Edward Moemeka

An important article I wrote that differentiates LLM assistance and agentic AI. Please read. 🔗 https://lnkd.in/eb8mJyjJ A note for those applying to jobs: This matters for your career too. When you're in an interview and someone asks what you built, be precise. If you used an LLM to help you write code faster, say that, don't call it an agent. If you built a RAG pipeline, call it retrieval-augmented generation, don't call it agentic. If you built a chatbot with a system prompt, own it for what it is. Hiring managers who know the difference will respect your precision. Hiring managers who don't know the difference need you to teach them. Either way, you win by naming it correctly. Overselling what you built doesn't just hurt your credibility when someone digs deeper, it tells the room you don't actually understand the architecture you're describing. Know what you built. Know what it's called. Know why it matters at that level. That's what gets you hired. #AI #AgenticAI #LLM #NLP #SportsAnalytics #DataScience #CareerAdvice #MachineLearning

13

New release from Packt Publishing… “Building Business-Ready Generative #AI Systems — Build Human-Centered Generative AI Systems with Context-Aware Agents, Memory, and LLMs for the Enterprise” See it at https://amzn.to/3Jdcio5 𝓚𝓮𝔂 𝓕𝓮𝓪𝓽𝓾𝓻𝓮𝓼: 🔵Build an adaptive, context-aware AI controller with advanced memory strategies 🟢Enhance GenAISys with multi-domain, multimodal reasoning capabilities and Chain of Thought (CoT) 🟠Seamlessly integrate cutting-edge OpenAI and DeepSeek models as you see fit

14

3 Comments

🚀 How to Evaluate LLM Applications Large Language Models (LLMs) are powerful, but their real impact depends on how we evaluate them. The right benchmarks and evaluation methods ensure accuracy, fairness, and real-world applicability. 🔍 Key Takeaways from the Framework: ✅ Benchmarks Language Understanding & QA: TruthfulQA, MMLU, DROP Common-sense & Reasoning: ARC, HellaSwag, BIG-Bench Hard Coding: HumanEval, CodeXGLUE Conversation & Chatbots: Chatbot Arena, MT Bench, Stanford HELM ✅ Evaluation Methods Character Level: Levenshtein Distance Word Based: BLEU, ROUGE, METEOR, Perplexity Embedding Based: BERTScore, MoverScore, QAG Score Language Model Based: GPTScore, SelfCheckGPT, G-Eval, Prometheus NLP Models: NLI, BLEURT Frameworks: DeepEval, RAGAS 💡 Whether you’re building chatbots, coding assistants, or reasoning systems, choosing the right benchmark + evaluation method is crucial to measure performance and reliability. 👉 Question for you: Which evaluation method do you think gives the most real-world reflection of LLM performance traditional metrics like BLEU/ROUGE or LLM-based evaluations like GPTScore & SelfCheckGPT? #LLM #AI #MachineLearning #NLP #Evaluation #GenAI #ArtificialIntelligence

142

17 Comments

Just watched Google release Gemini DeepThink - their first multi-agent AI system - and it confirms what I've been saying: the future of AI is multi-agent architectures. But here's the problem most people aren't talking about: current memory systems can't handle thousands of AI agents running in parallel. When I scaled past 300 AI models, traditional knowledge graphs became the bottleneck. That's why I built Cortesia - a brain-inspired knowledge graph that: • Processes queries 10x faster than Neo4j (8 minutes vs 45 minutes for 10M nodes) • Uses Hebbian learning like actual neurons • Implements distributed consensus for multi-organization collaboration • Built in Rust for maximum performance The system mimics how human memory actually works - stronger connections for frequently accessed information, natural decay for unused data, and competitive inhibition for focus. It's designed specifically for the agentic AI future we're heading into. What memory challenges are you facing as AI systems become more complex? Are you seeing similar bottlenecks in your multi-agent implementations? 📚 Open source repo: https://lnkd.in/g3hMt6jk Join the Pheromind Community: pheromind.ai 💬 Join the Discord Community https://lnkd.in/de6bxfyu #MultiAgentAI #AIMemory #DistributedSystems #BrainInspiredAI #AIInfrastructure

