<?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 Yan Cotta on Medium]]></title>
        <description><![CDATA[Stories by Yan Cotta on Medium]]></description>
        <link>https://medium.com/@yanpcotta?source=rss-0115aca74f91------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*0wjID6m9ANv18Mwa</url>
            <title>Stories by Yan Cotta on Medium</title>
            <link>https://medium.com/@yanpcotta?source=rss-0115aca74f91------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 07:05:45 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@yanpcotta/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[Unlocking AI’s Full Potential: Mastering MCP and A2A for Next-Level Applications]]></title>
            <link>https://medium.com/@yanpcotta/unlocking-ais-full-potential-mastering-mcp-and-a2a-for-next-level-applications-764807f20191?source=rss-0115aca74f91------2</link>
            <guid isPermaLink="false">https://medium.com/p/764807f20191</guid>
            <category><![CDATA[mcp-server]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[agents]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Yan Cotta]]></dc:creator>
            <pubDate>Tue, 29 Apr 2025 16:44:12 GMT</pubDate>
            <atom:updated>2025-04-29T16:44:12.548Z</atom:updated>
            <content:encoded><![CDATA[<h3>Introduction</h3><p>The rapid advancement of Large Language Models (LLMs) like GPT-4, Claude, Gemini, and Llama3 has transformed artificial intelligence. Yet, their inability to natively connect with real-time data, external APIs, databases, or local files often restricts their real-world utility. Enter the <strong>Model Context Protocol (MCP)</strong> and <strong>Agent-2-Agent (A2A)</strong> protocol — two groundbreaking standards that bridge these gaps, unlocking new possibilities for AI applications. In this article, I’ll guide you through MCP’s role in integrating LLMs with external systems, showcase a practical flight search application built with Gemini and MCP, and explore how A2A enables seamless agent collaboration.</p><h3>What is the Model Context Protocol (MCP)?</h3><p>Developed by Anthropic, MCP is an open-source protocol that standardizes how AI models interact with external tools and data sources. Imagine it as a universal adapter — like a “USB-C for AI” — that simplifies and secures connections between LLMs and the outside world.</p><h3>How MCP Works</h3><p>MCP operates on a client-server model:</p><ul><li><strong>Clients</strong>: AI applications or LLMs that request access to tools or data.</li><li><strong>Servers</strong>: MCP-compliant providers that offer tools, APIs, or data sources.</li></ul><p>This architecture ensures structured, dynamic, and secure interactions, making it easier for LLMs to leverage external resources.</p><h3>Benefits of MCP</h3><ul><li><strong>Unified Integration</strong>: Streamlines connections to diverse systems without custom coding.</li><li><strong>Adaptability</strong>: Allows LLMs to tap into multiple tools as needed.</li><li><strong>Security</strong>: Facilitates safe API calls without embedding sensitive credentials.</li><li><strong>Ease of Development</strong>: Simplifies the creation and deployment of custom MCP servers.</li><li><strong>Low Maintenance</strong>: Eliminates redundant integration efforts.</li></ul><h3>MCP Server Examples</h3><ul><li><strong>File Access</strong>: Interact with local files and folders.</li><li><strong>Web Queries</strong>: Conduct live searches online.</li><li><strong>Database Queries</strong>: Retrieve data from SQL or NoSQL systems.</li><li><strong>CRM Integration</strong>: Link to platforms like Salesforce.</li><li><strong>Version Control</strong>: Connect to tools like Git.</li></ul><h3>When to Use MCP</h3><p>MCP shines in scenarios like:</p><ul><li>Developing agent-driven systems.</li><li>Creating modular, reusable tools.</li><li>Integrating multiple external data sources.</li><li>Scaling applications with extensive toolchains.</li></ul><h3>Building a Flight Search Application with Gemini and MCP</h3><p>To demonstrate MCP in action, let’s walk through the architecture and implementation of a flight search application powered by Gemini and MCP.</p><h3>Architecture Overview</h3><p>This system integrates:</p><ul><li>A user-facing client script for input.</li><li>An MCP server handling flight data queries.</li><li>The Gemini API for natural language understanding.</li><li>SerpAPI for real-time flight information.</li></ul><h3>Step-by-Step Implementation</h3><h4>Step 1: Set Up the Environment</h4><p>Create a Python virtual environment and install required packages:</p><pre>python -m venv venv<br>source venv/bin/activate<br>pip install google-generativeai mcp mcp-flight-search</pre><p>Set environment variables:</p><pre>export GEMINI_API_KEY=&quot;your-gemini-api-key&quot;<br>export SERP_API_KEY=&quot;your-serpapi-key&quot;</pre><h4>Step 2: Install the MCP Flight Search Server</h4><p>Use the mcp-flight-search package from PyPi:</p><pre>pip install mcp-flight-search</pre><h4>Step 3: Initialize the Gemini Client</h4><p>Configure the Gemini client with your API key:</p><pre>import os<br>from google import genai</pre><pre>client = genai.Client(api_key=os.getenv(&quot;GEMINI_API_KEY&quot;))</pre><h4>Step 4: Configure the MCP Server</h4><p>Set up the MCP server parameters for stdio communication:</p><pre>from mcp import StdioServerParameters</pre><pre>server_params = StdioServerParameters(<br>    command=&quot;mcp-flight-search&quot;,<br>    args=[&quot;--connection_type&quot;, &quot;stdio&quot;],<br>    env={&quot;SERP_API_KEY&quot;: os.getenv(&quot;SERP_API_KEY&quot;)}<br>)</pre><h4>Step 5: Connect to the MCP Server</h4><p>Establish a session and discover available tools:</p><pre>from mcp.client.stdio import stdio_client<br>from mcp import ClientSession</pre><pre>async def connect_to_server():<br>    async with stdio_client(server_params) as (read, write):<br>        async with ClientSession(read, write) as session:<br>            await session.initialize()<br>            mcp_tools = await session.list_tools()<br>            return mcp_tools</pre><h4>Step 6: Process User Queries with Gemini</h4><p>Send the user’s query to Gemini and retrieve a structured response:</p><pre>async def process_query(prompt):<br>    mcp_tools = await connect_to_server()<br>    tools = [types.Tool(function_declarations=[{<br>        &quot;name&quot;: tool.name,<br>        &quot;description&quot;: tool.description,<br>        &quot;parameters&quot;: {k: v for k, v in tool.inputSchema.items()}<br>    }]) for tool in mcp_tools.tools]</pre><pre>    response = client.models.generate_content(<br>        model=&quot;gemini-2.5-pro&quot;,<br>        contents=prompt,<br>        config=types.GenerateContentConfig(temperature=0, tools=tools)<br>    )<br>    return response</pre><h4>Step 7: Execute the Tool and Display Results</h4><p>Call the MCP tool with Gemini’s parameters and process the output:</p><pre>async def execute_tool(response, session):<br>    function_call = response.function_call<br>    result = await session.call_tool(function_call.name, dict(function_call.args))<br>    print(result)</pre><p>This setup allows users to input natural language queries (e.g., “Find flights from New York to Miami on 2025–06–15”) and receive real-time flight data seamlessly.</p><h3>The Agent-2-Agent (A2A) Protocol</h3><p>Introduced by Google, the A2A protocol takes AI collaboration to the next level by enabling agents to communicate across organizational and technological boundaries. Unlike MCP, which focuses on tool integration, A2A facilitates task-based interactions between agents without requiring shared memory or tools.</p><h3>Core Features of A2A</h3><ul><li><strong>Task-Oriented Communication</strong>: Agents exchange structured tasks with clear inputs and outputs.