<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Ujwal Tickoo on Medium]]></title>
        <description><![CDATA[Stories by Ujwal Tickoo on Medium]]></description>
        <link>https://medium.com/@ujwaltickoo?source=rss-b4d2217d1709------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*U5TvvD0Jd_cF8YchqvVK9A.jpeg</url>
            <title>Stories by Ujwal Tickoo on Medium</title>
            <link>https://medium.com/@ujwaltickoo?source=rss-b4d2217d1709------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 11:51:06 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@ujwaltickoo/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Reasoning. Planning. Action. On Demand: How Agentic AI Drives Value from LLMs]]></title>
            <link>https://medium.com/@ujwaltickoo/reasoning-planning-action-on-demand-how-agentic-ai-drives-value-from-llms-983b879c07ee?source=rss-b4d2217d1709------2</link>
            <guid isPermaLink="false">https://medium.com/p/983b879c07ee</guid>
            <dc:creator><![CDATA[Ujwal Tickoo]]></dc:creator>
            <pubDate>Sun, 07 Jun 2026 08:11:00 GMT</pubDate>
            <atom:updated>2026-06-07T08:11:00.249Z</atom:updated>
            <content:encoded><![CDATA[<h3>Didn’t LLMs already solve AI? Why do we suddenly need Agentic AI?</h3><p>If you are confused about why we need Agentic AI after the massive wave of excitement around LLMs, you are not alone.</p><p>After all, LLMs can already answer questions, write code, summarize documents, generate content, and even reason through complex problems. So what exactly is Agentic AI adding?</p><p>The short answer is this:</p><blockquote><strong><em>LLMs can reason about a problem. Agents can reason about it, gather information, plan what to do, and take actions to solve it.</em></strong></blockquote><p>That distinction may sound subtle, but it represents one of the most important shifts in enterprise software and AI.</p><p>An LLM can help you think through a problem. It can suggest approaches, explain tradeoffs, and generate plans. But solving real-world problems usually requires more than reasoning. It requires access to information, interaction with systems, and the ability to take actions.</p><p>This is where Agentic AI enters the picture.</p><p>Agentic AI combines the reasoning capabilities of LLMs with tools, business systems, data sources, and actions. It transforms intelligence into outcomes.</p><p>The result is a new software paradigm where reasoning and actions become available on demand.</p><h3>The Journey from Software to LLMs to Agents</h3><p>To understand why Agentic AI matters, it helps to look at how software has evolved.</p><h3>Traditional Software</h3><p>Traditional software is built around predefined rules.</p><p>Developers anticipate user needs, define business logic, and create workflows.</p><p>The software can only do what somebody programmed it to do.</p><p>This model works well when the world is predictable.</p><p>But many business problems are not.</p><h3>LLMs</h3><p>LLMs introduced something fundamentally new.</p><p>Instead of relying entirely on predefined logic, they can reason about situations they have never explicitly encountered before.</p><p>Ask an LLM:</p><blockquote><em>“How would you compare these three product strategies?”</em></blockquote><p>It can analyze and synthesize information.</p><p>Ask:</p><blockquote><em>“What might be causing this issue?”</em></blockquote><p>It can evaluate possibilities and propose explanations.</p><p>This is a significant leap.</p><p>For the first time, software systems can reason about a problem rather than simply execute predefined logic.</p><p>But there is an important limitation.</p><p><strong><em>Reasoning about a problem is not the same as solving it.</em></strong></p><h3>Agents</h3><p>Agents extend the capabilities of LLMs beyond reasoning.</p><p>They can gather information, interact with systems, use tools, execute workflows, and take actions.</p><p>This is what transforms intelligence into outcomes.</p><p>An agent doesn’t simply answer a question.</p><p>It works toward achieving a goal.</p><p>That is why many people view Agentic AI as the next logical step in the evolution of AI systems.</p><h3>The Limits of Traditional Software and Automation</h3><p>Traditional automation follows a familiar pattern:</p><p><strong>Request → Workflow → Response</strong></p><p>For example:</p><ul><li>Reset a password</li><li>Check an account balance</li><li>Book an appointment</li><li>Track an order</li></ul><p>These are predictable tasks.</p><p>The workflow is known in advance.</p><p>But real-world customer interactions are rarely so predictable.</p><p>Consider questions such as:</p><blockquote><em>“What’s the fastest way to get to your nearest store from where I am right now?”</em></blockquote><blockquote><em>“Can you recommend another restaurant nearby if this one is full?”</em></blockquote><blockquote><em>“Which store near me has this product in stock?”</em></blockquote><blockquote><em>“How difficult will parking be near the restaurant tonight?”</em></blockquote><blockquote>The challenge is no longer executing a workflow. The challenge is determining what workflow should exist in the first place.</blockquote><p><strong>Traditional software excels at executing logic.</strong></p><p><strong>The problem is that someone must define the logic ahead of time.</strong></p><p><strong>As customer needs become more contextual and dynamic, the number of possible scenarios becomes virtually infinite.</strong></p><p><strong>This is where predefined workflows begin to break down.</strong></p><h3>Why LLMs Alone Are Not Enough</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AI1meeGe1V4QdmAOSBTgBQ.png" /></figure><p>This is where many discussions about AI stop too early. An LLM may understand a customer’s request perfectly. It may even reason about how the problem could be solved. But reasoning alone rarely solves the problem.Imagine a customer asks:</p><blockquote><em>“Find the nearest restaurant with available seating and give me directions.”</em></blockquote><p>The LLM understands the goal. It can identify the information required. But it cannot solve the problem without access to:</p><ul><li>Maps</li><li>Restaurant databases</li><li>Availability information</li><li>Directions services</li></ul><p>Similarly, if a customer asks:</p><blockquote><em>“Is parking available near the restaurant tonight?”</em></blockquote><p>The model can reason about the problem. But it cannot answer the question unless it can gather information from external sources. Understanding and reasoning are necessary. They are not sufficient. The missing piece is action.</p><h3>Agentic AI: Connecting Reasoning to Action</h3><p>This is where Agentic AI creates value. Agentic systems combine reasoning with tool usage and actions. They can:</p><ul><li>Access maps and navigation services</li><li>Query inventory systems</li><li>Search knowledge repositories</li><li>Retrieve information from enterprise applications</li><li>Call external APIs</li><li>Trigger workflows</li><li>Execute business processes</li></ul><p>The agent becomes an orchestration layer. It reasons about the problem, determines what information is needed, selects the appropriate tools, gathers information, and takes actions to achieve an outcome.</p><p>Instead of following a predefined workflow, it assembles the workflow dynamically based on the situation. That is the fundamental shift.</p><h3>A Simple Restaurant Example</h3><p>Imagine a customer asks:</p><blockquote><em>“Can I find parking near the restaurant tonight?”</em></blockquote><p>A traditional chatbot may not even have a predefined intent for parking availability. An LLM can reason about the problem. It can recognize that the customer is trying to determine whether visiting the restaurant will be convenient. But reasoning alone does not provide the answer.</p><p>An agentic system can:</p><ol><li>Identify the restaurant location.</li><li>Access mapping services.</li><li>Check nearby parking options.</li><li>Consider traffic conditions.</li><li>Look for local events.</li><li>Generate a recommendation.</li></ol><p>The response might be:</p><blockquote><em>“Street parking is limited tonight because of a nearby event. The parking garage two blocks away currently has availability and is your best option.”</em></blockquote><p><strong><em>Notice what happened. The value did not come from generating text. The value came from solving a problem. The text was simply the interface through which the solution was delivered.