<?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 Jian Ruan on Medium]]></title>
        <description><![CDATA[Stories by Jian Ruan on Medium]]></description>
        <link>https://medium.com/@jianruan18?source=rss-6205971bc42------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*MpCbd_hjEQ4R9PDYJ2CvzQ.jpeg</url>
            <title>Stories by Jian Ruan on Medium</title>
            <link>https://medium.com/@jianruan18?source=rss-6205971bc42------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 06:41:49 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@jianruan18/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[Version Control for Life: How AI Changed the Way I Grow]]></title>
            <link>https://levelup.gitconnected.com/version-control-for-life-how-ai-changed-the-way-i-grow-9ccb9f42ace8?source=rss-6205971bc42------2</link>
            <guid isPermaLink="false">https://medium.com/p/9ccb9f42ace8</guid>
            <category><![CDATA[mental-models]]></category>
            <category><![CDATA[personal-development]]></category>
            <category><![CDATA[claude-code]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Jian Ruan]]></dc:creator>
            <pubDate>Wed, 27 Aug 2025 14:36:32 GMT</pubDate>
            <atom:updated>2025-08-27T14:36:32.333Z</atom:updated>
            <content:encoded><![CDATA[<p>At the beginning of this year, I opened a markdown file titled “Q2 Fitness Goals” and realized something unsettling: it was identical to my Q1 Fitness goals. And Q4 from the previous year. The timestamps had changed, but I hadn’t.</p><p>As an engineer who ships code daily, this hit differently. At work, I live by the mantra “ship early, ship often.” Every bug fixed, every feature deployed — it’s all tracked, versioned, measured. My professional life runs on two-week sprints. But my personal life? It was running on a release cycle measured in years.</p><p>I sat there, staring at those unchanged goals, and asked myself: <em>What if I treated my personal growth like I treat my code?</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*DgAlEHmYxqdimln_" /><figcaption>Photo by <a href="https://unsplash.com/@yancymin?utm_source=medium&amp;utm_medium=referral">Yancy Min</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h3>The Experiment: Git for Life</h3><p>I did something that might sound strange to non-engineers: I created a private Git repository for my life.</p><p>Inside, markdown files for everything — career goals, financial targets, relationship aspirations, health metrics. Each update became a commit. Each commit, a snapshot of my evolving thinking.</p><p>At first, it felt ridiculous. Who tracks their life in version control? But after a month, the git history revealed something profound: <strong>commit frequency is a proxy for personal growth velocity</strong>. The areas where I was thriving had daily commits. The stagnant areas? Their files gathered digital dust with timestamps from months ago.</p><p>But tracking wasn’t enough. I needed a way to accelerate the iteration cycle. That’s when I started experimenting with AI as a growth partner.</p><h3>AI: The Mirror with an Amplifier</h3><p>I started using Claude Code as my weekly personal growth review partner in addition to my daily coding workload. What used to feel like heavy quarterly goal-setting sessions became lightweight Sunday check-ins.</p><p>The more I worked with AI, the more I realized <strong>AI is a mirror with an amplifier.</strong> When I’m clear about what I want, AI helps me move ten times faster. When I’m vague, it reflects that vagueness back, magnified. <strong>If clarity determines output quality in AI, maybe it determines life quality too.</strong></p><p>When I typed vague prompts like “Help me plan my fitness goals,” I got generic workout schedules. But when I got specific — “I want a plan aligned with major personal training methodologies like ACE and NSCA, focusing on strength training while learning anatomy and biomechanics” — suddenly I had a learning curriculum that actually excited me.</p><p>The pattern became clear: <strong>Clarity in equals power out.</strong> The quality of my personal growth became directly proportional to the precision of my thinking and how articulate I could convey it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/593/1*A0Qo3nbVbATaVoWNWFSmSg.png" /><figcaption>Figure 1: personal_training.md generated by Claude Code</figcaption></figure><h3>The Unified Cycle</h3><p>In software, we have a saying: “<strong>Shipping is the heartbeat of a company</strong>.” The development cycle is sacred:</p><h4>Software Development Cycle</h4><ol><li><strong>Discover</strong> → Explore the problem space</li><li><strong>Design</strong> → Define the architecture</li><li><strong>Build</strong> → Write and test code</li><li><strong>Deploy</strong> → Launch to production</li><li><strong>Support</strong> → Debug, refactor, scale</li></ol><p>Every loop teaches us something. Every iteration makes the product better. The faster we complete a cycle, the faster we improve.