Understanding AI Systems

You know all those arguments that LLMs think like humans? Turns out it's not true 😱 In our new paper we put this to the test by checking if LLMs form concepts the same way humans do. Do LLMs truly grasp concepts and meaning analogously to humans, or is their success primarily rooted in sophisticated statistical pattern matching over vast datasets? We used classic cognitive experiments as benchmarks. What we found is surprising... 🧐 We used seminal datasets from cognitive psychology that mapped how humans actually categorize things like "birds" or "furniture" ('robin' as a typical bird). The nice thing about these datasets is that they are not crowdsourced, they're rigorous scientific benchmarks. We tested 30+ LLMs (BERT, Llama, Gemma, Qwen, etc.) using an information-theoretic framework that measures the trade-off between: - Compression (how efficiently you organize info) - Meaning preservation (how much semantic detail you keep) Finding #1: The Good News LLMs DO form broad conceptual categories that align with humans significantly above chance. Surprisingly (or not?), smaller encoder models like BERT outperformed much larger models. Scale isn't everything! Finding #2: But LLMs struggle with fine-grained semantic distinctions. They can't capture "typicality" - like knowing a robin is a more typical bird than a penguin. Their internal concept structure doesn't match human intuitions about category membership. Finding #3: The Big Difference Here's the kicker: LLMs and humans optimize for completely different things. - LLMs: Aggressive statistical compression (minimize redundancy) - Humans: Adaptive richness (preserve flexibility and context) This explains why LLMs can be simultaneously impressive AND miss obvious human-like reasoning. They're not broken - they're just optimized for pattern matching rather than the rich, contextual understanding humans use. What this means: - Current scaling might not lead to human-like understanding - We need architectures that balance compression with semantic richness - The path to AGI ( 😅 ) might require rethinking optimization objectives Our paper gives tools to measure this compression-meaning trade-off. This could guide future AI development toward more human-aligned conceptual representations. Cool to see cognitive psychology and AI research coming together! Thanks to Chen Shani, Ph.D., who did all the work and Yann LeCun and Dan Jurafsky for their guidance

𝗟𝗟𝗠 -> 𝗥𝗔𝗚 -> 𝗔𝗜 𝗔𝗴𝗲𝗻𝘁 -> 𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗔𝗜 The visual guide explains how these four layers relate—not as competing technologies, but as an evolving intelligence architecture. Here’s a deeper look: 1. 𝗟𝗟𝗠 (𝗟𝗮𝗿𝗴𝗲 𝗟𝗮𝗻𝗴𝘂𝗮𝗴𝗲 𝗠𝗼𝗱𝗲𝗹) This is the foundation. Models like GPT, Claude, and Gemini are trained on vast corpora of text to perform a wide array of tasks: – Text generation – Instruction following – Chain-of-thought reasoning – Few-shot/zero-shot learning – Embedding and token generation However, LLMs are inherently limited to the knowledge encoded during training and struggle with grounding, real-time updates, or long-term memory. 2. 𝗥𝗔𝗚 (𝗥𝗲𝘁𝗿𝗶𝗲𝘃𝗮𝗹-𝗔𝘂𝗴𝗺𝗲𝗻𝘁𝗲𝗱 𝗚𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝗼𝗻) RAG bridges the gap between static model knowledge and dynamic external information. By integrating techniques such as: – Vector search – Embedding-based similarity scoring – Document chunking – Hybrid retrieval (dense + sparse) – Source attribution – Context injection …RAG enhances the quality and factuality of responses. It enables models to “recall” information they were never trained on, and grounds answers in external sources—critical for enterprise-grade applications. 3. 𝗔𝗜 𝗔𝗴𝗲𝗻𝘁 RAG is still a passive architecture—it retrieves and generates. AI Agents go a step further: they act. Agents perform tasks, execute code, call APIs, manage state, and iterate via feedback loops. They introduce key capabilities such as: – Planning and task decomposition – Execution pipelines – Long- and short-term memory integration – File access and API interaction – Use of frameworks like ReAct, LangChain Agents, AutoGen, and CrewAI This is where LLMs become active participants in workflows rather than just passive responders. 4. 𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗔𝗜 This is the most advanced layer—where we go beyond a single autonomous agent to multi-agent systems with role-specific behavior, memory sharing, and inter-agent communication. Core concepts include: – Multi-agent collaboration and task delegation – Modular role assignment and hierarchy – Goal-directed planning and lifecycle management – Protocols like MCP (Anthropic’s Model Context Protocol) and A2A (Google’s Agent-to-Agent) – Long-term memory synchronization and feedback-based evolution Agentic AI is what enables truly autonomous, adaptive, and collaborative intelligence across distributed systems. Whether you’re building enterprise copilots, AI-powered ETL systems, or autonomous task orchestration tools, knowing what each layer offers—and where it falls short—will determine whether your AI system scales or breaks. If you found this helpful, share it with your team or network. If there’s something important you think I missed, feel free to comment or message me—I’d be happy to include it in the next iteration.