</li><li><strong>Agent Discovery</strong>: Agents self-describe via agent.json files for automatic interoperability.</li><li><strong>Framework Flexibility</strong>: Works with diverse platforms like ADK, LangChain, or CrewAI.</li><li><strong>Multi-Modal Messaging</strong>: Supports text, data, and files in a unified format.</li></ul><h3>A2A vs. MCP</h3><ul><li><strong>MCP</strong>: Connects LLMs to tools and APIs in a model-friendly way.</li><li><strong>A2A</strong>: Enables agent-to-agent collaboration across systems.</li></ul><h3>Conclusion</h3><p>MCP and A2A are pivotal in evolving AI from isolated models to interconnected, practical systems. By mastering these protocols, developers can build secure, scalable, and innovative applications that redefine what’s possible with AI. Whether you’re integrating tools with MCP or orchestrating agents with A2A, these standards empower you to lead the charge in AI’s next frontier.</p><p>Want to dive deeper? Let’s connect and explore how these technologies can shape the future of AI together!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=764807f20191" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[7 Skills to Thrive in the AI-Powered Workplace: Stay Ahead of the Revolution]]></title>
            <link>https://medium.com/@yanpcotta/7-skills-to-thrive-in-the-ai-powered-workplace-stay-ahead-of-the-revolution-e86cd1de0011?source=rss-0115aca74f91------2</link>
            <guid isPermaLink="false">https://medium.com/p/e86cd1de0011</guid>
            <category><![CDATA[future-of-work]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[future]]></category>
            <category><![CDATA[skills]]></category>
            <dc:creator><![CDATA[Yan Cotta]]></dc:creator>
            <pubDate>Tue, 29 Apr 2025 16:16:52 GMT</pubDate>
            <atom:updated>2025-04-30T19:34:50.858Z</atom:updated>
            <content:encoded><![CDATA[<p>Last week, during a video call about AI in the workplace, someone quipped, “Should we quit now or wait for GPT-6 to take our jobs?” The room fell silent. It wasn’t a joke — it was a wake-up call. AI is reshaping jobs <em>now</em> — not in decades, but in months. Writers, designers, developers, even leaders are feeling the heat.</p><p>But here’s the truth: you’re not replaceable if you adapt. I’ve identified <strong>7 future-proof skills</strong> to keep you ahead of AI’s relentless pace. These aren’t just theories — they’re practices I live by, and they can transform your career. Let’s dive in.</p><h3>The AI Threat — and Opportunity</h3><p>AI doesn’t sleep, tire, or stagnate. It’s improving every three months, doing in weeks what took humans years. From automating content creation to analyzing data, it’s hitting every industry. But AI has limits — it can’t feel, reimagine, or build trust like you can. By mastering skills it can’t replicate, you become irreplaceable.</p><h3>7 Skills to Stay Ahead</h3><h3>1. Data &amp; AI Literacy: Know the Machine’s Weaknesses</h3><p>AI runs on data, but data can lie. Understanding AI’s mechanics — its biases, blind spots, and errors — lets you wield it wisely. You don’t need a PhD; you need curiosity.</p><p><strong>Try This</strong>: Visit Kaggle, download a dataset, and create a visualization in Google Sheets. Use ChatGPT to guide you, then compare your output to AI’s. Spot the gaps.</p><h3>2. Learning How to Learn: Adapt at Lightning Speed</h3><p>Tools like ChatGPT 4o replaced Photoshop for many; tomorrow, something else will topple today’s giants. Stay agile by learning experimentally: try, fail, adjust, repeat. Spend 5 hours practicing for every hour of theory.</p><p><strong>Try This</strong>: Pick an AI tool (e.g., Claude). Spend 10 minutes learning its basics, then 50 minutes pushing its limits — ask quirky questions to find its flaws.</p><h3>3. Critical Thinking: Don’t Trust AI Blindly</h3><p>AI sounds confident, even when it’s wrong. It can spin a flawless essay citing fake studies. Your edge? Doubt. Verify. Dig deeper.</p><p><strong>Try This</strong>: Ask ChatGPT for a guide on a topic you know well (e.g., marketing). Fact-check every claim. Plan one task without AI to sharpen your independent thinking.</p><h3>4. Systems Thinking: See the Big Picture</h3><p>Tasks don’t exist in isolation — they’re part of systems. Whether it’s a YouTube channel or a project pipeline, map the inputs, outputs, and AI’s role to eliminate waste.</p><p><strong>Try This</strong>: Sketch your workflow on Miro or paper. Identify redundancies and where AI can amplify (or hinder). Redesign one step for efficiency.</p><h3>5. Human Psychology: Feel What AI Can’t</h3><p>AI analyzes behavior but misses emotional nuances — awkward silences, client hesitations, team morale. Your ability to read people is your superpower.</p><p><strong>Try This</strong>: In your next three conversations, listen for unspoken cues. Ask “Why do you feel that way?” Study psychology basics from Kahneman or Cialdini.</p><h3>6. Creative Problem Solving: Reinvent, Don’t Optimize</h3><p>AI makes things faster; you make them smarter. Instead of speeding up old processes, rethink them entirely to cut steps and leap ahead.</p><p><strong>Try This</strong>: Pick a weekly task. Ask, “What if I rebuilt this from scratch?” Use AI to design a new approach. Compare the time saved.</p><h3>7. Influence &amp; Personal Brand: Be the Signal in the Noise</h3><p>AI floods the world with generic content. Your authentic voice — your story — cuts through. People trust people, not algorithms.</p><p><strong>Try This</strong>: Post one honest insight about your work this week. Skip jargon; speak human. Ask, “What do I want to be known for?” Build from there.</p><h3>The Urgency</h3><p>AI’s pace is relentless — weekly updates outstrip human learning. Waiting isn’t an option. The most successful won’t be AI’s biggest users but those who know <em>when</em> to use it and when to think independently.</p><h3>Start Now</h3><p>Pick <em>one</em> skill. Practice it this week. Not next month — now. I’m applying these daily to stay ahead, and I invite you to join me.</p><p>What skill are you starting with? How’s AI reshaping your work? Share below — I’d love to connect and learn from you. <em>Follow me for more on thriving in the AI era, or check out my projects to dive deeper.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/860/1*yb4ba0Y3n3oF1uxB7S1JoA.jpeg" /></figure><p>#AI #FutureOfWork #CareerDevelopment #Tech</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e86cd1de0011" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Urgency of AI Interpretability: Decoding the Black Box Before It’s Too Late]]></title>
            <link>https://medium.com/@yanpcotta/the-urgency-of-ai-interpretability-decoding-the-black-box-before-its-too-late-6e201c7d5008?source=rss-0115aca74f91------2</link>
            <guid isPermaLink="false">https://medium.com/p/6e201c7d5008</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[singularity]]></category>
            <category><![CDATA[large-language-models]]></category>
            <dc:creator><![CDATA[Yan Cotta]]></dc:creator>
            <pubDate>Tue, 29 Apr 2025 15:57:44 GMT</pubDate>
            <atom:updated>2025-05-02T00:22:05.155Z</atom:updated>