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ytcdEfKoyOEGEBx1kcuVbA.png" /><figcaption>Figure 2: The software development cycle</figcaption></figure><p>I realized I could apply the same framework to personal growth.</p><h4>Personal Growth Cycle:</h4><ol><li><strong>Discover</strong> → Self-awareness through journaling with AI</li><li><strong>Design</strong> → Goal-setting with clear metrics</li><li><strong>Build</strong> → Testing habits in controlled experiments</li><li><strong>Deploy</strong> → Executing in the real world</li><li><strong>Support</strong> → Reflecting and refining with AI assistance</li></ol><p>I connected Claude to my Linear workspace through MCP (yes, I use Linear for personal project management too). Every Sunday, I’d review my “life sprints” with AI as my pair programmer for personal growth.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ddmD3s88SiKb7GRBf6adqw.png" /><figcaption>Figure 3: The personal growth cycle</figcaption></figure><h3>Shipping Better Versions of Yourself</h3><p>Six months in, I find myself doing less but achieving more. My daily task count dropped while my impact soared.</p><p>The exhaustion was different too. Instead of the mindless grind of checking boxes, I was doing heavy mental work. Each clarified intention, each refined prompt, each judgment call — <strong>these were reps in a cognitive weight room</strong>.</p><p>Today, my personal growth repository has over a hundred commits. Some files change daily. Others stabilize once I’ve found what works. The git diff between who I was last year and who I am today is substantial.</p><p>But the real transformation isn’t in the tools — it’s in the mindset. I now think in iterations. I measure in commits, not quarters. I debug my habits like I debug my code.</p><p>AI didn’t give me this system. It revealed what was always possible: <strong>treating growth as a practice of continuous deployment, where every day is a chance to ship a slightly better version of yourself</strong>.</p><h3>The Heartbeat Continues</h3><p>Iteration is the heartbeat of growth — in software and in life.</p><p>For my next post, I’d love to share <strong>how I use AI agents in my daily coding workflow</strong>. But I wanted to start here, with something universal: the realization that we all have the power to version-control our lives.</p><p>We just need to start committing.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9ccb9f42ace8" width="1" height="1" alt=""><hr><p><a href="https://levelup.gitconnected.com/version-control-for-life-how-ai-changed-the-way-i-grow-9ccb9f42ace8">Version Control for Life: How AI Changed the Way I Grow</a> was originally published in <a href="https://levelup.gitconnected.com/">Level Up Coding</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Evals Are the Product: Building Useful AI Beyond the Benchmark]]></title>
            <link>https://levelup.gitconnected.com/evals-are-the-product-building-useful-ai-beyond-the-benchmark-8f31194d0b03?source=rss-6205971bc42------2</link>
            <guid isPermaLink="false">https://medium.com/p/8f31194d0b03</guid>
            <category><![CDATA[evaluation]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[product-development]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[llmops]]></category>
            <dc:creator><![CDATA[Jian Ruan]]></dc:creator>
            <pubDate>Wed, 11 Jun 2025 15:05:12 GMT</pubDate>
            <atom:updated>2025-06-11T15:05:12.495Z</atom:updated>
            <content:encoded><![CDATA[<p>Why we’re entering AI’s second half — and why evaluation is not just an afterthought, but the foundation of real-world utility.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*8QXHLX1x-dGwN7H7" /><figcaption>Photo by <a href="https://unsplash.com/@franganillo?utm_source=medium&amp;utm_medium=referral">Jorge Franganillo</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h3>The Second Half of AI</h3><p>Recently, I read an interesting article by OpenAI researcher Shunyu Yao titled <a href="https://ysymyth.github.io/The-Second-Half/"><em>The Second Half</em></a><em> </em>[1]. It makes a compelling case that the AI industry is now entering its “second half” — a shift from improving models and methods to building evaluation and benchmark.</p><p>In reinforcement learning, there are three components that drive performance: <strong>algorithm</strong>, <strong>environment</strong>, and <strong>prior knowledge</strong>. Over the past few years, we’ve seen rapid progress in the first two. Researchers have optimized architectures, tuning strategies, and training pipelines to squeeze more capability out of transformer models. At the same time, companies have fed these models massive amounts of data — common crawl, codebases, textbooks, product manuals, and more. That’s given us general-purpose LLMs that can pass bar exams, solve advanced math problems, and write well structured and accurate code.