? Many companies have simply discovered that generative AI tools like LLMs, while impressive, aren’t plug and play. Companies should consider a few suggestions when thinking of whether and how to onboard these tools: 1) choose performance over novelty 2) combine GenAI with tools like vector databases 3) never forget the human-in-the-loop 4) trace your data, and 5) have realistic expectations Wanna know more? Our latest Harvard Business Review article is out and together with my co-authors, Terence Tse, PhD, Danny Goh and Paul Lee, we invite you to read, what we think will be a compelling work. Link here: https://lnkd.in/dBT9dEv4

Everyone talks about AI. In fact, we also use #AI extensively in #VTDocs. But before delving into the nuances of AI and related technologies, it's important to understand the business context and the challenges it addresses. Let’s consider an example: When reviewing multiple documents, business analysts, project managers, and other professionals typically examine them for essential requirements: They identify dependencies within a document—perhaps paragraph three relates to paragraph 52 on page 19.  These interdependencies occur both within individual documents and across multiple ones.  Examples include standard operating procedures with specific policies, terms and conditions, or supplier agreements.  How does their process look like today? The analyst opens a document and skims through it from start to finish, noting any critical points or elements of interest.  They use the search function (Ctrl + F) to scan for terms like "warranty" or "indemnifications" in the contract. This method is time-consuming and ad-hoc. While experienced professionals might navigate quickly, those less experienced or new to the business need much more time and might overlook crucial details. VT Docs and its AI models offer a different approach: When uploading multiple documents into VT Docs, users instantly see common themes across documents – powered by AI.  Imagine a matrix where each column represents a document, and on the left, thematic frequencies are highlighted.  This visualization allows users to spot patterns quickly. For instance, if references to "indemnification" or variations of that appear across several documents, then VT Docs shows them automatically and provides an aggregated view. VT Docs utilizes an AI model specially designed for business and technical communications.  So, there's no "hallucination" or inaccuracy— the AI doesn't use a generative model (like ChatGPT) but relies on a model that can trace back to the precise text, ensuring 100% accuracy. This precision is non-negotiable and pivotal. Another consideration, especially concerning generative AI, is security and data integrity.  VT Docs and its AI model run completely isolated in the most secure locked-down environments. This ensures no data leaks or unauthorized learning from your data. AI has the potential to transform industries, but only if it maintains reliability and security and seamlessly integrates into existing workflows. At @VisibleThread, we aim to pioneer solutions that embody these principles, ensuring not just innovation but secure and dependable results. #documentreview #artificialintelligence #enterprise

Everyone talks about AI. In fact, we also use #AI extensively in #VTDocs. But before delving into the nuances of AI and related technologies, it's important to understand the business context and the challenges it addresses. Let’s consider an example: When reviewing multiple documents, business analysts, project managers, and other professionals typically examine them for essential requirements: They identify dependencies within a document—perhaps paragraph three relates to paragraph 52 on page 19.  These interdependencies occur both within individual documents and across multiple ones.  Examples include standard operating procedures with specific policies, terms and conditions, or supplier agreements.  How does their process look like today? The analyst opens a document and skims through it from start to finish, noting any critical points or elements of interest.  They use the search function (Ctrl + F) to scan for terms like "warranty" or "indemnifications" in the contract. This method is time-consuming and ad-hoc. While experienced professionals might navigate quickly, those less experienced or new to the business need much more time and might overlook crucial details. VT Docs and its AI models offer a different approach: When uploading multiple documents into VT Docs, users instantly see common themes across documents – powered by AI.  Imagine a matrix where each column represents a document, and on the left, thematic frequencies are highlighted.  This visualization allows users to spot patterns quickly. For instance, if references to "indemnification" or variations of that appear across several documents, then VT Docs shows them automatically and provides an aggregated view. VT Docs utilizes an AI model specially designed for business and technical communications.  So, there's no "hallucination" or inaccuracy— the AI doesn't use a generative model (like ChatGPT) but relies on a model that can trace back to the precise text, ensuring 100% accuracy. This precision is non-negotiable and pivotal. Another consideration, especially concerning generative AI, is security and data integrity.  VT Docs and its AI model run completely isolated in the most secure locked-down environments. This ensures no data leaks or unauthorized learning from your data. AI has the potential to transform industries, but only if it maintains reliability and security and seamlessly integrates into existing workflows. At @VisibleThread, we aim to pioneer solutions that embody these principles, ensuring not just innovation but secure and dependable results. #documentreview #artificialintelligence #enterprise