</em></strong></p><h3>The New Value Chain</h3><p>The progression looks something like this:</p><p><strong>Natural Language Understanding</strong><br>↓<br><strong>Reasoning About the Problem</strong><br>↓<br><strong>Information Gathering</strong><br>↓<br><strong>Tool Use</strong><br>↓<br><strong>Actions</strong><br>↓<br><strong>Problem Solving</strong><br>↓<br><strong>Business Value</strong></p><p>Each layer builds upon the previous one. Understanding alone is not enough. Reasoning alone is not enough. Information alone is not enough. Business value emerges when all of them work together to solve a customer’s problem.</p><h3>From Task Automation to Problem Automation</h3><p>For decades, businesses have focused on automating tasks. Agentic AI expands the opportunity to automating problem solving. That is a much larger opportunity.</p><p><strong><em>Tasks are predefined. Problems are often unique.</em></strong></p><p><strong><em>Tasks execute known workflows. Problems require discovering the right workflow in real time.</em></strong></p><p>This is why Agentic AI feels fundamentally different from previous generations of automation. <strong>It is not simply another workflow engine. It is a system that can reason about goals, gather information, and determine how to achieve outcomes.</strong></p><h3>Summary</h3><ul><li><strong>Reasoning</strong> → Understand the problem.</li><li><strong>Planning</strong> → Determine how to solve it.</li><li><strong>Action</strong> → Execute through tools and systems.</li><li><strong>On Demand</strong> → The key differentiator from traditional software, where logic must be predefined.</li></ul><h3>Final Thought</h3><p>The first wave of AI focused on understanding language. The current wave of LLMs introduced the ability to reason about problems. The next wave is about turning that reasoning into action. LLMs can reason about a problem. Agents can reason about it, gather information, and take actions to solve it. That may sound like a small distinction, but it is the difference between generating an answer and achieving an outcome.</p><p>The real value of Agentic AI is not that it can have a conversation. The real value is that it can connect reasoning, information, tools, and actions to solve problems that were previously too dynamic, too contextual, or too unpredictable to automate.</p><p>That is how AI moves beyond chatbots and becomes a genuine business capability.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=983b879c07ee" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Insights from Anthropic in Building Real Enterprise Systems — Beyond the Hype]]></title>
            <link>https://medium.com/@ujwaltickoo/insights-from-anthropic-in-building-real-enterprise-systems-beyond-the-hype-124c9ebc1b5f?source=rss-b4d2217d1709------2</link>
            <guid isPermaLink="false">https://medium.com/p/124c9ebc1b5f</guid>
            <category><![CDATA[automation]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[agents]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <dc:creator><![CDATA[Ujwal Tickoo]]></dc:creator>
            <pubDate>Sun, 01 Feb 2026 17:06:20 GMT</pubDate>
            <atom:updated>2026-02-01T17:06:20.157Z</atom:updated>
            <content:encoded><![CDATA[<h3>Insights from Anthropic in Building Real Enterprise Systems — Beyond the Hype</h3><p>Anthropic, the company behind Claude, recently published a candid and excellent perspective titled <a href="https://claude.com/blog/building-multi-agent-systems-when-and-how-to-use-them"><em>“Building Multi-Agent Systems: When (and When Not) to Use Them.”</em></a> What follows blends insights from that post with my own experience building and scaling real enterprise software systems.</p><p>There is significant noise right now. Many conversations around “agentic AI” are driven more by novelty than by experience in building real software or automation systems at scale. This often leads to architectural choices that look modern, but fail under production constraints.</p><p>Multi-agent AI systems are having a moment. From agent “swarms” to autonomous teams of LLMs collaborating on complex problems, it’s tempting to assume that more agents = more intelligence. In practice, that assumption is often wrong.</p><p>Multi-agent systems can be powerful — but only when used intentionally and sparingly. Most real-world applications are better served by a single, well-designed agent. This post lays out insights from Anthropic’s original post, married with my insights, on when multi-agent architectures make sense, when they don’t, and how to design them responsibly.</p><p><strong>What Is a Multi-Agent System?</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IgyRL46SdmdIweTxXv2XHQ.png" /><figcaption>When Multi-Agent sytems make sense</figcaption></figure><p>A multi-agent system consists of multiple AI agents (typically separate LLM instances), each with its own context and role, that coordinate through code to complete a task.</p><p>Instead of one agent reasoning end-to-end, work is split across agents that may:</p><p>- Run in parallel</p><p>- Specialize in different domains</p><p>- Hand off partial results to a coordinator</p><p>Think of it less like a “bigger brain” and more like a team of specialists with a manager.</p><h3>Start With One Agent (Seriously)</h3><p>The most important rule is also the most ignored:</p><p>Always start with a single agent.</p><p>A thoughtfully designed single agent — with:</p><p>- Clear instructions</p><p>- Strong tool access</p><p>- Structured outputs</p><p>- Guardrails</p><p>…can handle the majority of enterprise workflows.</p><p>Why single agents usually win:</p><p>- Lower cost: Multi-agent systems often consume 3–10× more tokens.</p><p>- Less complexity: Fewer prompts, fewer failure modes.</p><p>- Easier debugging: One reasoning chain instead of many interacting ones.</p><p><strong>If your system works with one agent, adding more agents usually makes it worse, not better.</strong></p><h3><strong>When Multi-Agent Systems Actually Add Value</strong></h3><p>Multi-agent architectures shine in three specific scenarios. Outside of these, they are usually overkill.</p><p><strong>1. Context Protection: Keeping Reasoning Clean</strong></p><p>LLMs degrade when their context becomes cluttered with irrelevant information. Multi-agent systems help by isolating context.</p><p>Example:</p><p>You’re building an enterprise research assistant:</p><p>- One agent gathers raw documents.</p><p>- Another summarizes them.</p><p>- A third produces an executive-level synthesis.</p><p>Each agent sees only what it needs. The final reasoning agent is protected from noise, PDFs, and raw logs — leading to higher-quality outputs.</p><p>Use agents when context pollution hurts reasoning quality.</p><p><strong>2. Parallelization: Exploring Many Paths at Once</strong></p><p>Some problems benefit from breadth, not just depth.</p><p>Example:</p><p>Market or strategy analysis:</p><p>- Agent A explores optimistic scenarios.</p><p>- Agent B explores pessimistic scenarios.</p><p>- Agent C looks for historical analogs.</p><p>- A coordinator compares and synthesizes insights.</p><p>A single agent can do this sequentially, but parallel agents:</p><p>- Cover more ground</p><p>- Surface more diverse ideas</p><p>- Reduce time to insight</p><p>Use agents when independent exploration improves results.</p><p><strong>3. Specialization: Different Tools, Different Minds</strong></p><p>As soon as an agent has too many tools or roles, performance suffers.</p><p>Example:</p><p>An internal enterprise platform:</p><p>- A data agent queries SQL and cleans data.</p><p>- A compliance agent checks regulatory constraints.</p><p>- A communication agent drafts user-facing explanations.</p><p>Each agent:</p><p>- Has a narrower prompt</p><p>- Uses fewer tools</p><p>- Makes fewer mistakes</p><p>Use agents when tool or domain specialization is real, not cosmetic.</p><h3>Common Pitfalls (Where Most Teams Go Wrong)</h3><p>Over-engineering early: Teams often build agent frameworks before proving the core task works. This slows learning and obscures real problems.</p><p>Excessive handoffs:</p><p>Every agent-to-agent handoff:</p><p>- Costs tokens</p><p>- Loses nuance</p><p>- Introduces interpretation errors</p><p>Magical thinking: agents do not “self-organize” simply because they are named. Coordination logic still needs to be designed explicitly.</p><p>A Practical Starting Architecture: Orchestrator + Sub-Agents</p><p>The simplest and most reliable pattern is:</p><p>One orchestrator agent + multiple specialist sub-agents</p><p>How it works:</p><p>1. The orchestrator understands the user goal.</p><p>2. It delegates clearly scoped tasks to sub-agents.</p><p>3. Sub-agents return structured results.</p><p>4. The orchestrator synthesizes the final output.