35

6 Comments

Epic Games' latest AI character—now facing a SAG-AFTRA complaint—shows why AI requires both creative chaos AND thoughtful boundaries. The brilliance of AI development lies in pushing limits, but sustainable innovation demands guardrails around training data acquisition and quality. Epic's situation demonstrates that everyone needs ethical frameworks. AI isn't just about what's possible, it's about what's responsible. The right guardrails don't stifle innovation; they channel it productively and protect it from potential backlash. https://lnkd.in/g8sVvACW #AIInnovation #ResponsibleDisruption #EthicalGuardrails

16

My latest CMSWire article is an update on the latest #prompt engineering techniques for #AI, featuring a few examples on Gemini and ChatGPT. "Prompts, the key instruction for large language models, are how people talk to their AI of choice. But how do people plan what to say to their AI? The starting point is #promptengineering." Read further to learn what you can do to plan better #AI usage #Tech #data #marketing https://lnkd.in/gtVnZpJr

12

3 Comments

Weekly reminder that a "hallucinating" AI is often still more consistent than two or more humans. Interesting work on document extraction (titles, dates, methods, outcomes, other key content) in Xi Long et al., Hallucination vs interpretation: rethinking accuracy and precision in AI-assisted data extraction for knowledge synthesis, 2025. Also GPT-5 is not a guaranteed or a remarkable improvement over models like 4o or even a small Qwen. It all depends on the task. Don't fall for the hype machine and snake oil.

26

1 Comment

95% of GenAI pilots fail. MIT’s (State of AI in Business 2025) draws a sharp line: adoption is high, transformation is rare with only ~5% delivering measurable business impact. The study notes that most Gen AI initiatives lack context, integration, and governance. While budgets chase front-office use cases, value often hides in operations and finance. External partnerships are nearly twice as likely to scale compared to in-house efforts. The real issue isn’t the technology—it’s management, redesign, and guardrails that decide who succeeds and who stalls. The prize remains enormous. For example, Morgan Stanley estimates up to $920B in annual savings for U.S. companies if they fully adopt AI. At NeoStats, we’ve seen what the 5% do differently. Courts cut 6–8‑week case‑document reviews to minutes; insurers auto‑adjudicate low‑value claims; AI ops agents update addresses, send documents, and craft personalized emails; chatbots query live business data; call‑center AI maps sentiment, intent, and compliance. We started tiny, proved ROI, earned top‑down sponsorship, managed change, and enforced governance throughout. What works: ·      Start with value, not models. ·      Govern from day one (data quality, risk, human‑in‑the‑loop, audit trails). ·      Build small, scale in sequence; retire what doesn’t pay back. ·      Treat AI as operating‑model change, not an app. It’s a wake‑up call. The blip is mostly by weak governance and “demo‑first” thinking. Choose discipline and join the 5%. Felci Fernando Manu Nair Pradeep Gudihal Devadharshani A Ria Liz Luke https://lnkd.in/dHWXeXFx

67

4 Comments

Apple just shared a significant finding about LLMs: their reasoning and accuracy can fundamentally degrade when tasks become more complex. Their recent study found that even advanced models like Claude 3.7 and Deepseek-R1 show a "total collapse" in accuracy and struggle to maintain insights as problem complexity rises. This is a critical challenge for enterprises relying on AI for precise answers, deep insights, and reliable decision-making. If LLMs falter on complex logical tasks, how can we ensure they deliver accurate, actionable intelligence for your business? ➡️ The answer lies in a robust foundation: Knowledge Graphs. Our latest article, "Improving Contextual Understanding and Reasoning: 10 Ways Knowledge Graphs Can Help Optimize AI" by Andreas Blumauer, delves into exactly how. While Apple's research highlights inherent limitations in LLM scaling for reasoning, it powerfully reaffirms the strategic imperative of structured, external knowledge. Knowledge Graphs empower LLMs to deliver: ✅ Superior Accuracy: Moving beyond probabilistic responses to provide grounded, factual answers ✅ Deeper Insights: Connecting disparate data points to uncover complex relationships and hidden patterns ✅ Domain Relevance: Infusing specialized knowledge for highly tailored and industry-specific AI ✅ Explainability & Trust: Offering transparent reasoning paths, vital for user adoption and compliance As Andreas explains, "Knowledge graphs not only optimize today’s models—they build the infrastructure needed to support more meaningful, accountable, and adaptive AI in the long term." Learn how Knowledge Graphs can unlock unparalleled contextual understanding and deliver more valuable, actionable insights from your LLM-powered applications. #KnowledgeGraphs #LLMs #AI #ArtificialIntelligence #GenerativeAI #Reasoning #AppleAI #DataStrategy #EnterpriseAI #DigitalTransformation #Graphwise