</p><p>However, Shunyu points out a critical insight: there’s a growing disconnect between benchmark performance and real-world utility value. As model intelligence improves, utility does not necessarily scale linearly. Why? Because we’re now bottlenecked not by the model’s capability, but by the <strong>environment</strong> — and in the context of AI applications, that means <strong>evaluations</strong>.</p><h3>What Are We Really Evaluating For?</h3><p>This is where evaluation — “<strong>evals</strong>” — becomes more than a performance check. It becomes the bridge between intelligence and usefulness.</p><p>This idea is echoed in another piece I found insightful: <a href="https://eugeneyan.com/writing/eval-process/"><em>An LLM-as-Judge Won’t Save the Product</em></a> by Eugene Yan [2]. His message is simple but critical: <strong>evals are not a one-time benchmark — they are iterative process</strong>.</p><p>In traditional software, we use test-driven development to catch regressions early, clarify requirements, and evolve our understanding of edge cases. AI systems are no different. The tests — the evals — must evolve with the system.</p><p>You don’t need to get it perfect from the start. In a talk at <a href="https://www.youtube.com/watch?v=4pYzYmSdSH4"><em>LangChain’s Interrupt conference</em></a>, Andrew Ng shared that he often begins evals with a rough, even not very helpful, metric — just something that gives signal [3]. That first draft helps him understand where the model fails. He then improves the evaluation criteria in response to those failures. <strong>That feedback loop is the real magic.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Tk60RWxEPnUWI5WzMVqxBg.png" /><figcaption><em>Figure 1: Applying the scientific method to building AI products [2]</em></figcaption></figure><h3>The X-Bench Approach: Tech-Market Fit for Agents</h3><p>One practical example of this shift is <a href="https://xbench.org/files/xbench_profession_v2.4.pdf">X-Bench</a>, a benchmark developed by VC firm HongShan (formerly Sequoia China) [4]. X-Bench doesn’t evaluate agents on abstract academic tasks. Instead, it focuses on whether AI agents can deliver <strong>economic value</strong> in <strong>professional domains</strong>.</p><p>X-Bench introduces the concept of <strong>Technology-Market Fit</strong>. It’s not just about asking, “Is this model smart?” — it’s asking, “Where does this model actually <em>work</em> in the market?”</p><p>To measure this, the benchmark selects domains like:</p><ul><li><strong>Recruitment</strong>: measures how well an agent can screen resumes, assess candidate fit, or draft interview questions.</li><li><strong>Marketing</strong>: evaluates an agent’s ability to generate targeted ad copy, customize outreach, or conduct competitive research.</li></ul><p>These metrics go beyond general intelligence. They reflect whether an agent can perform concrete tasks <em>within a professional context</em> — tasks that someone would otherwise pay a human to do.</p><p>This kind of <strong>profession-aligned evaluation</strong> is a big step forward. It offers companies, investors, and builders a more actionable way to judge readiness and opportunity. It also reminds us: <strong>performance only matters when it’s tied to outcomes that matter.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VV1zEiDRAuJJpd51LH2qsg.png" /><figcaption>Figure 2: Three stages of Tech-Market-Fit. [5]</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tZoVpoQqpUiUHuvR9zPUjQ.png" /><figcaption>Figure 3: Profession-aligned evaluation define domain agents, predict Tech-Market Fit (TMF) and track competition of agent products. [5]</figcaption></figure><h3>Why Coding Is an Outlier</h3><p>A question often comes up: why are coding agents like Cursor, Windsurf, Devin etc getting so much traction?</p><p>It’s not just that developers are building for themselves. It’s because <strong>coding has a built-in evaluation mechanism</strong>. You write code, and the compiler tells you immediately whether it runs. You write a test, and it either passes or fails. This instant, objective feedback loop makes iteration faster and evaluation easier.</p><p>Unfortunately, most domains aren’t like that. For example:</p><ul><li>In <strong>finance</strong>, how do you judge if an AI co-pilot gave the <em>right</em> advice?</li><li>In <strong>law</strong>, how do you ensure a chatbot stayed compliant with jurisdiction-specific rules?</li><li>In <strong>healthcare</strong>, what counts as a safe and helpful recommendation?</li></ul><p>These industries lack a clear “compiler.” Evaluation becomes a fuzzy area, requiring <strong>subject-matter experts</strong> to weigh in. But involving humans introduces friction — it’s expensive, slow, and hard to scale.</p><p>To get around this, some teams use LLMs to <em>evaluate</em> LLMs. But as Eugene Yan warns, <strong>using a model as a judge doesn’t absolve you from building a good evaluation process</strong>. If your underlying evaluation criteria and process are flawed, you’ll still get garbage-in, garbage-out behavior — even if the LLM judge sounds smart.