</p><p>This mirrors how real teams work — and is far easier to debug than peer-to-peer agent networks.</p><p>A Simple Design Test</p><p>Before adding another agent, ask:</p><blockquote><strong>Is the uncertainty about what to do — or how to do it?</strong></blockquote><p><strong>- If it’s what to do → agentic reasoning may help.</strong></p><p><strong>- If it’s how to do it → a deterministic workflow is often better.</strong></p><p>This single question prevents a large class of bad designs.</p><h3>Final Takeaways</h3><p>Multi-agent systems are not the default — they are a specialized tool. Start with a single agent and scale only when there’s clear evidence.</p><p>Use agents for:</p><p>- Context protection</p><p>- Parallel exploration</p><p>- Genuine specialization</p><p>- Prefer simple orchestration patterns over clever agent swarms.</p><p><strong>In enterprise AI especially, restraint is a feature. The best systems feel boring on the inside — and powerful on the outside.</strong></p><h3>Rule-based Process Automation vs Agentic AI vs Enterprise Workflows: A Leadership Framing</h3><p><strong>As agentic AI gains traction, many enterprises risk repeating an old mistake: rebranding complexity instead of clarifying intent.</strong></p><blockquote>Leaders need to anchor the conversation around uncertainty, not novelty.</blockquote><h3>The Core Distinction</h3><p>At a systems level, there are three fundamentally different problem types:</p><blockquote><strong>Deterministic “how” problems → RPA / Deterministic Workflows</strong></blockquote><blockquote><strong>Reasoning “what” problems → Agentic AI</strong></blockquote><blockquote><strong>Scaled decision systems → Multi-Agent Architectures</strong></blockquote><p>The mistake is treating these as competitors. They are complements.</p><p>RPA: Determinism at Scale</p><p>RPA and classical workflows excel when:</p><p>- The process is known in advance</p><p>- Inputs and outputs are structured</p><p>- Compliance, auditability, and predictability matter</p><p>Example:</p><p>- Invoice processing</p><p>- System-to-system reconciliation</p><p>- User provisioning</p><blockquote>Trying to replace this with agents is not innovation — it’s technical debt disguised as intelligence.</blockquote><p>If the question is “How do I do this every time?”</p><p>The answer is almost always automation, not agents.</p><p>Agentic AI: Reasoning Under Uncertainty</p><p>Agentic AI becomes valuable when:</p><p>- The next step is not known in advance</p><p>- The system must interpret intent, ambiguity, or trade-offs</p><p>- Human-like judgment is required</p><p>Example:</p><p>- “What response best addresses this customer escalation?”</p><p>- “Which option balances risk, cost, and speed right now?”</p><p>Here, static workflows break down, and agents add real leverage.</p><blockquote><strong>Multi-Agent Systems: When One Mind Isn’t Enough</strong></blockquote><p>Multi-agent systems need to sit above both RPA/deterministic workflows and single-agent AI — not replace them.</p><p>They are justified only when:</p><p>1. Decision context becomes too large or noisy</p><p>2. Independent reasoning paths improve outcomes</p><p>3. Specialized domains or tools must be isolated</p><p>Example: Enterprise Change Impact Analysis</p><p>- One agent scans technical dependencies</p><p>- One agent assesses compliance impact</p><p>- One agent evaluates customer-facing risk</p><p>- An orchestrator synthesizes the decision</p><p>Each agent reasons clearly because it is constrained.</p><h3><strong>A Practical Enterprise Pattern (That Actually Scales)</strong></h3><blockquote><strong>The winning architecture is layered, not agent-first:</strong></blockquote><p>1. RPA / Deterministic Workflows</p><p>2. Single Agent Decision Layer</p><p>3. Multi-Agent Reasoning Layer (Selective)</p><p>This keeps cost predictable, systems debuggable, and compliance intact.</p><p>The Leadership Question That Prevents Over-Engineering</p><blockquote>Before approving any agentic or multi-agent design, ask:</blockquote><blockquote>Is the uncertainty about what decision to make — or how to execute it?</blockquote><blockquote>- What → Agentic AI</blockquote><blockquote>- How → Workflow / RPA</blockquote><blockquote>- Many whats at once → Multi-Agent</blockquote><p>This question alone eliminates most architectural confusion.</p><h3><strong>Leadership Takeaway: Intelligence Is Not Free</strong></h3><p>Agentic systems introduce:</p><p>- Higher compute costs</p><p>- More failure modes</p><p>- Harder observability</p><p>- New governance risks</p><blockquote>The goal is not more agents.</blockquote><blockquote>The goal is less uncertainty per decision— with the simplest architecture that can achieve it.</blockquote><p>The most mature enterprises will:</p><p>- Use deterministic workflows where certainty exists</p><p>- Use agents where judgment is required</p><p>- Use multi-agent systems only when complexity truly demands it</p><p>That restraint — not novelty — is what separates experimentation from production-grade AI systems.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=124c9ebc1b5f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Derivatives and Integrals in AI Explained by Usain Bolt!]]></title>
            <link>https://medium.com/@ujwaltickoo/derivatives-and-integrals-in-ai-explained-by-usain-bolt-a0226e296b48?source=rss-b4d2217d1709------2</link>
            <guid isPermaLink="false">https://medium.com/p/a0226e296b48</guid>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[calculus]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Ujwal Tickoo]]></dc:creator>
            <pubDate>Sun, 09 Nov 2025 12:39:20 GMT</pubDate>
            <atom:updated>2025-11-09T12:39:20.765Z</atom:updated>
            <content:encoded><![CDATA[<p>If you’re starting out in <strong>AI, machine learning, or deep learning</strong>, you’ll quickly bump into phrases like <a href="https://youtu.be/8hJw4wzEOMc?si=BrL8zs2C-QBvKDz-"><em>gradient descent</em></a>, <em>differentiation</em>, and <em>integration</em>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_fvYlwlMhC4cVRtTT1JNtg.jpeg" /><figcaption>Usain Bolt and Calculus!</figcaption></figure><p>And if you happened to gloss over math in high school, the ghosts of calculus might suddenly come back to haunt you! Here is your ghostbuster spray for those worries. A plain English refresher on basic calculus to understand <a href="https://medium.com/@ujwaltickoo/what-the-model-in-llm-really-means-explained-simply-7b9c34e78fb9">AI models</a>. But before we get to Usain, lets get some basics out of the way:</p><h3><strong>Differentiation and Integration: Two Sides of the Same Curve</strong></h3><p>We often hear that <strong>differentiation </strong>and <strong>integration </strong>are opposites or mirror each other. But what does that really mean? Let’s unpack this from first principles using simple ideas of change — both average and instantaneous.</p><h3>Average vs Instantaneous Rate of Change</h3><p>When a function changes, we can describe that change in two ways:</p><p>• <strong>Average Rate of Change (AROC) </strong>measures how much something changes on average over an interval. AROC = (f(b) — f(a)) / (b — a). Think of it as the overall slope between two points.</p><p>• <strong>Instantaneous Rate of Change (IROC) </strong>zooms in to measure how fast something is changing at a single instant. IROC = lim(<strong>Δx</strong>→0) <strong>Δ</strong>f/<strong>Δ</strong>x = f’(x). This is the slope of the tangent — the true speed of change at that exact point.</p><p>Differentiation is how we calculate that instantaneous rate of change everywhere along the curve.</p><h3>Integration: Adding Back All the IROCs</h3><p>If differentiation tells us how fast things change, then integration tells us how much total change happened when we add up all those tiny changes.</p><p>You can think of integration as:</p><p><em>Adding back all the IROCs over a period of change to find the total effect.</em></p><p>In calculus terms, if f’(x) gives the instantaneous rate of change, then <br>∫f’(x) dx = f(x) + C. This means integration is the reverse of differentiation — it accumulates all the little changes to recover the original function.</p><h3>AROC and IROC explained by Usain Bolt</h3><p>If you watch this <a href="https://youtube.com/shorts/gDQnKp1QTX8?si=hwPvpTDTYn-JoxUv">youtube video of Usain Bolt</a>, you will notice that although he won the 100m race, he wasn’t the fastest off the blocks. Someone else was. Although <a href="https://youtu.be/EKvHQc3QEow?si=vKgvMdoiNIwIsjPM">Bolt’s AROC</a> (Average change of distance over time i.e. speed) is about 9.48 m/s in a 100 m sprint, his IROC instantaneous rate of change varies! He gets faster later. This is why IROC is very important and critical to watch-out for an athelete. A weak IROC — early or late — can mean the difference between victory and defeat when everyone’s running 100 m within 10 seconds!