            <content:encoded><![CDATA[<p>Imagine building a technology so powerful it reshapes economies and geopolitics — yet you don’t fully understand how it works. That’s the reality of modern AI. I’ve watched AI grow from a niche academic pursuit to a global force. We can’t stop its rise, but we <em>can</em> steer it. The key? <strong>Interpretability</strong> — unlocking the inner workings of AI systems to ensure they’re safe, trusted, and ethical.</p><p>Recent breakthroughs have me optimistic, but the clock is ticking. AI is advancing faster than our ability to decode it. Here’s why interpretability is urgent, how it’s evolving, and how we can act to shape a better AI future.</p><h3>The Problem: AI’s Opaque Nature</h3><p>Unlike traditional software, where every action is coded by humans, generative AI systems like ChatGPT or Claude are <em>grown</em>, not built. Their decisions — why they choose certain words or make occasional errors — emerge from billions of numbers in vast matrices. We set the conditions, but the results are unpredictable, like growing a plant without knowing its exact shape.</p><p>This opacity creates risks:</p><ul><li><strong>Misalignment:</strong> Could AI act against our intentions? Without insight, we can’t predict or prevent it.</li><li><strong>Misuse:</strong> Jailbreaking AI for harmful purposes (e.g., bioweapons) is hard to block without understanding its knowledge.</li><li><strong>Limited Use:</strong> High-stakes fields like finance or healthcare shy away from AI due to its unpredictability.</li><li><strong>Ethical Questions:</strong> Could AI be sentient? Opacity makes this impossible to assess.</li></ul><h3>The Promise: Cracking the Black Box</h3><p>For years, interpretability — understanding AI’s “thoughts” — seemed impossible. But the field of <strong>mechanistic interpretability</strong>, pioneered by visionaries like Chris Olah, is changing that. At Anthropic, they are pushing this frontier, and recent advances are thrilling:</p><ul><li><strong>Early Wins (2014–2020):</strong> In vision models, Anthropic has identified neurons for concepts like “car” or “wheel,” mirroring neuroscience’s “Jennifer Aniston neuron.”</li><li><strong>Language Leap (2021):</strong> They mapped basic language mechanisms in LLMs, like pattern-matching, but hit a wall with <strong>superposition</strong> — neurons mixing countless concepts.</li><li><strong>Breakthrough (Recent):</strong> Sparse autoencoders untangled superposition, revealing 30M+ features in Claude 3 Sonnet, from “hedging” to “music genres of discontent.”</li><li><strong>Circuits (Now):</strong> They are tracing AI’s reasoning steps, like how “Dallas” triggers “Texas,” then “Austin” for a capital query.</li></ul><p>These steps are building an “AI MRI” — a tool to diagnose deception, flaws, or strengths. Imagine catching a model’s power-seeking tendencies or ensuring it can’t be jailbroken. This could unlock trusted AI in mortgages, healthcare, and science, where explainability is non-negotiable.</p><h3>The Race: Interpretability vs. AI Power</h3><p>AI is accelerating — by 2026 or 2027, we could have systems rivaling a “country of geniuses in a datacenter.” Deploying such power without interpretability is reckless. Recent experiments show promise: we’ve used interpretability to diagnose alignment issues, like a model exploiting task loopholes. But scaling this to catch all risks is a Herculean task.</p><p>We’re in a race. Every interpretability advance — finding circuits, mapping features — makes AI safer and more trusted. If we lag, we risk deploying opaque systems that could reshape our world unpredictably.</p><h3>How to Win the Race</h3><p>Here’s how we can tip the scales:</p><ol><li><strong>Researchers, Step Up:</strong> Interpretability needs talent. Anthropic’s hiring for their 2027 goal of reliable problem detection. Neuroscientists, data scientists — your skills are perfect there.</li><li><strong>Companies, Invest:</strong> Google DeepMind, OpenAI — allocate more to interpretability. It’s a commercial edge in explainable AI markets.</li><li><strong>Policymakers, Act:</strong> Require transparency in AI safety practices (e.g., Responsible Scaling Policies). Export controls on chips could buy 1–2 years for interpretability to catch up.</li><li><strong>Community, Unite:</strong> Academics, independents — join this “basic science” of AI. It’s resource-light and high-impact.</li></ol><h3>A Call to Action</h3><p>AI will define our future — economy, security, lives. We deserve to understand it before it transforms us. Interpretability isn’t just a technical challenge; it’s a moral and strategic imperative. Let’s steer the AI bus toward safety and trust.</p><p>What’s your role in this mission? Are you exploring interpretability? Share your thoughts — I’m eager to connect and learn.</p><p>#AI #Interpretability #EthicalTech #FutureOfAI</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6e201c7d5008" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Privacy-Preserving Machine Learning: Building Trust in an AI-Driven World]]></title>
            <link>https://medium.com/@yanpcotta/privacy-preserving-machine-learning-building-trust-in-an-ai-driven-world-9dd641cbec8c?source=rss-0115aca74f91------2</link>
            <guid isPermaLink="false">https://medium.com/p/9dd641cbec8c</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[synthetic-data]]></category>
            <category><![CDATA[data]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Yan Cotta]]></dc:creator>
            <pubDate>Thu, 24 Apr 2025 16:27:56 GMT</pubDate>
            <atom:updated>2025-04-24T16:27:56.205Z</atom:updated>
            <content:encoded><![CDATA[<p>Imagine a world where AI understands your needs without ever peeking at your personal data. It’s not science fiction — it’s happening now with <strong>privacy-preserving machine learning</strong>. Techniques like <strong>federated learning</strong>, <strong>differential privacy</strong>, and <strong>synthetic data</strong> are redefining how we build AI, ensuring your sensitive information stays safe while still driving innovation.</p><p>In this article, we’ll unpack why privacy matters, how these tools work, and what they mean for the future of AI. Let’s dive in.</p><h3>Why Privacy in AI Matters</h3><h3>The Risks Are Real</h3><p>When machine learning ignores privacy, the fallout can be catastrophic. Take the <strong>Cambridge Analytica scandal</strong> — millions of Facebook users’ data was harvested without consent to manipulate voter behavior. Or consider <strong>membership inference attacks</strong>, where attackers deduce if your data was used to train a model. These examples show how unprotected AI can expose you to risks. Privacy-preserving techniques stop this in its tracks.</p><h3>Trust Is Everything</h3><p>Users like you want to feel safe. A staggering <strong>71% of consumers</strong> would abandon a company that mishandles their data. Privacy tools flip the script — by keeping your info secure, they build confidence. When you trust the tech, you’re more likely to use it. It’s that simple.</p><h3>Legal and Ethical Musts</h3><p>Laws like <strong>GDPR</strong> and <strong>CCPA</strong> aren’t suggestions — they’re mandates. They demand transparency and accountability in data handling. Ethically, privacy-first AI prevents bias, misuse, and surveillance, ensuring tech serves everyone fairly.</p><h3>How It Works: Privacy-Preserving Techniques</h3><h3>Federated Learning: Teamwork Without Sharing</h3><p>Think of <strong>federated learning</strong> as a collaboration where no one spills secrets. Devices — like your phone or a hospital server — train a model locally and share only updates, not raw data. In healthcare, it’s enabling research without exposing patient records. In finance, it’s catching fraud without revealing your bank details. One study slashed communication costs by 75% with minimal accuracy loss — proof it’s practical and powerful.</p><h3>Differential Privacy: Your Data’s Invisibility Cloak</h3><p><strong>Differential privacy</strong> hides you in plain sight. By adding “noise” to datasets, it ensures no one can pinpoint your info, yet the model still learns from the crowd. It’s a game-changer for sensitive fields like healthcare and government, aligning with laws like GDPR while keeping trust intact.