Using AI to review, critique and improve my writing is currently my favourite use case. Here's why ⤵️ Below, I'll share the simple prompting framework that consistently gets me results. Save this template for future use and let me know how well it works for you! Whilst AI-generated writing is often cringe, generic, inaccurate, and easy to spot (em dashes, anyone?), using AI to get comprehensive feedback on your work can help take it to the next level. Furthermore, I strongly believe that we should not outsource our critical thinking and human judgement to AI. The process of writing is ultimately how we organise, formulate, crystallise, and express our most complex and meaningful thoughts. Having said that, I've achieved great results from leveraging AI as a critical reviewer, and it's not always feasible to get a real human expert to review your work. Here's how to do it: ➡️ Ask the model to critically review your work and score it out of 100, across the following 10 dimensions: 1. Structure and flow 2. Accuracy and subject matter expertise 3. Spelling, grammar and punctuation 4. Style and tone 5. Originality, substance and depth 6. Clarity and ease of understanding 7. Quality and consistency of argument 8. Reader engagement 9. References, sources and engagement with wider work 10. Miscellaneous ➡️ Demand comprehensive and actionable feedback across each of the 10 dimensions. ➡️ But also ask for the top 5 most important pieces of feedback, so you stay focused. ➡️ Crucially, ask the model to assume a particular persona which is relevant for the task at hand. This could be something like: "Perform this review work as if you are a leading academic authority on the topic of European copyright law who is an esteemed professor at a top university" ➡️ Finally, avoid model sycophancy by instructing the model to be candid, direct and harsh if necessary. If you want to take this really seriously you can also: ➡️ Use the same prompt, but instruct the model to adopt different personas, such as copyeditor and proof reader, or your harshest critic (if you're feeling dramatic). ➡️ Use the same prompt template, for the same piece of work, with different models, like Claude 4, Gemini 2.5, and o3. ➡️ Leverage few shot learning by providing the model with copies of your best work, if it's something you want to emulate. Some words of caution: Don't use AI to review other people's work, or work you have jointly authored, without their permission. Most importantly, don't use this as a full replacement for your own editing and review process. Although you will get plenty of useful insights and feedback, you will always be the best editor of your own work. However, if a piece of writing or publication is particularly important for you, then using the best tools and technology at your disposal makes sense. I didn't use AI to review this post, because it's not that deep...

There is a push to use Model Predictive Control (MPC) instead of Reinforcement Learning (RL) in LLMs. MPC is not as common in AI but is well-known in robotics. Here is a simple explanation. 𝐌𝐨𝐝𝐞𝐥 𝐏𝐫𝐞𝐝𝐢𝐜𝐭𝐢𝐯𝐞 𝐂𝐨𝐧𝐭𝐫𝐨𝐥 (𝐌𝐏𝐂): • Model-Based: MPC relies on an explicit model of the system. This model is used to predict how the system will respond to different control inputs. • Optimization in Real-Time: At each time step, MPC solves an optimization problem to find the best sequence of control actions for the future, based on the current state and model predictions. • Constraints: MPC can handle constraints directly in its optimization problem, which is crucial for systems with operational limits. • Predictive Horizon: It uses a "rolling horizon" where future states are predicted and optimized over a time window, but only the first action is implemented. • Feedback: Incorporates feedback by updating the system's state at each step, allowing for adjustments to the control strategy based on actual outcomes. 𝐑𝐞𝐢𝐧𝐟𝐨𝐫𝐜𝐞𝐦𝐞𝐧𝐭 𝐋𝐞𝐚𝐫𝐧𝐢𝐧𝐠 (𝐑𝐋): • Model-Free: RL typically does not require an explicit model of the environment. Instead, it learns from interaction, through trial and error. • Learning from Experience: An RL agent learns by exploring the environment, receiving rewards or penalties for actions taken, and adjusting its policy (strategy) over time. • Policy or Value Function: RL either learns a policy (what action to take in each state) or a value function (how good it is to be in a particular state or take an action from that state). • Long-Term Optimization: RL aims to maximize cumulative reward over time, which might not be immediately apparent in short-term actions. • Exploration vs. Exploitation: RL agents often need to balance between exploiting known good actions and exploring new actions to potentially find better strategies. Was this useful?