</p><h3>When Change Stops Changing (The Second Derivative)</h3><p>Now, what happens when the rate of change itself stops increasing or decreasing? That’s when the <strong>second derivative</strong>, f’’(x), equals zero.</p><p>• f’’(x) measures how the IROC itself changes — like acceleration if f(x) is position.</p><p>• When f’’(x) = 0, the slope f’(x) stops changing — meaning you’ve reached a maximum or minimum on the original curve.</p><p>At that point, the hill flattens for a moment — the tangent slope (IROC) becomes zero — before the function either rises or falls again.</p><h3>Bringing It Together</h3><p>• Differentiation measures how fast change happens — the IROC.</p><p>• Integration adds back all those IROCs over time to recover the total change.</p><p>• When IROC stops changing, f’’(x) = 0, we find turning points — the peaks and valleys of the curve.</p><p>Together, these two ideas describe motion, growth, and change across nature and systems. Differentiation zooms into the moment. Integration zooms out to the whole journey.</p><h3>Visual Summary</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DmX9VCHjdzrmJEIGVcZRcw.png" /><figcaption>Derivatives and Integrals</figcaption></figure><p>The hill represents the function f(x). The tangent at any point shows f’(x) — the IROC. The shaded area beneath represents the integral, or total accumulated change. Where the tangent flattens, f’(x) = 0, the slope stops changing — a moment of balance.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a0226e296b48" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[“Attention is all you need” in plain English]]></title>
            <link>https://medium.com/@ujwaltickoo/attention-is-all-you-need-in-plain-english-b176d1ad4ada?source=rss-b4d2217d1709------2</link>
            <guid isPermaLink="false">https://medium.com/p/b176d1ad4ada</guid>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[artificial-neural-network]]></category>
            <dc:creator><![CDATA[Ujwal Tickoo]]></dc:creator>
            <pubDate>Sun, 02 Nov 2025 07:39:41 GMT</pubDate>
            <atom:updated>2025-11-02T07:39:41.528Z</atom:updated>
            <content:encoded><![CDATA[<p><em>How a simple mathematical idea transformed modern AI.</em></p><blockquote><a href="https://en.wikipedia.org/wiki/Attention_Is_All_You_Need">“Attention Is All You Need.”</a><em><br> The 2017 </em><a href="https://arxiv.org/abs/1706.03762"><em>Google research paper</em></a><em> that reshaped the entire field of Artificial Intelligence.<br> If you’ve worked around AI, ML, or Data Science, you’ve heard this phrase — but what does it really mean?</em></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*u2xMTBlYd-teSUz0VUlT8g.png" /></figure><h3>🧠 Pattern Intelligence — The Foundation of AI</h3><p>AI doesn’t think or feel like humans. AI is powered by <a href="https://medium.com/@ujwaltickoo/pattern-intelligence-is-ai-not-human-intelligence-5ef31e720291"><strong>Pattern Intelligence</strong> — the ability to find, weigh, and reproduce patterns across massive datasets.</a></p><p>This intelligence comes from <strong>mathematics</strong> (statistics, calculus, linear algebra) and <strong>computing power</strong> — all human creations. AI’s “intelligence” is built, not born. Understanding <strong>attention</strong> means understanding how AI uses mathematics to discover meaningful connections — like how words relate to each other in a sentence.</p><h3>📖 A Simple Analogy: Reading a Book</h3><p>Imagine you’re reading a long novel.<br> The plot stretches over hundreds of pages. What happened ten chapters ago matters at the very end.</p><p>If you forget those earlier parts, the story loses meaning.<br> Language works the same way — the meaning of a word depends on its context, sometimes far away in the sentence.</p><p>Before attention, older AI models read text one word at a time, forgetting earlier context as they moved forward.<br> They struggled to connect long-distance ideas — like trying to read one page a day for a year and remembering every detail. <em>Without remembering relationships across context, AI couldn’t capture the full story.</em></p><h3>💡 What “Attention” Does</h3><p>The <strong>Attention Mechanism</strong> changed that. It lets AI look at <strong>all words at once</strong> and decide which ones matter most to each other.</p><p>Mathematically, the model computes <strong>attention weights</strong> — numerical scores that indicate how strongly each word relates to another.</p><p>Example: <em>“The cat sat on the mat because it was tired.”</em></p><p>When the model processes “it,” attention helps it focus more on “cat” than on “mat,” since that connection makes sense.</p><p>Attention is essentially the model asking: <em>“Which other words should I care about while understanding this one?”</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*b0psG4Z7L5YYHTR-xkFJYA.jpeg" /><figcaption>Image created by David Joyner <a href="https://www.edx.org/bio/david-joyner">https://www.edx.org/bio/david-joyner</a></figcaption></figure><h3>🧩 Different Kinds of Attention</h3><p>There’s not just one type of attention. Transformers use several layers and flavors:</p><ul><li><strong>Self-Attention</strong><br>Each word attends to every other word in the same sequence to understand internal relationships. In “the dog chased the ball,” self-attention helps link “dog,” “chased,” and “ball.”</li><li><strong>Cross-Attention</strong><br> Used when two sequences interact — for example, aligning an English sentence with its French translation.</li><li><strong>Multi-Headed Attention</strong><br> The model splits attention into multiple “heads.”<br> Each head focuses on a different kind of relationship — grammar, emotion, object, sequence order — giving the model a multi-dimensional understanding of context.</li></ul><h3>🚀 Why “Attention Is All You Need”</h3><p><em>Earlier models relied on step-by-step memory.</em><br> The <strong>Transformer</strong> — introduced in the paper — replaced that with an <strong>attention-only</strong> system.</p><p>This breakthrough allowed models to:</p><ul><li>Process data in parallel (much faster)</li><li>Capture long-range context</li><li>Scale up efficiently to billions of parameters</li></ul><p>That’s why <strong>every modern AI model</strong> — GPT, Claude, Gemini, LLaMA — is built on the Transformer architecture.</p><h3>🧩 The lens of Math</h3><p>Attention made machines better at connecting patterns — not understanding them. It’s a mathematical lens, not a mind. While analogies to how humans pay attention across sentences and paragraphs are useful, what truly drives AI attention is mathematics under the hood.</p><h3>⚙️ In Summary</h3><ul><li>AI is <strong>Pattern Intelligence</strong>, not Human Intelligence.</li><li><strong>Attention</strong> lets AI find relationships and meaning structures within patterns.</li><li><strong>Transformers</strong> don’t <em>understand</em> context; they <em>model</em> or <em>capture dependencies</em> across long sequences <strong>using the attention mechanism</strong>. Their real innovation is the <strong>scalability</strong> of this mechanism — allowing parallel processing of input (tokens) and capturing relationships across large contexts efficiently.</li></ul><p>That’s the essence behind the phrase: “Attention Is All You Need.”</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b176d1ad4ada" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Pattern Intelligence is AI, not Human Intelligence]]></title>
            <link>https://medium.com/@ujwaltickoo/pattern-intelligence-is-ai-not-human-intelligence-5ef31e720291?source=rss-b4d2217d1709------2</link>
            <guid isPermaLink="false">https://medium.com/p/5ef31e720291</guid>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[mls]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[ai-product-management]]></category>
            <dc:creator><![CDATA[Ujwal Tickoo]]></dc:creator>
            <pubDate>Sun, 02 Nov 2025 06:26:00 GMT</pubDate>
            <atom:updated>2025-11-02T06:26:00.778Z</atom:updated>
