Stefan Suwelack’s Post

tl;dr: "there is a lot of hype and unrealistic expectations". Let's not mindlessly share the vacuous promises of those who desperately need to raise cash. There is no need to throw the baby out with the bathwater. By focusing on the areas where LLMs can offer a meaningful contribution, we will draw investments to productive improvements - long term.

"We want an AI agent that can automate XYZ. But we don't want it if we can't ensure quality. Can we do this with current tech?" - I am getting such requests and questions more frequently since the last couple months. And while it was easier for me to guess if a task would be easy to automate using current frameworks and models, it might seem like rocket science to business leaders who only have hype-driven articles and mixed opinions to rely on. Some other questions that I get in actual implementations are somewhere along these lines: * AI Agents v/s Generative AI: How is Agentic AI Different from Using LLMs for Specific Tasks? * Is Agentic AI Enough for Complex Software? * What's better - using Agentic Frameworks vs. Writing Your Own Code/Prompts So I decided to answer these questions and explain in detail the current capabilities and challenges of such systems in my recent blog post - https://lnkd.in/gs4aCm3W Do comment below if there are more lingering questions in your mind regarding the current state and near future of AI Agents!

#AgenticAI for a real-life #Marketing scenario AI Agents are the current #hype. I have written about them [1], and others are also discussing them. Overall, however, it does mean that there is a lot of confusion reg. what are #AIAgents? How are they different from #GenAI, or #LLMAgents? This has also led to a lot of players touting anything and everything as AI Agents. I am sorry but prompting an #LLM is not an AI Agent. It is also not sufficient to call a script making multiple LLM function calls as an AI Agent. Suddenly all #RPA scripts (which no doubt have their own value) have become AI Agents. Building an AI agent is not easy, so #AppStudios allowing you to string together a quick app should not also be classified as an AI Agent - it is #API orchestration at best. What are AI Agents then? We can endlessly debate this. So I thought the best would be to be to discuss the #functional / non-functional requirements of Agentic AI using an actual #enterprise use-case: - Given a user task, the goal of an agent platform is to identify an agent (group of agents) capable to executing that task. So the first component we need is an orchestration layer capable of #decomposing a task into sub-tasks, with execution of the resp. agents orchestrated by an #orchestration engine [2]. As of today, we prompt an LLM for the task decomposition [3]. So this is the overlap with Gen AI. Unfortunately, this also means that Agentic AI today is limited by the #reasoning capabilities of LLMs. For ex., the #GPT4 task decomposition of the above prompt is detailed in the fig: (Analyze products) - (Identify target audience) - (Create tailored email campaign) - Most complex tasks are long-running. Once the initial email campaign is launched, the agent monitors the campaign for 1-week. #Memory management [4] is key for Agentic AI systems for context sharing between tasks and maintaining execution context over long periods. The solution is to use #VectorDBs to store the agent memory externally, making it accessible as needed. - Next, we come to the most important capability of an Agentic AI framework, its #autonomy. It needs to monitor the execution / environment and #adapt autonomously. In this case, the agent realised that it is not going to achieve its #sales goal and autonomously added the tasks: (Find alternative products) - (Utilize customer data) - (Perform A/B testing) - Finally, integration with enterprise systems (#CRM in this case) will be needed for most use-cases. So plan for that. This is a live use-case running on the #AgentGPT platform, with some tweaking for memory and enterprise integration. The results are impressive - this is the task of a full-time #Marketer! However, it was achieved over a number of iterations. Manual supervision is still needed - #secure your CRM system before giving its access to the agent -:) We are currently working towards adding these safety & #ResponsibleAI guardrails to an #AgentOps pipeline [5].

I put together an Agentic AI shell system for my personal usage. Now what? This system allows legacy admins and lower-skill new admin use plain language to execute system commands. Would this be useful for your admins? This process could all be called via API to actually fix problems using existing tooling and allow for AI Agentic processes to be folded into existing workflows in IT. #ai #agentic #administration

Exciting advancements in AI research are paving the way for large multimodal models (LMMs), combining text, images, and more for enhanced capabilities. From assisting the visually impaired to revolutionizing industries like healthcare and e-commerce, the potential of multimodal AI is vast. With the ability to process diverse data types, LMMs offer a more human-like approach to problem-solving and decision-making. As this technology continues to evolve, the possibilities for innovation and transformation are endless. #MultimodalAI #LMMs #AI #Innovation #Technology #ArtificialIntelligence https://lnkd.in/gxUbq2tE

Have you heard of RAG (Retrieval-Augmented Generation) in AI? It's a game-changer for making LLM chatbots provide more precise information based on your data. But what if you're not a developer? Here's the thing: you don't always need complex integrations to leverage RAG-like functionalities! 🧠 Simple Workarounds for Non-Tech Users: 𝗗𝗼𝗰𝘂𝗺𝗲𝗻𝘁 𝗔𝘁𝘁𝗮𝗰𝗵𝗺𝗲𝗻𝘁: Many modern AI chatbots allow you to attach documents directly to your conversation. This creates a "mini-RAG" experience! 𝗖𝗼𝗽𝘆-𝗣𝗮𝘀𝘁𝗲 𝗔𝗽𝗽𝗿𝗼𝗮𝗰𝗵: Need the AI to consider specific information? Simply paste relevant excerpts into your prompt. 𝗦𝘂𝗺𝗺𝗮𝗿𝘆 𝗙𝗶𝗿𝘀𝘁: Before asking questions, provide a brief summary of the context you want the AI to consider. Real Example: The other day I needed to refresh my memory on what I need to do for my work travels. I uploaded the documentation created by our People Manager Silvia Repetto. 🎯Optimization at its simplest: the AI uses this context to provide more accurate, company-specific responses and both Silvia and I saved time. No coding required! 💡 Pro Tip: Be strategic about what information you provide. ⚠IMPORTANT: make sure you are 𝙚𝙭𝙩𝙧𝙚𝙢𝙚𝙡𝙮 aware of privacy policies before uploading any document directly on a chatbot. Make sure to protect all your and your company data from misuse and disclosure! 🤔 Why This Matters: Makes AI responses more accurate and context-aware Empowers non-technical users to leverage advanced AI features Improves efficiency without requiring technical expertise Have you found creative ways to make AI tools work better for your needs? Share your tricks in the comments! #ArtificialIntelligence #RAG #BusinessOptimization #ProductivityHacks #AIforAll