</p><h3>Case Study: MedHEL and Taxonomy-Driven Evals</h3><p>A great example of structured evaluation comes from Stanford’s <a href="https://crfm.stanford.edu/helm/medhelm/latest/#/leaderboard">MedHEL</a> project, where they present a structured, task-based approach to evaluation for LLMs in healthcare [6].</p><p>Rather than using generic metrics, MedHEL:</p><ul><li>Identifies <strong>real clinical workflows</strong>, like “clinical decision support” and “patient communication”.</li><li>Breaks each workflow into <strong>concrete tasks</strong>, such as “classify medical condition from notes” or “draft patient education messages.”</li><li>Defines <strong>who</strong> the task is for (e.g., clinician), <strong>when</strong> it occurs, <strong>what</strong> data is involved, and <strong>what success looks like</strong>.</li></ul><p>This creates a <strong>taxonomy-driven eval framework</strong>. It makes evaluation repeatable, interpretable, and grounded in professional reality.</p><p>This model could be replicated in other industries. For example, the same structure applied to finance industry:</p><ul><li><strong>Fraud Detection</strong> (transaction classification, anomaly scoring)</li><li><strong>Credit Risk</strong> (cashflow pattern modeling, default prediction)</li><li><strong>Customer Support</strong> (compliance-aware response generation)</li><li><strong>SMB Finance Copilots</strong> (forecasting cashflow, surfacing insights from bank data)</li></ul><p>The benefit? A clear map of what to evaluate, why it matters, and how it connects to business value.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kbroJwMVzDdt_FJ0G1pAlg.png" /><figcaption><em>Figure 4: An overview of the steps in creating MedHELM [6]</em></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Y9-z3oGPcwmjBOv13cnxhQ.png" /><figcaption>Figure 5: A scenario represents a use case and consists of a dataset of instances [6]</figcaption></figure><h3>A Mental Shift: Evals Are the Product</h3><p>If you can’t evaluate your model’s behavior — on the right tasks, for the right users, with meaningful metrics — then you don’t really have a product. You just have a demo.</p><p>To build useful AI systems, we need to:</p><ul><li>Design <strong>use-case–specific</strong> metrics</li><li>Validate outputs with <strong>real-world</strong> feedback</li><li>Improve iteratively alongside the model, UI, and data</li></ul><p>Evals are no longer just afterthought tooling. They’re the lens through which we understand whether our systems <em>actually work</em> for people.</p><h3>Conclusion</h3><p>We’re past the phase where better models alone drive better outcomes. We’re in the second half now — where success depends on whether your AI can <em>deliver value in context</em>.</p><p>You don’t need a perfect eval to get started. Instead</p><ul><li>Start simple</li><li>Iterate fast</li><li>Align with real-world outcomes</li></ul><p>Because in this next chapter of AI, <strong>how you evaluate is what you build.</strong></p><h4>Reference</h4><p>[1] Shunyu Yao, <em>The Second Half</em>, <a href="https://ysymyth.github.io/The-Second-Half/">https://ysymyth.github.io/The-Second-Half/</a></p><p>[2] Eugene Yan, <em>An LLM-as-Judge Won’t Save The Product — Fixing Your Process Will</em>, <a href="https://eugeneyan.com/writing/eval-process/">https://eugeneyan.com/writing/eval-process/</a></p><p>[3] Andrew Ng, <em>State of AI Agents | LangChain Interrupt</em>, <a href="https://www.youtube.com/watch?v=4pYzYmSdSH4">https://www.youtube.com/watch?v=4pYzYmSdSH4</a></p><p>[4] X-Bench Leaderboard, <em>“Evergreen” Benchmark for AI Agents</em>, <a href="https://xbench.org/">https://xbench.org</a></p><p>[5] HongShan (formerly Sequoia China), <em>X-BENCH: Tracking Agents Productivity — Scaling with Profession-Aligned Real-World Evaluations</em>, <a href="https://xbench.org/files/xbench_profession_v2.4.pdf">https://xbench.org/files/xbench_profession_v2.4.pdf</a></p><p>[6] Stanford Center for Research on Foundation Models (CRFM), <em>MedHEL: Holistic Evaluation of Large Language Models for Medical Tasks</em>, <a href="https://crfm.stanford.edu/helm/medhelm/latest/#/leaderboard">https://crfm.stanford.edu/helm/medhelm/latest/#/leaderboard</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8f31194d0b03" width="1" height="1" alt=""><hr><p><a href="https://levelup.gitconnected.com/evals-are-the-product-building-useful-ai-beyond-the-benchmark-8f31194d0b03">Evals Are the Product: Building Useful AI Beyond the Benchmark</a> was originally published in <a href="https://levelup.gitconnected.com/">Level Up Coding</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How I Made the Most of Google’s 5-Day Generative AI Intensive Course on Kaggle]]></title>
            <link>https://levelup.gitconnected.com/how-i-made-the-most-of-googles-5-day-generative-ai-intensive-course-on-kaggle-a858e634f44a?source=rss-6205971bc42------2</link>
            <guid isPermaLink="false">https://medium.com/p/a858e634f44a</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[google]]></category>