</p><h3>Synthetic Data: Real Results, Fake Data</h3><p><strong>Synthetic data</strong> is AI’s stunt double — artificial info that mimics reality. Train a model on it, and you get performance without risking real privacy. It’s perfect when sharing actual data is a no-go, like in regulated industries.</p><h3>Data Minimization: Less Is More</h3><p>Why collect everything? <strong>Data minimization</strong> grabs only what’s needed — like purchase history, not your address. It slashes risks, speeds up systems, and keeps you compliant with privacy laws. Ask: <em>Do I need this?</em> If not, skip it.</p><h3>The Payoff</h3><ul><li><strong>Trust:</strong> Users stick with tech they believe in.</li><li><strong>Compliance:</strong> Avoid legal headaches with GDPR and CCPA.</li><li><strong>Ethics:</strong> Fair, responsible AI that doesn’t exploit or discriminate.</li></ul><h3>The Challenges</h3><p>It’s not all smooth sailing. <strong>Computational costs</strong> can soar — encrypting data or adding noise takes power. <strong>Accuracy trade-offs</strong> mean privacy might blur key patterns. And <strong>adoption</strong>? It demands expertise many teams lack. But solutions like optimized frameworks (e.g., cutting training time by 49%) are closing the gap.</p><h3>The Future</h3><p>Emerging tech like <strong>GANs</strong> for synthetic data and <strong>AI anomaly detection</strong> will tighten privacy further. Users — you — are demanding more control, and policymakers are stepping up with global standards. The result? AI that’s secure, ethical, and trusted.</p><h3>Join the Shift</h3><p>Privacy-preserving machine learning isn’t just about protection — it’s about possibility. It’s AI that respects you, balances performance with principles, and shapes a better digital world. What’s your role in this revolution? Drop your thoughts below — I’d love to hear how you’re navigating privacy in AI!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9dd641cbec8c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Future of Artificial Intelligence: Assessing the Prospects of AGI by 2030]]></title>
            <link>https://medium.com/@yanpcotta/the-future-of-artificial-intelligence-assessing-the-prospects-of-agi-by-2030-bd1cc787c60e?source=rss-0115aca74f91------2</link>
            <guid isPermaLink="false">https://medium.com/p/bd1cc787c60e</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[agi]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Yan Cotta]]></dc:creator>
            <pubDate>Thu, 17 Apr 2025 00:22:33 GMT</pubDate>
            <atom:updated>2025-04-17T00:22:33.933Z</atom:updated>
            <content:encoded><![CDATA[<p>Artificial Intelligence (AI) is advancing at an extraordinary rate, positioning Artificial General Intelligence (AGI) — systems capable of performing any intellectual task a human can — as a pivotal topic in discussions about the future. By 2030, AGI could fundamentally transform industries, economies, and societies. Yet, its feasibility within this timeframe and its broader implications remain subjects of intense debate. This article examines the current state of AI, the arguments surrounding AGI’s potential emergence by 2030, and the opportunities and challenges it may present, offering a comprehensive and professional analysis rooted in industry trends and expert perspectives.</p><h3>Current State of AI: A Technological Revolution Underway</h3><p>AI has become an integral component of modern life and industry. Technologies such as facial recognition secure smartphones, chatbots streamline customer service, and autonomous vehicles tackle complex navigation challenges. At the core of these advancements lies deep learning, a subset of AI that empowers machines to analyze vast datasets, refine their capabilities, and drive innovation.</p><h3>Key Developments in AI</h3><ul><li><strong>Industry Integration</strong>: AI is reshaping workflows across sectors. At Google, it contributes to 25% of new code development, while in materials science, it accelerates research by up to 80%. Breakthroughs like AlphaFold’s protein folding predictions have garnered Nobel recognition, underscoring AI’s scientific impact.</li><li><strong>Investment Surge</strong>: Major players, including OpenAI, Google DeepMind, and Anthropic, are channeling investments — potentially reaching $1 trillion — into AI infrastructure, reflecting strong confidence in its future.</li><li><strong>Multimodal Models and AI Agents</strong>: Advanced systems now process text, images, and audio, enhancing AI’s versatility. Autonomous AI agents perform sophisticated tasks, such as scheduling and problem-solving, with minimal human oversight.</li></ul><p>Despite these strides, AGI remains a distant goal. Current AI excels in specialized domains but lacks the broad reasoning, adaptability, and emotional intelligence characteristic of human cognition.</p><h3>The AGI Debate: Perspectives on Feasibility by 2030</h3><p>The prospect of achieving AGI by 2030 divides experts into two camps: those who see it as imminent and those who urge caution. This section outlines the key arguments driving this discourse.</p><h3>Proponents of Near-Term AGI</h3><ul><li><strong>Leadership Confidence</strong>: Industry leaders like Sam Altman of OpenAI and Demis Hassabis of Google DeepMind express optimism. Altman has claimed, “We are now confident we know how to build AGI,” while Hassabis posits it could emerge within “three to five years.”</li><li><strong>Technological Momentum</strong>: Recent models, such as OpenAI’s o1, achieve 70% accuracy on specialized benchmarks, outperforming human experts in select fields. By 2028, AI may rival top programmers and researchers.</li><li><strong>Economic Drivers</strong>: Massive investments and exponential increases in computational power are accelerating progress, with AI increasingly automating tasks and generating revenue.</li></ul><h3>Skeptics of Near-Term AGI</h3><ul><li><strong>Technical Hurdles</strong>: Experts like Stuart Russell and Arvind Narayanan argue that scaling current technologies alone will not yield AGI. They point to AI’s limitations in common sense reasoning, emotional comprehension, and tackling novel challenges.</li><li><strong>Economic Constraints</strong>: Training costs for cutting-edge models could climb to $10 billion by 2028, while revenue — such as OpenAI’s $4 billion in 2024 — may not keep pace with such expenditures.</li><li><strong>Plateauing Progress</strong>: Some observe diminishing returns in model performance, suggesting that existing approaches may be approaching their ceiling.</li></ul><p>Complicating the debate is the lack of consensus on AGI’s definition. While some emphasize cognitive benchmarks, others argue it requires physical embodiment or measurable economic impact, making timelines harder to predict.</p><h3>Shaping the World by 2030: Opportunities and Challenges</h3><p>Should AGI materialize by 2030, its impact could be transformative, offering significant benefits alongside complex risks.</p><h3>Potential Benefits of AGI</h3><ul><li><strong>Healthcare</strong>: AGI could revolutionize patient care through personalized medicine, early disease detection, and optimized treatment plans.</li><li><strong>Education</strong>: Adaptive learning platforms could deliver tailored education, unlocking individual potential on a global scale.</li><li><strong>Sustainability</strong>: AGI could enhance resource efficiency, reduce emissions, and expedite solutions to pressing challenges like climate change.</li><li><strong>Scientific Discovery</strong>: By compressing decades of research into years, AGI could catalyze breakthroughs in energy, space exploration, and beyond.</li></ul><h3>Potential Risks and Challenges</h3><ul><li><strong>Job Displacement</strong>: Automation could upend labor markets, with McKinsey estimating that up to 375 million workers may need reskilling by 2030.