Tl;dr: next generation of AI agents will have dynamic tool generation capabilities. Current agentic AI systems (e.g. ChatGPT, Claude) are awesome, and we've just started to scratch the surface of their real-world usefulness.  However, they are still inherently constrained by their pre-defined set of "tools", i.e., the functions that the agent can execute. In other words, the ceiling of an agentic system is always determined by the existence and quality of the tools it has access to. In this sense, an agent is an entity that is designed to operate within a specific context/space which is narrowed down by its toolset. This leads me to think that the next generation of AI agents will likely be able to generate their own tools on-demand, per situation. Such a system would unlock endless possibilities by increasing its autonomy dramatically and eliminating the need for defining tools in advance. Now, from an engineering perspective, how can we build such systems?  To answer this, we need to figure out two things:  a. what are the components of a tool?  b. can it be automatically generated by AI? Generally speaking, a tool is enabled by the following: 1. Code: Representing the tool's logic and functionality, wrapped as a function. 2. Schema: Defining the function's parameters. 3. Function/Tool Calling mechanism: Enabling tool execution by extracting the function's arguments from the agent's context. Can we automatically generate these? The answer is simple, yes: - LLMs are not only capable of but excel at generating code from natural language descriptions. - The schemas can be derived directly from the function's code (or be generated using LLMs in trickier cases). - Function calling mechanisms are now supported in most LLMs. In fact, there is nothing essential that prevents us from building such systems today (which is a bit scary). The technology is already here, waiting to be unleashed. Interesting times ahead... #NextGenAIAgents

LLaMa 3.1 just killed proprietary AI models like ChatGPT/Claude, et al. ⤵️ 𝗛𝗲𝗿𝗲'𝘀 𝟰 𝗿𝗲𝗮𝘀𝗼𝗻𝘀 𝘄𝗵𝘆: 1. Customization is no longer a luxury, 𝗶𝘁'𝘀 𝗮 𝗻𝗲𝗰𝗲𝘀𝘀𝗶𝘁𝘆. If you're not tailoring AI to your specific needs, you're already behind. 2. "Open-source can't compete" is now a myth. 𝗟𝗟𝗮𝗠𝗮 𝟯.𝟭 doesn't just match closed models, it 𝗼𝘂𝘁𝗽𝗲𝗿𝗳𝗼𝗿𝗺𝘀 them. Ignore this at your peril. 3. The AI landscape is about to explode. Expect a flood of specialized models that will make generic solutions obsolete. 4. If you're still relying on off-the-shelf AI solutions, you're planning to fail. 𝗟𝗟𝗮𝗠𝗮 𝟯.𝟭 𝗶𝘀 𝘁𝗵𝗲 𝗻𝗲𝘄 𝘀𝘁𝗮𝗻𝗱𝗮𝗿𝗱. Adapt or become irrelevant. Second place is the first loser in the AI race. See the ARR difference between OpenAI and the next 3 "top models". ----- Hi, I'm Junaid, CEO at Ertiqah 👋 🔗 If you’re a small business owner, I share practical ways in which 𝘄𝗲 𝘂𝘀𝗲 𝗔𝗜 to be more productive and reduce costs 𝗶𝗻 𝗼𝘂𝗿 𝗼𝘄𝗻 𝗯𝘂𝘀𝗶𝗻𝗲𝘀𝘀, and how 𝘆𝗼𝘂 𝗰𝗮𝗻 𝗱𝗼 𝘁𝗵𝗲 𝘀𝗮𝗺𝗲. 𝗟𝗶𝗻𝗸 𝘁𝗼 𝗴𝗲𝘁 𝘁𝗵𝗲𝗺 𝗶𝗻 𝘆𝗼𝘂𝗿 𝗶𝗻𝗯𝗼𝘅 𝗲𝘃𝗲𝗿𝘆 𝘄𝗲𝗲𝗸: https://lnkd.in/ewkRZSin

More than a year after the launch of #ChatGPT, companies are still facing the same question when they first considered the technology: How do they actually go about putting it into business use? Many companies have simply discovered that #generativeAI AI tools like #LLMs, while impressive, aren’t plug and play. Companies should consider a few suggestions when thinking of whether and how to onboard these tools: 1) choose performance over novelty, 2) combine #GenAI with tools like vector databases, 3) never forget the human-in-the-loop, 4) trace your data, and 5) have realistic expectations.