            <category><![CDATA[learning]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Jian Ruan]]></dc:creator>
            <pubDate>Fri, 11 Apr 2025 00:25:57 GMT</pubDate>
            <atom:updated>2025-04-11T00:25:57.528Z</atom:updated>
            <content:encoded><![CDATA[<p>Recently, Google launched a <a href="https://rsvp.withgoogle.com/events/google-generative-ai-intensive"><strong>5-Day Generative AI Intensive Course</strong></a> on Kaggle. Whether you’re just getting started with LLMs or already building production-ready applications, this course offers a well-structured, fast-paced dive into the core concepts, tools, and real-world techniques behind modern generative AI.</p><p>Here’s how I did it 👇</p><h4>🎧 Step 1: Passive Listening While Moving</h4><p>Each unit in the course includes a short <strong>podcast</strong> episode, and I found these perfect for passive learning during walks, commutes, or doing chores. It’s kind of funny — we’re literally learning about AI… from AI-generated audio content.</p><h4>📚 Step 2: Interactive Reading in NotebookLM</h4><p>I downloaded all the <strong>whitepapers</strong> and imported them into <strong>NotebookLM</strong> (Google’s AI-powered research assistant). From there, I was able to generate:</p><ul><li><strong>Briefing Docs</strong> to summarize key concepts</li><li><strong>Mindmaps</strong> for visual learners like me (each node is clickable — triggering a deep-dive explanation via prompt!)</li><li><strong>Study Guides</strong> that includes quiz questions and answer keys</li><li><strong>Essay-style prompts</strong> — many of which are solid enough to turn into standalone Medium articles or study group write-ups</li></ul><p>Some examples of essay-style prompts:</p><ul><li><em>Compare inference-time optimization techniques like quantization, distillation, Flash Attention, and speculative decoding.</em></li><li><em>Evaluate the trade-offs between model size, training data size, and computational resources in the development of effective LLMs. Discuss implications of the Chinchilla scaling laws.</em></li></ul><p>✨ Feature wish: It would be great if NotebookLM could render flashcards natively. For now, I have to export the quiz text and load it into my flashcard app manually.</p><h4>👩‍💻 Step 3: Get Hands-On With Codelabs</h4><p>Each day includes Kaggle-hosted <strong>codelabs</strong> — interactive notebooks with zero setup required. I sometimes jumped straight into the codelabs first, then circled back to read the whitepapers for context.</p><p>Codelabs cover everything from prompting and vector search to fine-tuning LLMs and building agents with Gemini.</p><h4>🎥 Step 4: Catch the Livestreams (Optional)</h4><p>Each day also includes a <strong>livestream recording</strong>. These are less critical than the whitepapers or codelabs, but helpful if you’re curious how Google’s engineers frame the material. The Day 5 session, in particular, dives into <strong>Vertex AI and MLOps</strong> — great resource for anyone planning to take models into production.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*W9xXi0QtZrJGJ3sVefhdzw.png" /><figcaption>NotebookLM summary and mindmap.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eFfhJEkSuMAfz5PrfrbWSA.png" /><figcaption>Flashcard generated from Day 1 whitepaper on foundational LLMs.</figcaption></figure><h3>📅 What’s Inside the Course?</h3><h3>Day 1</h3><h4>Intro Unit — Foundational LLMs &amp; Text Generation</h4><ul><li><a href="https://www.youtube.com/watch?v=Na3O4Pkbp-U&amp;list=PLqFaTIg4myu_yKJpvF8WE2JfaG5kGuvoE&amp;index=2">Podcast | Foundational LLMs &amp; Text Generation</a></li><li><a href="https://www.kaggle.com/whitepaper-foundational-llm-and-text-generation">Whitepaper | Foundational Large Language Models &amp; Text Generation</a></li><li><a href="https://www.youtube.com/watch?v=WpIfAeCIFc0&amp;list=PLqFaTIg4myu-lbBTrUpoQQIzZZxvrOaP5&amp;index=1">Live Session | Day 1</a></li></ul><h4>Unit 1 — Prompt Engineering</h4><ul><li><a href="https://www.youtube.com/watch?v=CFtX0ZyLSAY&amp;list=PLqFaTIg4myu_yKJpvF8WE2JfaG5kGuvoE&amp;index=3">Podcast | Prompt Engineering</a></li><li><a href="https://www.kaggle.com/whitepaper-prompt-engineering">Whitepaper | Prompt Engineering</a></li><li><a href="https://www.kaggle.com/code/markishere/day-1-prompting">Codelab | Prompting</a></li><li><a href="https://www.kaggle.com/code/markishere/day-1-evaluation-and-structured-output">Codelab | Evaluation and structured output</a></li></ul><h3>Day 2</h3><h4>Unit 2 — Embeddings and Vector Stores/ Databases</h4><ul><li><a href="https://www.youtube.com/watch?v=xCAVsst6WJ8&amp;list=PLqFaTIg4myu_yKJpvF8WE2JfaG5kGuvoE&amp;index=4">Podcast | Embeddings &amp; Vector Stores</a></li><li><a href="https://www.kaggle.com/whitepaper-embeddings-and-vector-stores">Whitepaper | Embeddings &amp; Vector Stores</a></li><li><a href="https://www.kaggle.com/code/markishere/day-2-document-q-a-with-rag">Codelab | Document Q&amp;A with RAG</a></li><li><a