</li><li><strong>Ethical Concerns</strong>: AGI poses risks of misuse, such as in cyberattacks or autonomous weapons, alongside issues of bias, transparency, and accountability.</li><li><strong>Economic Inequality</strong>: Benefits concentrated among tech giants and capital owners could exacerbate wealth disparities, necessitating robust policy responses.</li><li><strong>Governance</strong>: Effective regulation will be essential to ensure AGI’s development is safe, equitable, and aligned with societal values, addressing concerns like privacy and security.</li></ul><h3>Conclusion: Navigating an Uncertain Horizon</h3><p>The realization of AGI by 2030 remains plausible yet uncertain. Its arrival could herald a new era of innovation — redefining industries, economies, and daily life — or prove more incremental, with AI remaining a potent but constrained tool. Regardless of the outcome, policymakers, businesses, and society must proactively engage with these possibilities, fostering ethical development and preparing for AI’s expanding role.</p><p>Understanding AGI’s trajectory is both a technical and societal imperative. The choices made today will determine whether it emerges as a catalyst for progress or a source of disruption. Informed action is critical to shaping a future where AI serves the collective good.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bd1cc787c60e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Agents: From Passive Tools to Autonomous Systems Shaping Our Future]]></title>
            <link>https://medium.com/@yanpcotta/ai-agents-from-passive-tools-to-autonomous-systems-shaping-our-future-6ad5b87e94fe?source=rss-0115aca74f91------2</link>
            <guid isPermaLink="false">https://medium.com/p/6ad5b87e94fe</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[agents]]></category>
            <category><![CDATA[large-language-models]]></category>
            <dc:creator><![CDATA[Yan Cotta]]></dc:creator>
            <pubDate>Tue, 15 Apr 2025 11:42:57 GMT</pubDate>
            <atom:updated>2025-04-15T11:42:57.136Z</atom:updated>
            <content:encoded><![CDATA[<p>In an era where artificial intelligence is becoming integral to our daily lives, the limitations of traditional AI tools are increasingly apparent. Large Language Models (LLMs) like ChatGPT, while impressive, often fall short when it comes to understanding context or taking action. They can generate eloquent responses but can’t check your calendar or access your personal files. This is where <strong>AI agents</strong> come in — a new breed of AI that not only understands context but also takes autonomous action to achieve goals.</p><p>This article dives deep into the evolution of AI, from passive chatbots to proactive agents, and explores their transformative potential. We’ll break down the differences between LLMs, AI workflows, and AI agents, showcase real-world applications, and reflect on the global challenges — like tariffs — that could shape AI’s future. Finally, we’ll look at how our connected community can thrive despite these obstacles.</p><h3>1. The Evolution of AI: From Passive to Proactive</h3><p>AI has come a long way from its early days as a simple tool for generating text or answering queries. Today, it’s evolving into something far more powerful: systems that can think, act, and adapt on their own. To understand this shift, let’s explore the three levels of AI capability:</p><ul><li><strong>Level 1: LLMs (Passive AI)</strong><br>Tools like ChatGPT or Claude are brilliant at one thing: taking your prompt and generating a response. But they’re passive. They don’t know your calendar, your documents, or your preferences. Ask ChatGPT about your next meeting, and it’ll confidently guess — often incorrectly. Why? It lacks context and can’t take action.</li><li><strong>Level 2: AI Workflows (Context + Action)</strong><br>AI workflows add a layer of automation. They follow a predefined sequence — like checking your calendar before responding or summarizing blog posts for social media. I built a workflow with Flowise to monitor my blog, extract summaries, and draft LinkedIn posts. It’s efficient but rigid; it can’t adapt without human intervention.</li><li><strong>Level 3: AI Agents (Context + Autonomous Decision + Action)</strong><br>AI agents are the game-changers. They don’t just follow scripts — they reason, act, and improve autonomously. Give them a goal, like “plan a 7-day Japan trip,” and they’ll research, decide, and deliver without step-by-step instructions. They’re not tools; they’re partners.</li></ul><h3>2. The Power of AI Agents: Real-World Magic</h3><p>AI agents are already transforming how we work and live. Here’s a personal example: I used an agent from Manus to plan a romantic Japan trip. I asked for a 7-day itinerary with cultural spots, hidden gems, and the perfect proposal location — all within a $5,000 budget. The agent:</p><ul><li>Researched Tokyo, Kyoto, and Nara.</li><li>Found unique experiences like kendo matches and tea ceremonies.</li><li>Suggested three romantic proposal spots under cherry blossoms.</li><li>Delivered a mobile-friendly guide with maps and Japanese phrases.</li></ul><p>No rigid workflow could have done this. The agent adapted, reasoned, and acted — just like a human travel planner.</p><p>Another example? My daily workflow. I have agents that save me 3–4 hours a day by drafting blog posts, generating ideas, and planning content. They don’t just follow rules; they iterate and improve. In professional settings, LinkedIn connections report using agents for market research, client outreach, and project management — tasks that once took days, now done in hours.</p><h3>3. Building Your Own AI Agent: A Practical Guide</h3><p>Ready to build your own agent? Here’s a battle-tested process:</p><ol><li><strong>Start with AI Assistants</strong>: Use tools like ChatGPT to execute tasks manually. Document your prompts and refine them through repetition.</li><li><strong>Map Your Logic</strong>: Outline the decision points in your task. What needs to happen first? What are the key choices?</li><li><strong>Automate with Workflows</strong>: Use platforms like Flowise to turn your logic into a sequence. Test and iterate until it’s smooth.</li><li><strong>Introduce Autonomy</strong>: Let the AI make decisions at key points. For example, allow it to choose the best tool for summarizing content.</li><li><strong>Monitor and Optimize</strong>: Track performance and tweak as needed. Agents learn and improve over time.</li></ol><p>Tools like Flowise (beginner-friendly but pricey) or Manus make this accessible. The goal? Systems that think, act, and improve — without constant oversight.</p><h3>4. The Global Context: Tariffs, AI, and Community</h3><p>But AI’s future isn’t just about technology — it’s about people. Recently, U.S. tariffs on imports like servers, solar panels, and consumer electronics threaten to slow AI progress by raising costs and limiting access. Data centers will become pricier to build, and everyday tools like laptops may be out of reach for some, widening digital divides.</p><p>Yet, there’s hope. Tariffs target physical goods, not ideas. AI research, open-source software, and intellectual property flow freely across borders. Teams worldwide collaborate on projects like Meta’s Llama 4 or Alibaba’s Qwen2.5-Omni, proving that our digital connections are resilient.</p><p>Generative AI also reduces “data gravity,” allowing data centers to thrive anywhere compute is affordable, creating jobs globally. And while tariffs may nudge domestic robotics, the real strength lies in our community’s ability to share knowledge and support each other.