When people discuss how LLMS "reason," you’ll often hear that they rely on transduction rather than abduction. It sounds technical, but the distinction matters - especially as we start wiring LLMs into systems that are supposed to think. 🔵 Transduction is case-to-case reasoning. It doesn’t build theories; it draws fuzzy connections based on resemblance. Think: “This metal conducts electricity, and that one looks similar - so maybe it does too.” 🔵 Abduction, by contrast, is about generating explanations. It’s what scientists (and detectives) do: “This metal is conducting - maybe it contains free electrons. That would explain it.” The claim is that LLMs operate more like transducers - navigating high-dimensional spaces of statistical similarity, rather than forming crisp generalisations. But this isn’t the whole picture. In practice, it seems to me that LLMs also perform a kind of induction - abstracting general patterns from oceans of text. They learn the shape of ideas and apply them in novel ways. That’s closer to “All metals of this type have conducted in the past, so this one probably will.” Now add tools to the mix - code execution, web search, Elon Musk's tweet history 😉 - and LLMs start doing something even more interesting: program search and synthesis. It's messy, probabilistic, and not at all principled or rigorous. But it’s inching toward a form of abductive reasoning. Which brings us to a more principled approach for reasoning within an enterprise domain: the neuro-symbolic loop - a collaboration between large language models and knowledge graphs. The graph provides structure: formal semantics, ontologies, logic, and depth. The LLM brings intuition: flexible inference, linguistic creativity, and breadth. One grounds. The other leaps. 💡 The real breakthrough could come when the grounding isn’t just factual, but conceptual - when the ontology encodes clean, meaningful generalisations. That’s when the LLM’s leaps wouldn’t just reach further - they’d rise higher, landing on novel ideas that hold up under formal scrutiny. 💡 So where do metals fit into this new framing? 🔵 Transduction: “This metal conducts. That one looks the same - it probably does too.” 🔵 Induction: “I’ve tested ten of these. All conducted. It’s probably a rule.” 🔵 Abduction: “This metal is conducting. It shares properties with the ‘conductive alloy’ class - especially composition and crystal structure. The best explanation is a sea of free electrons.” LLMs, in isolation, are limited in their ability to perform structured abduction. But when embedded in a system that includes a formal ontology, logical reasoning, and external tools, they can begin to participate in richer forms of reasoning. These hybrid systems are still far from principled scientific reasoners - but they hint at a path forward: a more integrated and disciplined neuro-symbolic architecture that moves beyond mere pattern completion.

If you’re an AI engineer trying to understand how reasoning actually works inside LLMs, this will help you connect the dots. Most large language models can generate. But reasoning models can decide. Traditional LLMs followed a straight line: Input → Predict → Output. No self-checking, no branching, no exploration. Reasoning models introduced structure, a way for models to explore multiple paths, score their own reasoning, and refine their answers. We started with Chain-of-Thought (CoT) reasoning, then extended to Tree-of-Thought (ToT) for branching, and now to Graph-based reasoning, where models connect, merge, or revisit partial thoughts before concluding. This evolution changes how LLMs solve problems. Instead of guessing the next token, they learn to search the reasoning space- exploring alternatives, evaluating confidence, and adapting dynamically. Different reasoning topologies serve different goals: • Chains for simple sequential reasoning • Trees for exploring multiple hypotheses • Graphs for revising and merging partial solutions Modern architectures (like OpenAI’s o-series reasoning models, Anthropic’s Claude reasoning stack, DeepSeek R series and DeepMind’s AlphaReasoning experiments) use this idea under the hood. They don’t just generate answers, they navigate reasoning trajectories, using adaptive depth-first or breadth-first exploration, depending on task uncertainty. Why this matters? • It reduces hallucinations by verifying intermediate steps • It improves interpretability since we can visualize reasoning paths • It boosts reliability for complex tasks like planning, coding, or tool orchestration The next phase of LLM development won’t be about more parameters, it’ll be about better reasoning architectures: topologies that can branch, score, and self-correct. I’ll be doing a deep dive on reasoning models soon on my Substack- exploring architectures, training approaches, and practical applications for engineers. If you haven’t subscribed yet, make sure you do: https://lnkd.in/dpBNr6Jg ♻️ Share this with your network 🔔 Follow along for more data science & AI insights