            <content:encoded><![CDATA[<p>The biggest misunderstanding about Artificial Intelligence lies in the word <em>“intelligence”</em> itself.<br> It makes us assume we’re building something like the human mind — a machine that sees, thinks, and feels like we do. But that’s not what’s happening.</p><p>AI doesn’t have human intelligence. It has <em>pattern intelligence.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IQ4Qp5CNGQ3j_FipdAYfyw.png" /><figcaption>AI is Pattern Intelligence, Not Human Intelligence</figcaption></figure><p>The source of this pattern intelligence lies in <strong>mathematics</strong> — statistics, calculus, matrices, and vectors — and in the <strong>hardware</strong> that computes them at scale. All of these are human inventions. AI’s “intelligence” is built upon the foundations of human mathematics and computing power — not consciousness.</p><p>The label “intelligence” itself muddles our understanding of AI. It blurs the clarity we need to see what we’re actually dealing with and what, at large, society needs to learn thoughtfully.</p><h3>The Illusion of Human-Like Labels</h3><p>We call it <em>Computer Vision</em>, but there’s no vision.<br> We say <em>Language Understanding</em>, but there’s no comprehension.</p><p>These terms borrow from human experience — and that’s where confusion begins.</p><p>AI doesn’t see or understand like humans. It processes data — text, images, audio — and identifies patterns that correlate with what humans label as “seeing,” “speaking,” or “writing.”<br> Its intelligence lies in the <em>structure of patterns</em>, not in awareness or meaning.</p><h3>So What Kind of Intelligence Is AI?</h3><p><strong>It Is Pattern Intelligence.</strong></p><p>Every AI system — from ChatGPT to Gemini to self-driving cars — learns by detecting and reproducing <strong>patterns</strong> in massive datasets.</p><p>When trained on books, videos, and conversations, AI learns how words and concepts tend to appear together. It doesn’t <em>understand</em> a story about loss; it recognizes a pattern connecting <em>mother</em>, <em>child</em>, <em>death</em>, and <em>pain.</em></p><p>When it generates a new story, it doesn’t <em>feel</em> grief — it reproduces the underlying pattern of <em>pain through separation.</em></p><p>That’s what makes AI powerful — and also fundamentally <em>not human.</em></p><h3>Patterns? What Patterns?</h3><p>Data from the world — books, videos, audio, images — are fed into machine learning systems. AI algorithms use powerful hardware to process this data and detect patterns.</p><p>Once patterns are found and encoded, machines can <em>behave</em> like humans — they can exhibit behaviors that resemble understanding. But AI doesn’t read a book, grasp meaning, or feel empathy.</p><p>Its ability to write stories or respond intelligently doesn’t make it human.<br>It simply reveals the depth of its <strong>pattern recognition</strong>.</p><p>That is <em>Pattern Intelligence. </em>It is not human at all. There is no feeling, no consciousness — only pattern understanding.</p><h3>The Risk of Seemingly Conscious AI (SCAI)</h3><p><a href="https://mustafa-suleyman.ai/seemingly-conscious-ai-is-coming">Mustafa Suleyman calls this phenomenon <em>Seemingly Conscious AI (SCAI)</em></a> to highlight that we are dealing with a very different kind of intelligence — one that appears to think and feel but doesn’t.</p><p>His warning is simple: if we start treating <em>pattern intelligence</em> as <em>person intelligence</em>, we risk blurring moral and regulatory boundaries. The societal and global risks of such confusion are immense.</p><p><em>“My central worry is that many people will start to believe in the illusion of AIs as conscious entities so strongly that they’ll soon advocate for AI rights, model welfare and even AI citizenship. This development will be a dangerous turn in AI progress and deserves our immediate attention.</em></p><p><em>We must build AI for people; not to be a digital person. AI companions are a completely new category, and we urgently need to start talking about the guardrails we put in place to protect people and ensure this amazing technology can do its job of delivering immense value to the world. I’m fixated on building the most useful and supportive AI companion imaginable. But to succeed, I also need to talk about what we, and others, shouldn’t build.</em></p><p><em>To some this discussion will feel ungrounded, more science fiction than reality. To others it may feel unnecessarily alarmist. Such emotional reactions are the tip of the iceberg given what lies ahead. It’s highly likely that some people will argue that these AIs are not only conscious, but that as a result they may suffer and therefore deserve our moral consideration.</em></p><p><em>To be clear, there is zero evidence of this today, and some argue there are strong reasons to believe it will not be the case in the future. Yet the consequences of many people starting to believe an SCAI is actually conscious deserve our immediate attention. We have to be extremely cautious here and encourage real public debate and begin to set clear norms and standards. This is about how we build the right kind of AI — not AI consciousness. Clearly establishing this difference isn’t an argument about semantics, it’s about safety. Personality without personhood. And this work must start now.”</em></p><p>Recognizing this difference isn’t just semantics; it’s about <strong>safety</strong>.<br> The illusion of consciousness can lead to misplaced trust, emotional dependence, and flawed governance.</p><h3>Why Clarity Matters</h3><p><a href="https://en.wikipedia.org/wiki/Turing_test">The Turing Test </a>measured whether a machine could <em>behave</em> like a human, not whether it was <em>conscious. </em>We are testing behavior — not awareness.</p><p>Today’s models pass many behavioral tests, but they remain deeply mechanical: learning and predicting patterns, not experiencing meaning.</p><p>As product builders and policymakers, understanding this distinction helps us design responsibly:</p><ul><li>Focus on <strong>reliability over relatability</strong></li><li>Measure <strong>pattern accuracy</strong>, not moral intention</li><li>Build <strong>guardrails</strong> that protect people from mistaking patterns for empathy</li></ul><h3>Clarity About Pattern Intelligence</h3><p>The future of AI depends on how we define and understand it.</p><p>If we see it as <em>pattern intelligence</em> — a tool that mimics human expression without possessing human experience — we can harness its power without losing clarity about what it truly is.</p><p>AI doesn’t need to be human to be transformative.<br> It just needs to be understood.</p><blockquote><em>Clarity about pattern intelligence is not philosophical — it’s practical.<br> It’s how we build, govern, and trust the next generation of AI responsibly.</em></blockquote><p>About the author: Ujwal Tickoo is a product leader exploring the intersection of automation, AI, philosophy, and human decision-making.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5ef31e720291" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Zen of Inner Peace]]></title>
            <link>https://medium.com/@ujwaltickoo/the-zen-of-inner-peace-5037466be841?source=rss-b4d2217d1709------2</link>
            <guid isPermaLink="false">https://medium.com/p/5037466be841</guid>
            <dc:creator><![CDATA[Ujwal Tickoo]]></dc:creator>
            <pubDate>Sun, 17 Aug 2025 00:54:26 GMT</pubDate>
            <atom:updated>2025-08-17T02:48:22.866Z</atom:updated>
            <content:encoded><![CDATA[<p>Zen and Inner Peace</p><p><em>Kung Fu Panda</em> is a delightful movie. One of its funniest paradoxes is the scene where a character seeks “inner peace” by closing his eyes, holding a pose, and chanting the words over and over. It’s funny — but also revealing.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/500/0*t8DICw-dyrkKbJPc" /></figure><p>Because, real inner peace can’t be summoned by a command, a pose, or a mantra. It can’t be found in food, fragrance, or fleeting pleasures. Yes, the scent of a rose can bring a moment of calm — but it fades as soon as you leave the garden. That’s not <em>inner</em> peace. That’s outer.</p><p>True inner peace shows up in unexpected places. Take parenting, for example. Even in the messy struggle of changing diapers filled with smelly poop, a parent can feel a deep calm rooted in love. That peace flows from within, not from the circumstances.</p><p>So what is this <em>within</em>?<br> It’s not the senses, and it’s not the mind that processes what the senses perceive. Touch, taste, smell, sight — all these give only temporary relief or pleasure. They cannot provide lasting stillness. For peace that endures, we have to look deeper — into a part of us untouched by external conditions.</p><p>There has to be something inside us that holds a steady source of peace, no matter where we are: at the office, in a crowded mall, or in the quiet of a garden.</p><p>Yoga teacher <a href="https://www.instagram.com/joshkrameryoga/?hl=en">Josh Kramer</a> puts it well:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/836/1*rZ2Htl3PGZ2Wh6C40Mu9UA.jpeg" /></figure><blockquote><strong>“</strong>Here’s a thought…If you search for peace in things that are temporary in nature, your peace will be temporary.</blockquote><blockquote>True peace has to be nurtured…it has to come from within.</blockquote><blockquote>When you recognise this, you will be able to find peace in any situation.”</blockquote><p>I can’t hand you your answer. And that’s the point. The source of inner peace has to be discovered by you, within you.</p><p>My invitation is simple: instead of looking outside for the next fragrance, distraction, or comfort, turn inward. Explore the quiet space inside yourself. Does it even exist? Where? The fountain of inner peace can only be sought by you, within you.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5037466be841" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[UI v/s API v/s UAI v/s Human-LLM interface]]></title>