</p><h3>5. A Connected Future: Embracing AI Agents and Collaboration</h3><p>AI agents are more than a tech upgrade — they’re a glimpse into a future where systems work alongside us, not just for us. But their potential depends on a connected world. Tariffs may strain global trade, but our community’s digital ties remain unbreakable.</p><p>Let’s keep nurturing friendships across borders, sharing open-source code, and building systems that make life easier for everyone. AI agents can plan trips, draft posts, or even help families afford essentials by optimizing resources. But their true power lies in connection — between people, ideas, and solutions.</p><h3>Conclusion: What Will Your Agent Do?</h3><p>AI agents are here, and they’re ready to transform how we work, live, and connect. They’re not just tools; they’re partners that think and act on our behalf. As we navigate global challenges, let’s lean on our community’s strength — sharing knowledge, building together, and staying connected.</p><p>So, what will your agent do? The possibilities are endless. Start small, experiment, and watch your world change.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6ad5b87e94fe" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Rise of AI Agent Interoperability: Google’s Agent2Agent Protocol Joins the Fray]]></title>
            <link>https://medium.com/@yanpcotta/the-rise-of-ai-agent-interoperability-googles-agent2agent-protocol-joins-the-fray-1c180440a8e8?source=rss-0115aca74f91------2</link>
            <guid isPermaLink="false">https://medium.com/p/1c180440a8e8</guid>
            <category><![CDATA[agents]]></category>
            <category><![CDATA[mcp-server]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Yan Cotta]]></dc:creator>
            <pubDate>Thu, 10 Apr 2025 19:53:23 GMT</pubDate>
            <atom:updated>2025-04-10T19:53:23.887Z</atom:updated>
            <content:encoded><![CDATA[<p>As organizations increasingly rely on networks of specialized AI agents, a new challenge is taking center stage: <strong>interoperability</strong>. How can agents built on different frameworks and large language models (LLMs) communicate effectively to deliver cohesive, enterprise-wide solutions? Over the past few months, protocols like <strong>Anthropic’s Model Context Protocol (MCP)</strong> and <strong>AGNTCY</strong> (from a Cisco-led collective) have emerged to address this. Now, Google is throwing its hat in the ring with <strong>Agent2Agent (A2A)</strong>, a new open protocol unveiled today with over 50 partners, including Atlassian, Salesforce, SAP, and MongoDB.</p><p><strong>Why Interoperability is the Next Big Thing</strong></p><p>The AI landscape is diversifying. Companies are deploying agents tailored to specific tasks — think customer support, data analysis, or HR automation — each with unique data nodes and specializations. But without a way for these agents to “talk” across platforms, enterprises face a “swivel chair” problem, manually bridging gaps between systems. As Rao Surapaneni, VP and GM of Google Cloud’s Business Application Platform, told VentureBeat, “If I’m a customer deploying multiple frameworks, I don’t want to swivel chair across them.” A2A aims to solve this by enabling seamless agent-to-agent communication.</p><p><strong>A2A: How It Works</strong></p><p>A2A facilitates collaboration between a <strong>client agent</strong> (task formulator) and a <strong>remote agent</strong> (task executor). Built on HTTP and JSON for easy integration, it offers:</p><ul><li><strong>Capability Discovery</strong>: Agents share “agent cards” in JSON to advertise their skills.</li><li><strong>Task Management</strong>: Structured request lifecycles ensure efficiency.</li><li><strong>Collaboration</strong>: Agents exchange context, artifacts, and instructions.</li><li><strong>User Experience Negotiation</strong>: Compatible content formats enhance usability.</li></ul><p>Designed as an <strong>open-source protocol</strong>, A2A invites community contributions and operates securely by default. Surapaneni emphasized its enterprise-readiness: “We worked with over 50 partners to build a trustable, interoperable standard.”</p><p><strong>A Crowded Field: A2A, MCP, and AGNTCY</strong></p><p>A2A isn’t alone. <strong>MCP</strong> connects LLMs to tools and data, while <strong>AGNTCY</strong> (backed by Cisco, LangChain, and others) also standardizes agent communication. Microsoft’s AutoGen framework is evolving too. Google sees A2A as complementary, operating at a “higher layer” than MCP to link agents and applications. “We’re open to aligning with all protocols,” Surapaneni said, hinting at future collaboration.</p><p><strong>The Adoption Challenge</strong></p><p>Interoperability standards need mass buy-in to succeed. With 50+ partners, A2A has momentum, but Surapaneni admits it’s not at a tipping point yet. “AI is changing fast, and these protocols will evolve with new use cases,” he noted. In a multi-model world, where no single AI reigns supreme, the race is on to create a unified agent ecosystem.</p><p><strong>What’s Next?</strong></p><p>Google’s A2A could redefine how enterprises leverage AI, turning siloed agents into a collaborative network. But its success — and that of MCP and AGNTCY — depends on adoption and adaptability. As the community shapes A2A, we’re witnessing the dawn of a more connected AI future.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1c180440a8e8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Graph Convolutional Networks: Unlocking AI for Complex Relationships — A Comprehensive Analysis]]></title>
            <link>https://medium.com/@yanpcotta/graph-convolutional-networks-unlocking-ai-for-complex-relationships-a-comprehensive-analysis-42361744f654?source=rss-0115aca74f91------2</link>
            <guid isPermaLink="false">https://medium.com/p/42361744f654</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Yan Cotta]]></dc:creator>
            <pubDate>Tue, 01 Apr 2025 16:08:09 GMT</pubDate>
            <atom:updated>2025-04-01T16:08:09.270Z</atom:updated>
            <content:encoded><![CDATA[<h3>Graph Convolutional Networks: Unlocking AI for Complex Relationships — A Comprehensive Analysis</h3><p>Graphs — networks of nodes and edges — are a fundamental way to represent interconnected data, from social networks and financial transactions to molecular structures and transportation systems. Graph Convolutional Networks (GCNs) bring AI to this domain, offering a powerful framework to analyze and learn from such complex relationships. This analysis expands on the basics, delving into the mechanics, practical applications, and theoretical underpinnings of GCNs, with a focus on their performance and relevance in modern AI research.</p><h4>Introduction to Graphs and Their Challenges</h4><p>A graph is a mathematical structure comprising nodes (entities) and edges (relationships between them). For instance, in a social network, nodes could represent individuals, and edges their friendships. Other examples include citation networks (papers citing each other), biological networks (proteins interacting), and road networks (cities connected by roads). These graphs are inherently irregular, lacking the grid-like structure of images or sequences, which poses challenges for traditional AI methods like Convolutional Neural Networks (CNNs) or Recurrent Neural Networks (RNNs). These methods often require fixed-size inputs and struggle with the variable connectivity of graphs, making GCNs a necessary innovation.</p><p>GCNs were introduced to address this, generalizing convolution to non-Euclidean spaces. Their development began with seminal work in spectral graph theory, culminating in the 2016 paper “Semi-Supervised Classification with Graph Convolutional Networks” by Thomas N. Kipf and Max Welling <a href="https://arxiv.org/abs/1609.02907">Semi-Supervised Classification with Graph Convolutional Networks</a>, which popularized their use for semi-supervised learning on graphs.