Agentic #RAG for Developers! Agentic RAG emerges as an evolution of traditional RAG, integrating AI agents to enhance the RAG approach. This approach employs autonomous agents to analyze initial findings and strategically select effective tools for data retrieval. These AI agents have the capability to breakdown the complex task into several subtasks so it becomes easy to handle. They also possess the memory (like chat history) so they know what has happened and what steps needs to be taken further. Also, these AI agents are so smart they can call any API or tool whenever there is a requirement to solve the tasks. The agents can come up with logic, reasoning and take actions accordingly. This is what makes an agentic RAG approach so prominent. The system deconstructs complex queries into manageable segments, assigning specific agents to each part while maintaining seamless coordination. Know more: https://lnkd.in/g8iAjH6r Building agentic RAG pipelines with Vectorize: https://vectorize.io/

Why isn’t anyone talking about how DeepSeek is transforming the landscape of Agentic AI testing and evaluation? With its innovative open-source approach and advanced models like DeepSeek-V3 and DeepSeek-R1, the platform is setting new standards for transparency, efficiency, and accessibility. DeepSeek’s impact is driven by: 1. Cost-Effective Solutions: Delivering competitive performance on a fraction of the budget. 2. Open-Source Transparency: Fully inspectable code ensures compliance with GDPR/CCPA and fosters trust in AI systems. 3. Innovative Architecture: The Mixture-of-Experts (MoE) design activates only relevant parameters during queries, optimizing speed and computational efficiency. 4. Exceptional Performance: DeepSeek-R1 surpasses leading models like OpenAI’s o1 in benchmarks for coding and mathematical tasks. 5. Broad Accessibility: Offering unlimited free usage, it empowers individuals and organizations to leverage advanced AI capabilities without financial barriers. At RagaAI Inc, we’re thrilled by DeepSeek’s potential to redefine Agentic AI testing, aligning with our own vision of building transparent, reliable, and impactful AI technology. #opensource #ai

𝗔𝗶𝗱𝗮 & 𝗔𝗜 𝗔𝘀𝘀𝗶𝘀𝘁: 𝗛𝗼𝘄 𝗧𝗵𝗲𝘆 𝗖𝗮𝗻 𝗛𝗲𝗹𝗽 𝗬𝗼𝘂 𝘄𝗶𝘁𝗵 𝗬𝗼𝘂𝗿 𝗜𝗻𝘁𝗲𝗴𝗿𝗮𝘁𝗶𝗼𝗻 𝗡𝗲𝗲𝗱𝘀 🌟 Aida and AI Assist are Exalate's way of harnessing #AI to make your integration process smoother and faster. Here’s how they each play their part: 𝙒𝙝𝙖𝙩’𝙨 𝙩𝙝𝙚 𝘿𝙞𝙛𝙛𝙚𝙧𝙚𝙣𝙘𝙚? 💫 𝗔𝗶𝗱𝗮: Think of her as your 24/7 Exalate expert. Aida quickly answers questions about features, setup, and usage. Perfect for new users, or anyone who needs fast, clear directions. Skip the support line and just ask Aida! 💫 𝗔𝗜 𝗔𝘀𝘀𝗶𝘀𝘁: For those using Exalate’s Script Mode, AI Assist is your co-pilot for creating sync scripts. Instead of spending hours on complex scripting, AI Assist turns it into simple prompts, helping you achieve powerful customizations with minimal effort. Prompt. Review. Sync. It’s that easy. 𝙒𝙝𝙮 𝘽𝙤𝙩𝙝?  Use Aida to understand your options and requirements, then let AI Assist fine-tune and customize your integration with ease. Together, they cover all the bases. 👌 No more documentation diving or script debugging until midnight! 👇 Read the full article for all the details on how to use Aida and AI Assist to make your integrations smoother than ever! #AI #Integration #Exalate #AIAssist #AIinTech #ProductivityHacks

Earlier this week the team wrote a detailed specification for how we'll support a Brightflag customer's business process that spans multiple systems. We wanted to include a flow chart, and we asked to AI to create it. We gave the same prompt to a general-purpose AI product and to Lucid Software, a product purpose-built for visualizing data. The difference in the outputs is stark. General-purpose AI products are *incredible*—I use them every day—but they aren't a realistic alternative to purpose-built ones. (Note I redacted confidential information.) We're operating in a world in which AI is and will continue to be everywhere. The distinction between "AI product" and "not an AI product" is disappearing fast. But—and here's the kicker—the gap between "valuable AI product" and "not valuable AI product" is growing. This is an important part of the thesis behind Ask Brightflag, our conversational user interface launching this month. Can you dump all your legal matter and spend data into a general-purpose AI product and ask it questions? Sure, but the answers will be gibberish (we've tried!). #legalops #legaltech #ai #genai