            <link>https://medium.com/@ujwaltickoo/ui-v-s-api-v-s-uai-v-s-human-llm-interface-a07a64461cd9?source=rss-b4d2217d1709------2</link>
            <guid isPermaLink="false">https://medium.com/p/a07a64461cd9</guid>
            <category><![CDATA[genai]]></category>
            <category><![CDATA[ux]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[Ujwal Tickoo]]></dc:creator>
            <pubDate>Thu, 14 Aug 2025 05:12:29 GMT</pubDate>
            <atom:updated>2025-08-14T05:12:29.752Z</atom:updated>
            <content:encoded><![CDATA[<p>Here are my thoughts about beyond UI, API, and UAI — towards Direct Human–Model Interaction.</p><blockquote>UI v/s API v/s UAI<br>To start with <a href="https://www.joshbeckman.org/blog/practicing/ui-vs-api-vs-uai">Josh Beckman</a> has written a lovely post that explains these ideas well :<br>“First we built a <strong>user interface (UI)</strong> when the application was just going to be operated by humans. There’s a whole branch of study around good and bad patterns in that design practice, focusing on how we make things easily operable by humans.</blockquote><blockquote>Then we added an <strong>application programmable interface (API)</strong> when we wanted the application to be operated by other applications — integrated programs. Similarly, there are whole conferences and books about what good design looks like for these interfaces, aimed at how we make programs easily operable by other applications (and developers of those applications).</blockquote><blockquote>These past few months I think we’re seeing <strong>user agent interface (UAI)</strong> design emerge from all the discussions and usage of agents (LLM and otherwise) to perform actions on behalf of a human.”</blockquote><p>The Human-LLM Influence Interface (HLII)</p><p>With LLMs and model-centric computing, what I sense missing is not simply human access to functionality but actual human access to influence Model training and weights and drive experiments there. Beyond a RAG type augmentation — a new kind of Human-Machine input level interface.</p><p>When we talk about UI, API, and UAI, we’re usually talking about <em>access</em> — how humans, applications, or agents can tap into software or LLM capabilities.</p><p>But what feels missing is a way for humans to go beyond <strong>accessing</strong> a model to actually <strong>shaping</strong> it. I’m not talking about RAG or prompt engineering. I mean the ability for a non-developer to influence an LLM’s <em>training process</em> — tuning its weights, running experiments, reshaping its latent space — through a human-machine interface designed for that purpose.</p><p>Think of it as <em>teaching the model directly</em>, not just evaluating or wrapping it in APIs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/866/1*YmUzfiZnElPqXkeoZPT0Aw.png" /><figcaption>Human-LLM influence interface</figcaption></figure><p>So far, I am reading online about emerging prototypes of this concept — intriguing but nowhere near mainstream. If this became a mature capability, it could redefine what “human–AI collaboration” means.</p><p>Have you come across research or tools that make this possible? I’d love to see examples or your reflection if Human-LLM influence interface is going to be valuable or is it naive to believe this is possible/valuable?</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a07a64461cd9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What the “Model” in LLM Really Means — Explained Simply]]></title>
            <link>https://medium.com/@ujwaltickoo/what-the-model-in-llm-really-means-explained-simply-7b9c34e78fb9?source=rss-b4d2217d1709------2</link>
            <guid isPermaLink="false">https://medium.com/p/7b9c34e78fb9</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[neural-networks]]></category>
            <category><![CDATA[math]]></category>
            <dc:creator><![CDATA[Ujwal Tickoo]]></dc:creator>
            <pubDate>Sun, 22 Jun 2025 08:30:59 GMT</pubDate>
            <atom:updated>2025-06-22T08:30:59.978Z</atom:updated>
            <content:encoded><![CDATA[<h3>What the “Model” in LLM Really Means — Explained Simply</h3><p>It’s Called a Model for a Reason, let’s try understanding the basics of LLM Magic. Very specifically, understanding the “Model” in LLMs Matters for AI Product Builders.</p><p>LLMs are not powerful simply because they’ve read vast amounts of text. What makes them “models” is the <strong><em>training</em></strong> process that uses this text to learn <strong><em>probabilistic patterns of language. </em></strong>This pattern could be code in Python, or poem in English or an email at work. .</p><p>There are lots of LLM primers regurgitating terms like “next word prediction”, “text ingestion”, “model weights”, along with “neural networks”. That is a lot of geek to digest for a non-tech professional! Let’s take a simpler approach to understanding how LLMs work.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*r1WfjLh0JSketdW5DY-Xtg.jpeg" /><figcaption><strong><em>“Understanding the Model in Training vs. Usage”</em></strong><em><br> During training, the LLM learns patterns in language by adjusting its internal parameters. D</em>uring usage (also called inference), it applies what it has learned to generate responses<em>.</em></figcaption></figure><h3>Why LLMs Are Called Models (And Not Just Giant Text Readers)</h3><p><strong><em>An LLM is not called a ‘model’ simply because it has ingested large volumes of text.</em></strong></p><blockquote><strong><em>Failing to internalize this, can lead to a lot of confusion about why LLMs are being used in AI application development and how they work or how to write good prompts for ChatGPT.</em></strong></blockquote><p><strong>So, about the ‘model’ part — and if you hated math in school, sorry! This is all about math and stats :)</strong></p><p>At a high level the “Model” part refers to the LLM’s ability to use math (statistics) to predict the next word (token is the technically accurate term). But how? The Model has learned from <strong><em>Patterns</em></strong> in text — inherent patterns of reasoning, logic, and thinking implicit in written text. <em>Pattern recognition — especially the ability to predict the next token </em>— is key to how the model works.</p><h3>Training, Patterns, and Prediction: Demystifying the LLM “Model”</h3><p>A <em>pattern in this case refers to statistical function </em>— for example, the likelihood that the word ‘forecast’ follows ‘sales’ in business-related text.</p><p>It becomes a model <em>because</em> it has learned parameters (weights) during training, just like a machine learning model that predicts prices, customer churn, or product recommendations. (A review of regression in high-school math/statistics course will conceptually clarify ideas like weights and functions very well.)</p><blockquote><strong><em>Think of it like this:</em></strong><em> You trained a sales forecast model using historical sales data. The LLM is trained using huge amounts of text data instead.</em></blockquote><p>Let’s break this down further:</p><h3>Why is an LLM a “model” — Understanding Training</h3><p>Technically speaking, the model bit, refers to <strong>a trained mathematical function </strong>— typically a neural network — that, once trained on massive datasets, can generate likely sequences of text based on prior input.</p><h3>Does it digest a lot of text?</h3><p>Yes. Think massive volumes of text like wikipedia, books, blogs, articles, code-repositories and more. It is trained on massive amounts of textual data — <strong>learning patterns in the process.</strong></p><h3>Patterns versus Understanding</h3><h4>Here is where it is important to pause and reflect.