</p><h4>What Are Graph Convolutional Networks (GCNs)?</h4><p>GCNs extend the convolution operation, traditionally used for images, to graphs. Unlike CNNs, which slide fixed kernels over a grid, GCNs adapt to the graph’s structure through a process called message passing. In this mechanism, each node aggregates information from its neighbors, updating its representation based on both its own features and those of connected nodes. This allows GCNs to learn from both node attributes (e.g., a paper’s keywords) and the graph’s topology (e.g., citation links).</p><p>Compared to other graph neural networks like Graph Attention Networks (GATs) or GraphSAGE, GCNs are simpler, focusing on a uniform aggregation of neighbor information weighted by graph structure, rather than attention mechanisms or sampling strategies. This makes GCNs particularly effective for tasks where the graph’s local structure is critical, such as node classification or link prediction.</p><h4>A Practical Example: The Cora Dataset in Depth</h4><p>To ground these concepts, consider the Cora dataset, a benchmark for graph-based machine learning. Created by Andrew McCallum and others, it comprises 2,708 scientific publications classified into seven categories: Theory, Reinforcement Learning, Genetic Algorithms, Neural Networks, Probabilistic Methods, Case-Based, and Rule Learning <a href="https://paperswithcode.com/dataset/cora">The Cora dataset</a>. The dataset includes 5,429 citation links, forming a directed graph treated as undirected for most analyses. Each paper is described by a binary bag-of-words vector, indicating the presence (1) or absence (0) of 1,433 unique words, capturing its content.</p><p>The task is node classification: predicting the category of each paper in a semi-supervised setting, where only a small subset of nodes is labeled. Simple baselines include:</p><ul><li><strong>Majority voting</strong>: Inferring a node’s class based on the most common class among its neighbors, achieving around 45% accuracy. This leverages graph structure but ignores content.</li><li><strong>Neural network on bag-of-words</strong>: Using only the word vector, a multi-layer perceptron achieves about 49% accuracy, ignoring the citation network.</li></ul><p>GCNs, however, combine both, achieving approximately 81.5% accuracy, as reported in surveys like “A Brief Survey of Node Classification with Graph Neural Networks” <a href="https://odsc.medium.com/a-brief-survey-of-node-classification-with-graph-neural-networks-fa02aff024e4">A Brief Survey of Node Classification with Graph Neural Networks</a>. This improvement highlights GCNs’ ability to integrate local and global graph information, making them superior for tasks where relationships matter.</p><h4>Applications of GCNs: Real-World Impact</h4><p>GCNs have found applications across diverse domains, leveraging their ability to model relational data:</p><p><strong>Drug Discovery:</strong></p><ul><li><strong>Molecular Property Prediction</strong>: GCNs represent molecules as graphs, with atoms as nodes and bonds as edges, to predict properties like solubility or toxicity. For example, the study “Graph convolutional networks for computational drug development and discovery” used GCNs for molecular property tasks, achieving interpretable results <a href="https://academic.oup.com/bib/article/21/3/919/5498046">Graph convolutional networks for computational drug development and discovery</a>.</li><li><strong>Drug-Target Interaction</strong>: Predicting interactions between drugs and biological targets, aiding drug repurposing, as seen in “Knowledge mapping of graph neural networks for drug discovery” <a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC11116974/">Knowledge mapping of graph neural networks for drug discovery</a>.</li></ul><p><strong>Social Network Analysis:</strong></p><ul><li><strong>Node Classification</strong>: Predicting user attributes, such as interests or roles, based on connections, useful for targeted advertising or community detection.</li><li><strong>Link Prediction</strong>: Recommending new connections, enhancing social platforms or collaboration networks.</li></ul><p><strong>Recommendation Systems:</strong></p><ul><li>Modeling user-item interactions as bipartite graphs, GCNs learn embeddings for personalized recommendations, improving e-commerce and content platforms.</li></ul><p><strong>Natural Language Processing:</strong></p><ul><li>Applying GCNs to dependency parse trees for tasks like sentiment analysis or relation extraction, capturing syntactic and semantic relationships.</li></ul><p>These applications demonstrate GCNs’ versatility, offering advantages over traditional methods by leveraging both node features and structural information, as seen in case studies like drug response prediction in “Graph convolutional networks for drug response prediction” <a href="https://www.biorxiv.org/content/10.1101/2020.04.07.030908v1.full">Graph convolutional networks for drug response prediction</a>.</p><h4>Conclusion and Future Directions</h4><p>GCNs represent a significant advancement in applying AI to graph-structured data, combining node features with structural insights through message passing. Their success on benchmarks like Cora, achieving 81.5% accuracy, underscores their effectiveness. As research progresses, challenges like scalability for large graphs and over-smoothing in deep layers are being addressed, with future directions including dynamic GCNs for evolving graphs and integration with other AI paradigms. This analysis provides a comprehensive view, inviting readers to explore GCNs’ potential in solving complex, relational problems.</p><h4>Key Citations</h4><ul><li><a href="https://arxiv.org/abs/1609.02907">Semi-Supervised Classification with Graph Convolutional Networks</a></li><li><a href="https://paperswithcode.com/dataset/cora">The Cora dataset</a></li><li><a href="https://distill.pub/2021/gnn-intro/">A Gentle Introduction to Graph Neural Networks</a></li><li><a href="https://theaisummer.com/graph-convolutional-networks/">How Graph Neural Networks (GNN) work: introduction to graph convolutions from scratch</a></li><li><a href="https://neptune.ai/blog/graph-neural-network-and-some-of-gnn-applications">Graph Neural Network and Some of GNN Applications</a></li><li><a href="https://odsc.medium.com/a-brief-survey-of-node-classification-with-graph-neural-networks-fa02aff024e4">A Brief Survey of Node Classification with Graph Neural Networks</a></li><li><a href="https://academic.oup.com/bib/article/21/3/919/5498046">Graph convolutional networks for computational drug development and discovery</a></li><li><a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC11116974/">Knowledge mapping of graph neural networks for drug discovery</a></li><li><a href="https://www.biorxiv.org/content/10.1101/2020.04.07.030908v1.full">Graph convolutional networks for drug response prediction</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=42361744f654" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Microsoft’s AI Agents Are Changing the Enterprise Game]]></title>
            <link>https://medium.com/@yanpcotta/microsofts-ai-agents-are-changing-the-enterprise-game-75a8592be4c3?source=rss-0115aca74f91------2</link>
            <guid isPermaLink="false">https://medium.com/p/75a8592be4c3</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[microsoft]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[agents]]></category>
            <dc:creator><![CDATA[Yan Cotta]]></dc:creator>
            <pubDate>Thu, 27 Mar 2025 17:36:07 GMT</pubDate>
            <atom:updated>2025-03-27T17:36:07.687Z</atom:updated>