href="https://www.kaggle.com/code/markishere/day-2-embeddings-and-similarity-scores">Codelab | Embeddings and similarity scores</a></li><li><a href="https://www.kaggle.com/code/markishere/day-2-classifying-embeddings-with-keras">Codelab | Classifying embeddings with Keras</a></li><li><a href="https://www.youtube.com/watch?v=AjpjCHdIINU&amp;list=PLqFaTIg4myu-lbBTrUpoQQIzZZxvrOaP5&amp;index=2">Live Session | Day 2</a></li></ul><h3>Day 3</h3><h4>Unit 3a — Generative AI Agents</h4><ul><li><a href="https://www.youtube.com/watch?v=D3Kaqz7VW28&amp;list=PLqFaTIg4myu_yKJpvF8WE2JfaG5kGuvoE&amp;index=4">Podcast | Generative AI Agents</a></li><li><a href="https://www.kaggle.com/whitepaper-agents">Whitepaper | Generative AI Agents</a></li><li><a href="https://www.kaggle.com/code/markishere/day-3-function-calling-with-the-gemini-api">Codelab | Function calling with the Gemini API</a></li><li><a href="https://www.kaggle.com/code/markishere/day-3-building-an-agent-with-langgraph/">Codelab | Building an agent with LangGraph</a></li><li><a href="https://www.youtube.com/watch?v=g6MVIEzFTjY&amp;list=PLqFaTIg4myu-lbBTrUpoQQIzZZxvrOaP5&amp;index=3">Live Session | Day 3</a></li></ul><h4>Advanced 3b — Agents Companion</h4><ul><li><a href="https://www.youtube.com/watch?v=7rbSwt-7odQ&amp;list=PLqFaTIg4myu_yKJpvF8WE2JfaG5kGuvoE&amp;index=6">Podcast | Agents Companion</a></li><li><a href="https://www.kaggle.com/whitepaper-agent-companion">Whitepaper | Agents Companion</a></li></ul><h3>Day 4</h3><h4>Unit 4 — Domain-Specific LLMs</h4><ul><li><a href="https://www.youtube.com/watch?v=MWqspvVvNzA&amp;list=PLqFaTIg4myu_yKJpvF8WE2JfaG5kGuvoE&amp;index=7">Podcast | Solving Domain-Specific Problems Using LLMs</a></li><li><a href="https://www.kaggle.com/whitepaper-solving-domains-specific-problems-using-llms">Whitepaper | Solving Domain-Specific Problems Using LLMs</a></li><li><a href="https://www.kaggle.com/code/markishere/day-4-fine-tuning-a-custom-model">Codelab | Fine tuning a custom model</a></li><li><a href="https://www.kaggle.com/code/markishere/day-4-google-search-grounding">Codelab | Google Search grounding</a></li><li><a href="https://www.youtube.com/watch?v=AN2tpHi26OE&amp;list=PLqFaTIg4myu-lbBTrUpoQQIzZZxvrOaP5&amp;index=4">Live Session | Day 4</a></li></ul><h3>Day 5</h3><h4>Unit 5 — MLOps for Generative AI</h4><ul><li><a href="https://www.youtube.com/watch?v=Hbk8UXavHrk&amp;list=PLqFaTIg4myu_yKJpvF8WE2JfaG5kGuvoE&amp;index=8">Podcast | Operationalizing Generative AI on Vertex AI using MLOps</a></li><li><a href="https://www.kaggle.com/whitepaper-operationalizing-generative-ai-on-vertex-ai-using-mlops">Whitepaper | Operationalizing Generative AI on Vertex AI using MLOps</a></li><li><a href="https://github.com/GoogleCloudPlatform/agent-starter-pack">GitHub Repo | Agent Starter Pack — Production-ready Generative AI Agent templates built for Google Cloud</a></li><li><a href="https://www.youtube.com/watch?v=eZ-8UQ_t4YM&amp;list=PLqFaTIg4myu-lbBTrUpoQQIzZZxvrOaP5&amp;index=5">Live Session | Day 5</a></li></ul><h3>Final Thoughts</h3><p>This course is far more than just a content dump — it’s a <strong>well-crafted, multi-format learning experience</strong>. I walked away with a deeper understanding of GenAI, practical tools, and several threads of inspiration to explore further.</p><p>It’s clear that Google is making a strong push to help developers adopt Gemini and bring generative AI into real-world production workflows. I’d love to see more production-ready templates and implementation patterns in the future.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a858e634f44a" width="1" height="1" alt=""><hr><p><a href="https://levelup.gitconnected.com/how-i-made-the-most-of-googles-5-day-generative-ai-intensive-course-on-kaggle-a858e634f44a">How I Made the Most of Google’s 5-Day Generative AI Intensive Course on Kaggle</a> was originally published in <a href="https://levelup.gitconnected.com/">Level Up Coding</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Beyond the Math: Visual Resources to Help You Build an Intuitive Understanding of LLMs]]></title>
            <link>https://levelup.gitconnected.com/beyond-the-math-visual-resources-to-help-you-build-an-intuitive-understanding-of-llms-397139ddd706?source=rss-6205971bc42------2</link>
            <guid isPermaLink="false">https://medium.com/p/397139ddd706</guid>
            <category><![CDATA[visualization]]></category>
            <category><![CDATA[intuition]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[matrix-multiplication]]></category>
            <dc:creator><![CDATA[Jian Ruan]]></dc:creator>
            <pubDate>Wed, 12 Mar 2025 01:07:36 GMT</pubDate>