</h4><p><strong>Again, the key isn’t just reading data — it’s the learning of statistical patterns during training that makes it a model. Learning patterns is the core functionality — learning statistical patterns across language. Next word prediction is the output.</strong></p><p><em>We need to internalize the following:</em></p><p>· Distinguish <strong>data ingestion (during training) </strong>from “<strong>learning”</strong></p><p>· Internalize <strong>statistical pattern recognition</strong></p><p><strong>The statistical pattern recognition </strong>grounds the math function of the model in <strong>next-token prediction</strong>, which is the root mechanism.</p><h3>Does the LLM truly ‘reason” or ‘think’ like humans?</h3><p>No. It mimics reasoning patterns but doesn’t understand or think like humans do. It mimics reasoning based on patterns — not human comprehension of input or text it was trained on.</p><h3>Emergent Abilities: Where the Magic Feels Real</h3><p>· During both training and inference, LLMs <strong>predict the next token</strong> given previous ones. That’s the fundamental task.</p><p>· All emergent behaviors — summarization, translation, question answering, reasoning — are <strong>side-effects</strong> of this ability to predict text that “makes sense” in context.</p><p>Just like a child exposed to lots of books begins to write coherent stories, the LLM, trained on vast text data, starts demonstrating capabilities that weren’t explicitly programmed — like answering questions or summarizing articles.</p><p><strong>Learning patterns is the core functionality — learning statistical patterns across language.</strong></p><p>· The entire power of an LLM is in its ability to detect and generalize <strong>patterns in language</strong> — including grammar, tone, reasoning structures, and even style.</p><p>· These patterns are <strong>statistical</strong>, meaning the model has no understanding or grounding in meaning — it’s simply picking the most probable word(s) based on prior context.</p><p>Back to the Model-bit. It is <strong>not just the digestion of large volumes of text, but really the pattern learning during training that creates the model.</strong></p><p>· Merely ingesting text during training doesn’t do it. The LLM becomes a model <strong>through training a complex mathematical function</strong>, where it learns <strong>statistical relationships</strong> between tokens (words/pieces of words).</p><p>· <strong>Pattern learning</strong> (by adjusting weights in the neural network based on prediction errors) is the process by which the model’s intelligence emerges.</p><h3>Summary: Why It’s Called a Model</h3><p>An LLM becomes a “model” through training — learning statistical patterns across language by adjusting weights based on prediction errors. This enables it to generate contextually relevant, human-like text responses. But it doesn’t think or understand like humans; it simply models the probability of what comes next.</p><p>If you want a clear understanding on Agents and Agentic AI check this post <a href="https://medium.com/@ujwaltickoo/understanding-agentic-reasoning-and-thinking-for-product-and-bmanagers-c94c11f77b60">https://medium.com/@ujwaltickoo/understanding-agentic-reasoning-and-thinking-for-product-and-bmanagers-c94c11f77b60</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7b9c34e78fb9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Understanding Agentic Reasoning and Thinking for Product and BManagers]]></title>
            <link>https://medium.com/@ujwaltickoo/understanding-agentic-reasoning-and-thinking-for-product-and-bmanagers-c94c11f77b60?source=rss-b4d2217d1709------2</link>
            <guid isPermaLink="false">https://medium.com/p/c94c11f77b60</guid>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[business-strategy]]></category>
            <category><![CDATA[agents]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[business-operations]]></category>
            <dc:creator><![CDATA[Ujwal Tickoo]]></dc:creator>
            <pubDate>Mon, 09 Jun 2025 06:05:36 GMT</pubDate>
            <atom:updated>2025-06-15T06:15:21.532Z</atom:updated>
            <content:encoded><![CDATA[<h3>Understanding Agentic Reasoning and Thinking for Product and Business Managers</h3><p><strong>How do we make sense of ideas like thinking and reasoning in Agentic AI when there is no consciousness involved? </strong>Take, for example, the following words from an article from McKinsey on Agentic AI:</p><p>“Agentic systems don’t just execute tasks — they “think.” They reason across time horizons, learn from outcomes, and collaborate with other AI agents in areas such as fraud detection, compliance, and capital allocation to continuously optimize performance” — <a href="https://www.mckinsey.com/capabilities/operations/our-insights/when-can-ai-make-good-decisions-the-rise-of-ai-corporate-citizens">McKinsey article on Agentic AI in Business Operations</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/924/1*VgH_EFid2jYLU5VkKfyVvg.jpeg" /></figure><p>A simple way to interpret “reasoning” of agents is that they ingest data (over time) and improve decisions based on updated patterns in data by updating the underlying math of their model characteristics. Thus <em>reasoning means discerning patterns, making changes to existing patterns of thought to achieve specific goals.</em></p><p>Thinking here implies the ability to start with ingesting data — then compare old and new patterns to drive new actions around specific goals including calling other agents with relevant skills that an agent may not have.</p><p>Let’s walk through the original statement from the McKinsey article — breaking things down to the most fundamental ideas.</p><p><strong>Step-by-Step Breakdown</strong></p><p><strong>1. “Agentic systems don’t just execute tasks — they <em>think</em>.”</strong></p><p><strong>Classically coded systems (e.g. automations) </strong>follow rules (e.g., RPA or basic AI). They don’t <em>adapt</em> unless reprogrammed. But, with <strong>new data ingestion</strong>, agentic systems are able to <em>adapt without being re-programmed as they aren’t based on fixed deterministic rules and coding.</em></p><p><strong>Agentic systems</strong>, thus, have the ability to <em>sense, evaluate, and act with autonomy</em>.</p><p><strong>Breakdown</strong>: <strong><em>Thinking</em> (for an AI system) doesn’t mean consciousness</strong>. It means:</p><ul><li>Comparing current data to past data (pattern matching)</li><li>Evaluating multiple options or hypotheses (simulation or planning)</li><li>Choosing based on an internal goal (like cost-saving, fraud avoidance, etc.)</li></ul><p><strong>In Summary</strong>: “Ingesting data and comparing patterns” is a simplification of how“thinking” happens in machines.</p><p><strong>2. “They reason across time horizons.”</strong></p><p><strong>Reasoning across time</strong> means:</p><ul><li>Making decisions not just for <em>now</em>, but predicting or simulating <em>future outcomes</em>.</li><li>Adapting based on both <strong>short-term feedback</strong> and <strong>long-term strategy</strong>.</li></ul><p><strong>Breakdown</strong>: To reason across time, a system must:</p><ul><li>Model cause-effect relationships (e.g., if I delay a payment today, it affects credit score tomorrow).</li><li>Track historical data and project future states.</li></ul><p><strong>In simple terms</strong>: The agent doesn’t just say “What’s best now?” but “What will be best over time?”</p><p><strong>3. “Learn from outcomes.”</strong></p><p>This is classic <strong>reinforcement</strong> or <strong>feedback learning</strong>:</p><ul><li>The agent tries something.</li><li>It observes the result.</li><li>It updates its “belief” or model to do better next time. The “belief” part refers to the math and workings of the model that drives an agent. The agent updates outcomes of the model based on patterns observed from newly ingested data.</li></ul><p><strong>Breakdown</strong>: Thus <strong>Learning means, adjusting internal models based on feedback.</strong></p><p><strong>4. “Collaborate with other AI agents.”</strong></p><p>This is the <strong>multi-agent</strong> concept:</p><ul><li>One agent might be good at fraud detection.</li><li>Another might specialize in compliance.</li><li>They <em>interact</em> or pass information to each other.</li></ul><p><strong>Breakdown</strong>: Specialized agents can communicate and divide responsibility, just like human teams.</p><p><strong>5. “Optimizing performance.”</strong></p><p>This is the <strong>goal-directed nature</strong> of agents.</p><ul><li>They aren’t just passive tools.