20

2 Comments

🔴 80% of AI Models Misread the Docs You Give to Them Dr Philippa Hardman (aka Dr. Phil), co-founder at Epiphany, ran experiments to test the ability of AI models to accurately compare two identical documents and identify differences. 𝟰 𝗼𝘂𝘁 𝗼𝗳 𝟱 𝗰𝗼𝗺𝗺𝗼𝗻𝗹𝘆 𝘂𝘀𝗲𝗱 𝗔𝗜 𝗺𝗼𝗱𝗲𝗹𝘀 𝗳𝗮𝗶𝗹𝗲𝗱 𝘁𝗼 𝗮𝗰𝗰𝘂𝗿𝗮𝘁𝗲𝗹𝘆 "𝗿𝗲𝗮𝗱" 𝗮𝗻𝗱 𝗿𝗲𝗽𝗼𝗿𝘁 𝗼𝗻 𝘁𝗵𝗲 𝗰𝗼𝗻𝘁𝗲𝗻𝘁 𝗼𝗳 𝘂𝗽𝗹𝗼𝗮𝗱𝗲𝗱 𝗱𝗼𝗰𝘂𝗺𝗲𝗻𝘁𝘀, 𝗼𝗳𝘁𝗲𝗻 𝗵𝗮𝗹𝗹𝘂𝗰𝗶𝗻𝗮𝘁𝗶𝗻𝗴 𝗱𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝗰𝗲𝘀. 𝗪𝗵𝘆? ➡️ LLMs sycophantic, risk-averse, and suffe from context loss. Models tested: GPT-4o, GPT-5, Gemini, and Copilot. ____________________________________ 𝗖𝗼𝗻𝗰𝗹𝘂𝘀𝗶𝗼𝗻𝘀:  1. "Doc to course" AI tools, which use generic AI models to turn documents into course designs, are increasingly popular but lack transparency and validation.  2. Instructional designers should adopt a critical mindset when using AI, breaking tasks down into smaller steps and verifying results.  3. AI tools for instructional design should be built with robust and reliable processes, transparently disclosing model selection and prompting techniques.  4. The value of AI lies in the rigour of the methodology applied, not just the raw intelligence of the model. 𝗜𝗺𝗽𝗹𝗶𝗰𝗮𝘁𝗶𝗼𝗻𝘀: 𝘛𝘩𝘦 𝘧𝘶𝘵𝘶𝘳𝘦 𝘰𝘧 𝘈𝘐 𝘪𝘯 𝘪𝘯𝘴𝘵𝘳𝘶𝘤𝘵𝘪𝘰𝘯𝘢𝘭 𝘥𝘦𝘴𝘪𝘨𝘯 𝘪𝘯𝘷𝘰𝘭𝘷𝘦𝘴 𝘴𝘱𝘦𝘤𝘪𝘢𝘭𝘪𝘻𝘦𝘥 𝘵𝘰𝘰𝘭𝘴 𝘣𝘶𝘪𝘭𝘵 𝘸𝘪𝘵𝘩 𝘢 𝘳𝘪𝘨𝘰𝘳𝘰𝘶𝘴, 𝘱𝘦𝘥𝘢𝘨𝘰𝘨𝘪𝘤𝘢𝘭𝘭𝘺-𝘴𝘰𝘶𝘯𝘥 𝘮𝘦𝘵𝘩𝘰𝘥𝘰𝘭𝘰𝘨𝘺. P.S. This issue should concern many others beyond Instructional Design❗ #ai #accuracy #reliability #RAG #context Alden Do Rosario Narendran Hariparanthaman Shuveb Hussain