            <atom:updated>2025-03-12T01:07:36.284Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*F3qyeN3f0FqrGUuF" /><figcaption>Photo by <a href="https://unsplash.com/@thisisengineering?utm_source=medium&amp;utm_medium=referral">ThisisEngineering</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>For many, understanding how LLMs work isn’t as straightforward as reading through math formulas. While a strong math foundation is valuable, <strong>equations alone often fail to convey the deeper intuition behind these models</strong>. Instead of getting lost in abstract notation, I find that visualizations and interactive tools offer a much clearer way to grasping core ideas before diving into the details.</p><p>The real challenge isn’t just understanding the math — it’s about building an intuitive sense of why these models recognize patterns, generate coherent responses, and even seem to reason. Visual representations can bridge the gap between complex theory and practical understanding — because at the end of the day, it all boils down to <strong>matrix multiplication</strong>.</p><p>Here are some of the best visual resources I’ve found that help build a deep, intuitive understanding of LLMs.</p><h3>3Blue1Brown —Understanding Transformers Through Math-Driven Animations</h3><p>I’ve always been amazed by how Grant Sanderson (aka 3Blue1Brown) turns abstract math concepts into something visually intuitive and elegant. His <a href="https://www.3blue1brown.com/topics/neural-networks">neural network series</a>, particularly the transformer breakdown, is one of the best visual explanations out there [1]. He walks through concepts like gradient descent, back-propagation, and attention mechanisms in a way that just clicks. Plus, his other videos on linear algebra and probability are worth checking out — they provide a strong foundation for understanding the math behind these models.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RROk96L3bQQMR4d-uV0s9Q.png" /><figcaption>Break down of key concepts of LLM. Source: 3Blue1Brown [1].</figcaption></figure><h3>Brendan Bycroft ‘s <em>3D </em>Transformer Visualization</h3><p>Software engineer Brendan Bycroft has created an <a href="https://bbycroft.net/llm">incredible 3D visualization</a> of GPT like transformer models [2]. His interactive tool lets you zoom, rotate, and explore a transformer model, like you’re peeling down a neural network layer by layer.</p><p>What I love about this tool is how it walks you through the entire model architecture step by step. Instead of staring at abstract block diagrams, you can literally explore how attention layers, MLPs, and other components stack together. It gives you a big-picture view while also letting you zoom in on the computation happening inside — hover over each dot in the matrix, and you can see the calculations happening at different stages.</p><p>You can even compare different model sizes, from <a href="https://github.com/karpathy/nanoGPT">nanoGPT</a> to much larger architectures like GPT-3 [3]. Scaling up feels like copy-pasting the same modular building blocks, which reinforces the idea that these models grow by stacking more of the same fundamental units. If you want to build an intuitive, almost Lego-like mental model of transformers, this is a fantastic place to start.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/536/1*3zxI15WS-QuJRY4z07vikw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*a8u5WVU8vCqsNVKrt5S2dQ.png" /><figcaption><em>3D Visualization of GPT-style LLMs. Source: </em>Brendan Bycroft [2].</figcaption></figure><h3>Basil Hosmer’s Matrix Multiplication Visualizer (mm)</h3><p>Matrix multiplication is at the core of LLMs, and Basil Hosmer’s <a href="https://bhosmer.github.io/mm/ref.html"><em>mm</em></a> package makes it easier to see what’s happening under the hood [4]. In his PyTorch blog post, <a href="https://pytorch.org/blog/inside-the-matrix/"><em>Inside the Matrix: Visualizing Matrix Multiplication, Attention and Beyond</em></a>, he describes how <em>mm</em> package helps visualize the entire computational process — starting from simple dot products and building up to matrix-vector multiplications, attention heads in nanoGPT, and even large-scale partitioning techniques used in transformer models [5].</p><p>What makes this tool so inspiring is how it treats <strong>matrix multiplication (</strong>matmuls) <strong>as a fundamentally three-dimensional operation</strong>. Instead of representing matmuls as flat grids, <strong><em>mm</em> folds them into 3D structures, like origami, </strong>making it much easier to see how values interact across different dimensions and the relationships between input and output.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/768/1*6uUh-CEoBgx_2pFTvcJRqA.gif" /><figcaption>3D animation of matrix-vector multiplication. Source: Basil Hosmer [5].</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*U3PSz5QAfPuAktpA.jpg" /><figcaption>A chain of matrix multiplication. Source: Basil Hosmer [5].</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*eK5dEUWS2Tz0WZKw.jpg" /><figcaption>Attention head visualized with 3D matrix multiplication. Source: Basil Hosmer [5].</figcaption></figure><h3>Prof. Tom Yeh — Building Transformers in Excel</h3><p>One of the most surprising yet effective ways I’ve come across to build an intuition for LLMs is <a href="https://www.youtube.com/watch?v=idF6TiTGYsE">Prof. Tom Yeh’s Excel-based DeepSeek breakdown</a> [6]. Instead of explaining transformers through code or equations, he manually constructs matrices inside a spreadsheet, walking through dimensions and computations step by step.