</li><li>They continuously measure how well they’re doing against a target (e.g., minimizing fraud, maximizing returns).</li><li>They self-correct and self-improve.</li></ul><p><strong>Breakdown</strong>: Optimization = Adjusting behavior over time to get better results using feedback and planning.</p><p>Goal directed behavior is in-itself complex and multi-layered. A good example to understand is in the well-documented example of Google’s AlphaGo. It used <a href="https://medium.com/@kinwo/a-brief-introduction-of-alphago-and-deep-learning-how-it-works-76e23f82fe99">three components — Policy network, Value Network, and Tree Search</a> to get to the <em>goal of </em><a href="https://www.youtube.com/watch?v=WXuK6gekU1Y&amp;t=5248s"><em>game-winning</em> against world’s best Go Player Lee Sedol.</a></p><p><strong>Summary:</strong></p><p>Agentic systems are not just <strong>rule followers</strong>. They are:</p><ul><li><strong>Perceivers</strong>: They sense the world (via data).</li><li><strong>Evaluators</strong>: They compare current and past situations.</li><li><strong>Forecasters</strong>: They simulate or predict outcomes across time.</li><li><strong>Learners</strong>: They update behavior based on outcomes.</li><li><strong>Collaborators</strong>: They call on specialized peers to act together.</li><li><strong>Optimizers</strong>: They tune their behavior to meet a goal.</li></ul><p>Note: Human senses — eyes, skin and ears get — visual, temperature/air, and audio data from the world. Perception in agents is based on data gathering and ingestion without the incumbent consciousness of humans.</p><p>This is a simple way to understand, how AI agents think and reason. It is based on data ingestion, pattern comparison, model updates, goal maximization, feedback loops, and more. All this, without human consciousness.</p><p>If you are new to Agentic AI here is my post that will help you get started in less than 5 minutes! <a href="https://medium.com/@ujwaltickoo/a-business-centric-understanding-of-agentic-ai-b474719fed81">https://medium.com/@ujwaltickoo/a-business-centric-understanding-of-agentic-ai-b474719fed81</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c94c11f77b60" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A business-centric understanding of Agentic AI]]></title>
            <link>https://medium.com/@ujwaltickoo/a-business-centric-understanding-of-agentic-ai-b474719fed81?source=rss-b4d2217d1709------2</link>
            <guid isPermaLink="false">https://medium.com/p/b474719fed81</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[agents]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Ujwal Tickoo]]></dc:creator>
            <pubDate>Wed, 16 Apr 2025 05:23:33 GMT</pubDate>
            <atom:updated>2025-04-16T05:26:01.931Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*S0IE5MYi402F4ppjEyFbVg.png" /></figure><h3>Understanding Agentic AI Without the Tech Jargon</h3><p>Is it possible to understand Agentic AI without diving into technical jargon or the latest buzzwords — and still make it relevant for business? Absolutely. If you’ve ever used ChatGPT, you’re already halfway there. In just a few minutes of reading, you’ll have a practical grasp of what Agentic AI is and why it matters.</p><h3>What Is Agentic AI?</h3><p>Agentic Artificial Intelligence isn’t brand-new. It’s an evolution — building on technologies like Large Language Models (LLMs), such as those used in ChatGPT. The key difference is that Agentic AI introduces <strong>agency</strong>: the capacity to reason and act independently in pursuit of defined goals. In essence, it transforms an AI model from a passive responder into an <strong>Agent</strong> — one that can take action on behalf of a human or a system.</p><h3>Why Do We Need Agentic AI?</h3><p>Weren’t GenAI and LLMs powerful enough already? It’s a fair question.</p><p>The answer is that Agentic AI isn’t a completely new technology — it’s a <strong>natural progression</strong>. While LLMs like those behind ChatGPT are powerful, they are limited in two important ways when it comes to real-world business use:</p><ol><li><strong>They are passive</strong> — LLMs wait for a prompt and respond. For example, the flow in ChatGPT is:Prompt → LLM → Output</li><li><strong>They lack access to proprietary data</strong> — LLMs do not inherently know what’s inside your organization’s private systems. They cannot access personal calendars, internal documents, or secure databases. Jeff Su, in one of his YouTube videos, offers a great example of this limitation: an LLM that can’t access a private calendar simply can’t help with scheduling or availability-based decisions.</li></ol><h3>Hallucination and Its Challenges</h3><p>Because LLMs operate with limited context and data, they can sometimes “hallucinate” — generate information that sounds plausible but is factually incorrect. This is a serious limitation for business-critical decisions.</p><p>To give some scale: Google’s search index is believed to cover only about 5–10% of the total internet. OpenAI has shared that ChatGPT is trained using:</p><ul><li>(1) Publicly available internet data</li><li>(2) Data licensed or provided by third parties</li><li>(3) Information generated by users or human trainers</li></ul><p>This means that <strong>LLMs alone are not enough</strong> for reliable, data-rich business operations.</p><h3>How Agentic AI Solves Real Business Problems</h3><p>To address limitations like passivity and restricted data access, a new generation of applications has emerged. These connect LLMs with business systems, tools, and real-time data sources.</p><p>Such systems are often called <strong>AI workflows</strong>. The most advanced forms — where AI systems operate with independence — are known as <strong>autonomous Agentic AI applications</strong>.</p><p>In simple terms:<br> <strong>Agentic AI = LLM + Tools + Data + Goals</strong></p><p>These systems are capable of acting on your behalf, making decisions, and continuously learning.</p><h3>Going Beyond One LLM</h3><p>Agentic AI doesn’t have to rely on just one model. Multiple LLMs can be linked together, forming a chain that mimics collaborative human reasoning. For example, one model might draft a response, another checks its accuracy, and a third ensures compliance.</p><p>This setup enables:</p><ul><li>Maker-checker frameworks</li><li>Complex reasoning chains</li><li>More trustworthy and robust AI behavior</li></ul><p>[<a href="https://youtu.be/O0GNrvO7wD0?si=0uYTFjLeGFnj1ngT">Don Woodlock explains this idea well in his YouTube video about chaining LLMs for reasoning</a>.]</p><h3>Agents vs. Workflows</h3><p>Not all business needs require full-fledged agents. Sometimes, <strong>AI-powered workflows</strong> — more structured and predictable — are sufficient. <a href="https://youtu.be/Qd6anWv0mv0?si=KLTBcwyQQ_uMQ2zS">Google Tech has a short video explaining the difference between workflows and agents</a>, which is worth watching if you’re exploring practical applications.</p><h3>Human and Ethical Considerations</h3><p>As exciting as Agentic AI, GenAI, and LLMs are, it’s essential to avoid becoming overly focused on the tech. AI solutions must be <strong>contextualized with ethical, human, and organizational considerations</strong> in mind.</p><p>An excellent article by <a href="https://sloanreview.mit.edu/article/why-ai-demands-a-new-breed-of-leaders/">Faisal Hoque, Thomas H. Davenport, and Erik Nelson in <em>MIT Sloan Management Review</em> </a>provides examples of what can go wrong:</p><blockquote><em>“When leaders fail to think through the strategic and organizational consequences of their AI plans, the results can be catastrophic. Zillow’s failed attempt to use AI-generated property valuations led to $300 million in losses and a 20% drop in stock value.”</em></blockquote><p>They also note the importance of considering user impact:</p><blockquote><em>“When Air Canada used a generative AI chatbot to help travelers book flights, the bot made mistakes about bereavement fares — causing undue stress for grieving passengers. This shows the importance of human oversight and thoughtful deployment.”</em></blockquote><h3>Closing Thoughts</h3><p>Agentic AI represents a powerful shift — from passive interaction to active intelligence. When done right, it can:</p><ul><li>Improve efficiency and enable new forms of automation</li><li>Connect systems</li><li>Support decision-making</li></ul><p>But like any powerful tool, it must be wielded thoughtfully — with an eye on the <strong>business need</strong>, the <strong>human experience</strong>, and the <strong>ethical boundaries</strong>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b474719fed81" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>