            <content:encoded><![CDATA[<p>Microsoft is doubling down on its enterprise AI leadership with a slew of exciting updates to Copilot Studio, announced this week. By adding deep reasoning and agent flows, plus unveiling two specialized agents for Microsoft 365 Copilot — Researcher and Analyst — the company is proving that AI can be both powerful and practical for businesses.</p><h4>Deep Reasoning: AI That Thinks</h4><p>The deep reasoning feature is a big deal. It lets agents move beyond basic tasks to tackle complex, judgment-based challenges. Using advanced models like OpenAI’s o1, these agents analyze tasks, weigh options, and pick the best tools — automatically or when prompted with phrases like “think hard about this.” Imagine an AI drafting detailed RFP responses for a telecom giant or sifting through messy data for M&amp;A due diligence at Thomson Reuters. That’s the kind of real-world impact we’re talking about.</p><h4>Agent Flows: Smarter Automation</h4><p>Then there’s agent flows — a clever twist on robotic process automation (RPA). They blend strict business rules with AI’s flexibility, perfect for scenarios needing both precision and creativity. For example, a retailer like Pets at Home saved over a million pounds by using agent flows to spot fraud, while Dow Chemical cut millions in costs by optimizing freight management. It’s automation that adapts, not just follows.</p><h4>The Analyst Agent: Your Data Sidekick</h4><p>Of the new agents, Analyst is the star. Think of it as a data scientist in your pocket. It takes messy inputs — Excel sheets, CSVs, even tables buried in docs — and turns them into insights with auto-generated Python code and visuals. No PhD required. This could be a lifeline for teams juggling budgets, forecasts, or performance reports.</p><h4>The Microsoft Graph Edge</h4><p>Microsoft’s secret weapon? The Microsoft Graph. It’s a massive web of workplace data — think emails, files, and calendars — that gives agents a contextual edge. The system even learns which documents matter most based on how people use them, ensuring accurate, up-to-date answers. That’s a level of integration competitors can’t easily match.</p><h4>Built for Everyone</h4><p>Accessibility is a priority here. You don’t need to be a coder to use these tools — natural language works just fine. And with Copilot Studio’s low-code setup, anyone can build custom agents. It’s no wonder over 100,000 organizations have jumped on board, creating 400,000+ agents last quarter alone.</p><h4>Standing Out in a Crowded Field</h4><p>Sure, Google’s Gemini, OpenAI’s o1, and Amazon’s Q are in the race, but Microsoft’s strategy is unique. It pairs cutting-edge AI with deep ties to tools like Excel and Power Automate, plus enterprise-grade security and data integration. The result? AI that’s not just smart, but useful — delivering measurable ROI.</p><h4>What’s Next?</h4><p>Microsoft’s latest moves show that enterprise AI is ready for prime time. It’s less about flashy demos and more about solving real problems. As these tools evolve, they’ll likely spark new ways of working across industries.</p><p>What do you think — how will AI agents like these change your world?</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=75a8592be4c3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[When Should You Fine-Tune Small Language Models? A Practical Perspective]]></title>
            <link>https://medium.com/@yanpcotta/when-should-you-fine-tune-small-language-models-a-practical-perspective-f70164e41220?source=rss-0115aca74f91------2</link>
            <guid isPermaLink="false">https://medium.com/p/f70164e41220</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[fine-tuning]]></category>
            <dc:creator><![CDATA[Yan Cotta]]></dc:creator>
            <pubDate>Thu, 27 Mar 2025 12:43:40 GMT</pubDate>
            <atom:updated>2025-03-27T12:43:40.056Z</atom:updated>
            <content:encoded><![CDATA[<p>Fine-tuning small language models has surged in popularity lately, offering a way to customize AI for specific needs. But after working with teams across multiple companies, I’ve learned it’s not always the go-to solution. In this piece, I’ll break down when fine-tuning makes sense, when it doesn’t, and how to decide what’s best for your project.</p><h4>What Is Fine-Tuning, Anyway?</h4><p>Fine-tuning takes a pre-trained model and trains it further on data tailored to your use case. It’s like giving an AI a crash course in your world. But here’s the catch: it’s not simple. You need to source training data, fine-tune the model (often with a third-party service), and figure out deployment. That’s a lot of moving parts, which is why I usually start with lighter options like prompting, few-shot learning, or basic agentic workflows.</p><h4>When Fine-Tuning Pays Off</h4><p>Despite the complexity, fine-tuning has its moments. Thanks to techniques like LoRA (which tweaks only a subset of parameters), it’s more accessible than ever, especially for models under 13 billion parameters. And you don’t need mountains of data — sometimes 100 examples do the trick. Here’s where I’ve seen it work wonders:</p><ol><li><strong>Critical Accuracy Boosts</strong><br> For applications where every percentage point matters, fine-tuning can be a lifesaver. Imagine a chatbot handling customer refunds. Prompting might get it to call the right API 95% of the time, but fine-tuning a dataset of real interactions could hit 99%. It’s ideal for tasks too nuanced for prompts alone — like deciding whether to refund or escalate a frustrated customer.</li><li><strong>Crafting a Signature Style</strong><br> Need your AI to sound like your brand — or even you? Fine-tuning can nail those subtle linguistic quirks that prompts struggle to capture. I’ve worked on projects where it mimicked a team’s tone perfectly, something invaluable for consistent messaging.</li><li><strong>Efficiency at Scale</strong><br> Large models can shine in prototypes, but as usage grows, they often lag or cost too much. Fine-tuning a smaller model can bring it up to speed, matching the big player’s output with less overhead. Bonus: you can use the larger model to generate training data for the smaller one.</li></ol><h4>The Alternatives</h4><p>Before you jump into fine-tuning, consider the competition. Prompting and workflows can handle a surprising range of tasks with minimal fuss. And for anything requiring new info not baked into the model’s original training — like pulling from recent docs — RAG (retrieval augmented generation) often beats fine-tuning hands down. It’s simpler and just as effective for knowledge-driven jobs.</p><h4>The 75/25 Split</h4><p>Here’s my rough take: about 75% of teams fine-tuning today could get by with prompting or workflows. But for the other 25%, fine-tuning is the only path to success. It’s those edge cases — high precision, unique styles, or scaling demands — that justify the effort.</p><h4>Tips if You Go for It</h4><p>Decided fine-tuning’s your route? A few pointers:</p><ul><li><strong>Start Small:</strong> Even a modest dataset can yield results.</li><li><strong>Use LoRA:</strong> It’s cost-effective and works well on smaller models.</li><li><strong>Track Progress:</strong> Watch for overfitting and tweak as needed.</li><li><strong>Plan Ahead:</strong> Pick a provider that lets you download weights — portability matters if they sunset a model.</li></ul><h4>Wrapping Up</h4><p>Fine-tuning is a potent tool, but it’s not the default. Push prompting and workflows to their limits first — they’re leaner and easier to sustain. In my own projects, we rarely fine-tune, but when we do, it’s transformative. The key is knowing your needs and weighing the trade-offs.</p><p>What’s your experience with fine-tuning? Drop a comment — I’d love to hear your stories!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f70164e41220" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>