💥 “Ohh, but it looks like you used ChatGPT (a Large Language Model) to write this...” 💥 Ever heard that one? It’s all too common – and here’s why it happens: If you ask an LLM, “Write a 500-word blog on [topic],” you’re going to get an average response. One prompt, one answer – no iterations, no refinement, no analysis. It's like expecting a gourmet meal from a microwave. 😛 But what if you could level up with Agentic Workflows? 🧠 Here’s where advanced prompt engineering transforms the game. An Agentic Workflow isn’t just about throwing a prompt and hoping for the best. It’s a well structured, multi-agent approach that breaks tasks into smaller, manageable steps, incorporating techniques like chain of thought, task planning, and self-reflection to get outputs that actually hit the mark. Each AI agent works on a specific part of the task, ensuring every step is reviewed, improved, and precisely executed. 🌟 Take blog writing as an example: 🔹 Traditional Zero-Shot Prompt: “Write a blog on [topic].” The AI gives one-shot output – quick but lacking depth, no quality checks, no room for iterative improvement. 🔹 Agentic Workflow: Research credible sources Create a detailed outline with headings/sub-headings Review and modify the structure Draft, edit, and refine – step by step Each step’s output serves as the next input, transforming a single shot into a thoughtful, polished piece. That’s AI-driven quality control in action. Why does this matter? ➡️ Quality gains: Less guesswork, more refinement ➡️ Increased relevance: No one-off solutions; each agent fine-tunes the output to fit the needs ➡️ Strategic insight: Enables a comprehensive, multi-layered response to complex challenges The future of AI lies not in single-shot prompts but in collaborative, agentic workflows that add real value to every output. Imagine scaling this in business processes across industries! 😍 So, the next time you hear, "Ohh, but you just used AI for that,” you can smile and know your output went through a rigorous, intelligent process – not a one-prompt wonder. #AI #AgenticWorkflows #PromptEngineering #FutureOfAI #LLMs #Productivity

Everyone talks about AI. In fact, we also use #AI extensively in #VTDocs. But before delving into the nuances of AI and related technologies, it's important to understand the business context and the challenges it addresses. Let’s consider an example: When reviewing multiple documents, business analysts, project managers, and other professionals typically examine them for essential requirements: They identify dependencies within a document—perhaps paragraph three relates to paragraph 52 on page 19.  These interdependencies occur both within individual documents and across multiple ones.  Examples include standard operating procedures with specific policies, terms and conditions, or supplier agreements.  How does their process look like today? The analyst opens a document and skims through it from start to finish, noting any critical points or elements of interest.  They use the search function (Ctrl + F) to scan for terms like "warranty" or "indemnifications" in the contract. This method is time-consuming and ad-hoc. While experienced professionals might navigate quickly, those less experienced or new to the business need much more time and might overlook crucial details. VT Docs and its AI models offer a different approach: When uploading multiple documents into VT Docs, users instantly see common themes across documents – powered by AI.  Imagine a matrix where each column represents a document, and on the left, thematic frequencies are highlighted.  This visualization allows users to spot patterns quickly. For instance, if references to "indemnification" or variations of that appear across several documents, then VT Docs shows them automatically and provides an aggregated view. VT Docs utilizes an AI model specially designed for business and technical communications.  So, there's no "hallucination" or inaccuracy— the AI doesn't use a generative model (like ChatGPT) but relies on a model that can trace back to the precise text, ensuring 100% accuracy. This precision is non-negotiable and pivotal. Another consideration, especially concerning generative AI, is security and data integrity.  VT Docs and its AI model run completely isolated in the most secure locked-down environments. This ensures no data leaks or unauthorized learning from your data. AI has the potential to transform industries, but only if it maintains reliability and security and seamlessly integrates into existing workflows. At @VisibleThread, we aim to pioneer solutions that embody these principles, ensuring not just innovation but secure and dependable results. #documentreview #artificialintelligence #enterprise

Automation + AI Attended the IMDA X SPARKS Executive Toundtable on Operationalising A.I. For Enterprise Scale. A few points resonated with me. ✅How can we operationalise AI so that everyday office workers can access AI? ✅The future of smart AI requires context: build up from User Input, Canonical Knowledge to include Personalisation and Profile Data and Derived Insights. ✅in 2023, the Marketing department benfitted most from analytical AI in terms of increase in revenue! Check my ChatGPT generated picture of AI! https://ow.ly/jH8N50PT5HJ