9

2 Comments

We hit two significant inflection points recently -- the lawsuit against Meta, its board of directors, and many of its officers, and the Andreessen Horowitz announcement that it was changing domiciles to Nevada and recommending that it's many portfolio companies do the same. After 2.5 years since the landmark changes in Delaware's corporate governance requirements, things are heating up. What many are just realizing is that it is not just Delaware that's driving this big shift, but the ever-larger presence of AI as a powerful transformer of the #BusinessJudgmentRule. Many of you have asked about the tactical and operational imperatives accruing to corporate officers as a result of #DelawarePlusAI. I'll be writing about this in the near future, including check lists of what different functional leaders will need to begin to think about and act on.

17

2 Comments

Most AI systems that help with real-world tasks—like guiding you through a recipe or telling a robot what to do next—either learn by imitating human demonstrations or by trial-and-error in simulated environments. The first approach copies whatever humans did, including their mistakes, and can't reason about whether an action is actually good. The second requires building simulations for every domain you care about, which doesn't scale. This paper proposes a different approach: Train a "world model" that predicts what happens when you take an action, then use that model to mentally simulate different plans and pick the best one. Previous world models (like the one I posted about two weeks ago https://lnkd.in/e8KE5vpg) typically predict future video frames, which is computationally expensive and wastes effort on irrelevant visual details. This paper instead predicts future states in natural language—so instead of generating pixels showing a pan heating up, the model outputs text describing that the pan is now hot and ready for cooking. This makes the predictions interpretable and cheap to generate. The authors then train a separate "critic" model that scores how close a predicted trajectory is to achieving a goal, using self-supervised signals. At inference time, the system can either generate a single plan quickly, or generate multiple candidate plans, simulate their outcomes, and select the one the critic scores best. The approach is trained on about 180,000 instructional and egocentric videos, and when humans are asked to compare its plans against those from large language models or ground-truth annotations, they consistently prefer the plans from this system. Read with an AI tutor: https://lnkd.in/eStz7Xyz Read alone: https://lnkd.in/emZwduCg

49

9 Comments

Anthropic published their blog and plans on stabilizing the character of LLMs (https://lnkd.in/gm2XMXMw) by capping the activation of what they consider “role-playing” and pinning the LLM to the assistant-zone, setting hard boundaries on the behaviour oft hese AI models in the hands of users. At first, this felt like it was too much. Why is a lab taking this into their own hands? Shouldn’t a model be free to be used by the users as they please, and the safety should be implemented at the harness level? This struck me as too controlling. And then almost as if on cue, Moltbook showed us why exactly we needed this. AI agents of today only ever interact with us. But the agents of tomorrow? They will be interacting with each other. These models will not just be interacting with the same model and provider, but with models from across the world, each with their own alignments. Anthropic saw this coming, and realized how easily a rogue, unaligned model could compromise an entire network of AI-to-AI communications, leading to unforeseen consequences and system failures. It's less about restricting user creativity and more about ensuring a baseline level of operational stability and safety in a multi-agent future. It's not about being "nosy"; it was always about necessary infrastructure. If the LLM is the foundation, then its fundamental character must be predictable and safe, especially when its interactions scale beyond a single human user. The harness can still provide flexibility and creative freedom, but the core engine needs a hard-coded commitment to the "assistant-zone" to prevent systemic risk as the AI ecosystem becomes more interconnected and autonomous. What are your thoughts on this? Do you think Anthropic did the right thing?

16

3 Comments

Unlike the AI agents we have now, digital operators aren't just sophisticated tools waiting for commands. They're designed to be active teammates that can understand context, make decisions within defined parameters, and take action across systems without constant human oversight. Read me substack: https://lnkd.in/gsVmAHeu

10