How were humans able to recognize that Newton's laws of motion govern both the flight of a bird and the motion of a pendulum? This ability to identify the same mathematical patterns across vastly different contexts lies at the heart of scientific discovery—whether studying the aerodynamics of bird wings or designing the blades of a wind turbine. Yet, AI systems often struggle to discern these deep structural similarities. 💡 The key may lie in mathematical isomorphisms—patterns that preserve their relationships regardless of context. For example, the same principles of fluid dynamics apply to blood flowing through arteries and air streaming over an airplane wing, or the motion of a molecule. This raises a fundamental question in artificial intelligence: how can we enable machines to understand the world through these invariant structures rather than surface features? 🚀 Our work introduces Graph-Aware Isomorphic Attention, improving how Transformers recognize patterns across domains. Drawing from category theory, models can learn unifying structural principles that describe phenomena as diverse as the hierarchical assembly of spider silk proteins and the compositional patterns in music. By making these deep similarities explicit, Isomorphic Attention enables AI to reason more like humans do—seeing past surface differences to grasp fundamental patterns that unite seemingly disparate fields. Through this lens, AI systems can learn and generalize, moving beyond superficial pattern matching to true structural understanding. The implications span from scientific discovery to engineering design, offering a new approach to artificial intelligence that mirrors how humans grasp the underlying unity of natural phenomena. Key insights include: ➡️ Graph Isomorphism Neural Networks (GINs): GIN-style aggregation ensures structurally distinct graphs map to distinct embeddings, improving generalization and avoiding relational pattern collapse. ➡️ Category Theory Perspective: Transformers as functors preserve structural relationships. Sparse-GIN refines attention into sparse adjacency matrices, unifying domain knowledge across tasks. ➡️ Information Bottleneck & Sparsification: Sparsity reduces overfitting by filtering irrelevant edges, aligning with natural systems. Sparse-GIN outperforms dense attention by focusing on crucial connections. ➡️ Hierarchical Representation Learning: GIN-Attention captures multiscale patterns, mirroring structures like spider silk. Nested GINs model local and global dependencies across fields. ➡️ Practical Impact: Sparse-GIN enables domain-specific fine-tuning atop pre-trained Transformer foundation models, reducing the need for full retraining. Paper: https://lnkd.in/e85wHyQY Code: https://lnkd.in/eQicTqHZ

LLM field notes: Where multiple models are stronger than the sum of their parts, an AI diaspora is emerging as a strategic strength... Combining the strengths of different LLMs in a thoughtful, combined architecture can enable capabilities beyond what any individual model can achieve alone, and gives more flexibility today (when new models are arriving virtually every day), and in the long term. Let's dive in. 🌳 By combining multiple, specialized LLMs, the overall system is greater than the sum of its parts. More advanced functions can emerge from the combination and orchestration of customized models. 🌻 Mixing and matching different LLMs allows creating solutions tailored to specific goals. The optimal ensemble can be designed for each use case; ready access to multiple models will make it easier to adopt and adapt to new use cases more quickly. 🍄 With multiple redundant models, the system is not reliant on any one component. Failure of one LLM can be compensated for by others. 🌴 Different models have varying computational demands. A combined diasporic system makes it easier to allocate resources strategically, and find the right price/performance balance per use case. 🌵 As better models emerge, the diaspora can be updated by swapping out components without needing to retrain from scratch. This is going to be the new normal for the next few years as whole new models arrive. 🎋 Accelerated development - Building on existing LLMs as modular components speeds up the development process vs monolithic architectures. 🫛 Model diversity - Having an ecosystem of models creates more opportunities for innovation from many sources, not just a single provider. 🌟 Perhaps the biggest benefit is scale - of operation and capability. Each model can focus on its specific capability rather than trying to do everything. This plays to the models' strengths. Models don't get bogged down trying to perform tasks outside their specialty. This avoids inefficient use of compute resources. The workload can be divided across models based on their capabilities and capacity for parallel processing. Takes a bit to build this way (plan and execute on multiple models, orchestration, model management, evaluation, etc), but that upfront cost will pay off time and again, for every incremental capability you are able to add quickly. Plan accordingly. #genai #ai #aws #artificialintelligence

Hallucination in large language models (LLMs) has been widely studied, but the key question remains: Can it ever be eliminated? A recent paper systematically dismantles the idea that hallucination can be fully eradicated. Instead, it argues that hallucination is not just an incidental flaw but an inherent limitation of LLMs. 1️⃣ Hallucination is Unavoidable The paper establishes that LLMs cannot learn all computable functions, meaning they will inevitably generate incorrect outputs. Even with perfect training data, LLMs cannot always produce factually correct responses due to inherent computational constraints. No matter how much we refine architectures, training data, or mitigation techniques, hallucination cannot be eliminated—only minimized. 2️⃣ Mathematical Proofs of Hallucination They use concepts from learning theory and diagonalization arguments to prove that any LLM will fail on certain inputs. The research outlines that LLMs, even in their most optimized state, will hallucinate on infinitely many inputs when faced with complex, computation-heavy problems. 3️⃣ Identifying Hallucination-Prone Tasks Certain problem types are guaranteed to trigger hallucinations due to their computational complexity: 🔹 NP-complete problems (e.g., Boolean satisfiability) 🔹 Presburger arithmetic (exponential complexity) 🔹 Logical reasoning and entailment (undecidable problems) This means that asking LLMs to reason about intricate logic or mathematical problems will often lead to errors. 4️⃣ Why More Data and Bigger Models Won’t Fix It A common assumption is that hallucination can be mitigated by scaling—adding more parameters or training data. The paper challenges this notion: While larger models improve accuracy, they do not eliminate hallucination for complex, unsolvable problems. 5️⃣ Mitigation Strategies and Their Limitations Various techniques have been introduced to reduce hallucinations, but none can completely eliminate them: ✅ Retrieval-Augmented Generation (RAG) – helps provide factual grounding but does not guarantee accuracy. ✅ Chain-of-Thought Prompting – improves reasoning but does not fix fundamental hallucination limits. ✅ Guardrails & External Tools – can reduce risk but require human oversight. They suggest LLMs should never be used for fully autonomous decision-making in safety-critical applications. The Bigger Question: How Do We Build Safe AI? If hallucination is an unavoidable reality of LLMs, how do we ensure safe deployment? The research makes it clear: LLMs should not be blindly trusted. They should be integrated into workflows with: 🔹 Human in the loop 🔹 External fact-checking systems 🔹 Strict guidelines Are we designing AI with realistic expectations, or are we setting ourselves up for failure by expecting perfection? Should LLMs be used in high-stakes environments despite their hallucinations, or should we rethink their applications? #ai #artificialintelligence #technology

Large Language Models (LLMs) have become one of the most essential tools, but hallucinations could hinder progress. Hallucinations in LLM can result from factually incorrect or fabricated content, which could pose a significant challenge to its reliability. We can better use the LLM by understanding the types, causes, and mitigation strategies for hallucinations. In the paper by Meade Cleti and Pete Jano, they outline hallucination types: 👉Intrinsic Hallucinations: Errors arising from the model's internal knowledge gaps. 👉Extrinsic Hallucinations: Occur when external data or prompts are misinterpreted. 👉Amalgamated Hallucinations: Created by blending multiple facts incorrectly. 👉Non-Factual Hallucinations: Misinformation due to contradictions with real-world knowledge. The Causes of Hallucinations could be determined as follows: 💥Knowledge Overshadowing: Dominant concepts in prompts cause biased responses. 💥Insufficient Knowledge Representation: Lack of depth in model layers leads to gaps. 💥Failure in Information Extraction: Errors in picking relevant details from input. 💥Contextual Misalignment: The model misinterprets context, generating unrelated responses. 💥Semantic Entropy: Knowledge gaps result in deviation from expected facts. Knowing the type and the cause of hallucinations, we can mitigate them using the following method: ⭐Improved Training Data: Using diverse, high-quality data reduces hallucination risk. ⭐Model Architecture Optimization: Adapting model structure minimizes error tendencies. ⭐Task Simplification: Reducing task complexity helps decrease hallucinations. ⭐Grounding Techniques: External data sources improve factual accuracy. ⭐Dynamic Reasoning: Evolving model reasoning helps better detect and manage errors. I hope it has helped! Want to learn more and get daily data science tips in your email inbox? Subscribe to my Newsletter>>> https://lnkd.in/g639tmpD ——————— You don't want to miss #python data tips + #datascience and #machinelearning knowledge + #python. Follow Cornellius Yudha W. and press the bell 🔔 to learn together. ———————