</p><p>His video explores how DeepSeek’s V2 and V3 models implement:</p><ul><li><strong>Multi-Head Latent Attention (MLA)</strong> — Using compressed matrices to optimize computation.</li><li><strong>Mixture of Experts (MoE)</strong> — How routers dynamically distribute tokens to different expert layers.</li><li><strong>Computation Flow</strong> — Following matrix transformations through attention heads and expert networks.</li></ul><p>Seeing transformers modeled in Excel makes everything feel far more tangible. If you like hands-on learning and want a new way to think about attention mechanisms, this approach is well worth a look.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2AzJ8CVJloWhlWvFGQ2NeQ.jpeg" /><figcaption>Breakdown of DeepSeek architecture in Excel spreadsheet. Source: Created by author with template provided by Prof. Tom Yeh [6].</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*acUXD4GVBF6BRAvTLLMO7g.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*A0FdbiXkKrVwJ07ea7rd8A.png" /><figcaption>Data flow of matrix multiplication (in blue lines). Source: Prof. Tom Yeh [6].</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1016/1*ZYiLUsERVuSefpGOYbWopw.png" /><figcaption>Illustration of the MoE &amp; MLA architecture of DeepSeek-V2. Source: DeepSeek [7].</figcaption></figure><h3>Other Fun Projects Worth Exploring</h3><h4>Transformer Explainer</h4><p>This <a href="https://poloclub.github.io/transformer-explainer/">waterflow-style</a> chart makes it easy to visualize how text tokens move through embeddings, self-attention, query-key-value (QKV) matrices, MLP layers, and the final probability distribution [8].</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TZruvQ2sI4zrnVSP_ehICA.png" /><figcaption>LLM data flow. Source: Transformer Explainer [8].</figcaption></figure><h4>Machine Learning University (MLU) — Visual explanations of core machine learning concepts</h4><p>Amazon’s <a href="http://mlu-explain.github.io/">Machine Learning University (MLU)</a> offers great interactive visualizations for breaking down key ML concepts. Whether it’s ROC/AUC, cross-validation, or reinforcement learning, their step-by-step approach makes these ideas easy to grasp [9].</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yGwqhzs5lh3sy_-eEwMv7Q.gif" /><figcaption>Neural network animation. Source: Machine Learning University [9].</figcaption></figure><h4><em>MNIST</em> <em>Digit Recognition Visualization</em></h4><p>The MNIST handwritten digit recognition is like the <em>Hello World </em>project of neural networks. This <a href="https://okdalto.github.io/VisualizeMNIST_web/">simple interactive tool</a> lets you draw digits and watch activation patterns light up in real-time [10].</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/768/1*IIr62bK5N_-H6dR1PDW0mw.gif" /><figcaption><em>Visualize handwritten digit recognition</em>. Source: Seonghyeon Kim [10].</figcaption></figure><h4>Interactive Cartoon Neural Networks</h4><p>This is probably the cutest way to introduce neural networks to kids — or anyone new to the concept. With a <a href="http://aegeorge42.github.io/">playful cartoon style</a>, it makes abstract ML ideas like backpropagation, activation functions, and training feel intuitive and fun [11].</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*G1oaBFSUTAvWvgActuiP5A.png" /><figcaption><em>Interactive cartoon neural networks. Source: </em>Allison George [11].</figcaption></figure><h3>Conclusion</h3><p>Understanding the inner workings of LLMs doesn’t have to be a dry exercise in linear algebra and dense research papers. These visual and interactive resources bring complex concepts to life, making it easier to develop an intuitive grasp of how transformers process information, how attention mechanisms work, and how matrix multiplications form the foundation of it all. Whether it’s a 3D model you can rotate, a spreadsheet-based breakdown, or a playful cartoon explainer, each of these tools offers a unique way to see machine learning in action.</p><p>If you know of any other great visualization tools, I’d love to hear about them — there’s always more to explore!</p><h4>Reference</h4><p>[1] Grant Sanderson, 3Blue1Brown, <em>Neural Networks Series.</em> <a href="https://www.3blue1brown.com/topics/neural-networks">https://www.3blue1brown.com/topics/neural-networks</a></p><p>[2] Brendan Bycroft, <em>3D Visualization of GPT-style LLMs</em>. <a href="https://bbycroft.net/llm">https://bbycroft.net/llm</a></p><p>[3] Andrej Karpathy, <em>nanoGPT.</em> GitHub*.* <a href="https://github.com/karpathy/nanoGPT">https://github.com/karpathy/nanoGPT</a></p><p>[4] Basil Hosmer. PyTorch Blog, <em>Inside the Matrix: Visualizing Matrix Multiplication, Attention and Beyond.</em> <a href="https://pytorch.org/blog/inside-the-matrix/">https://pytorch.org/blog/inside-the-matrix/</a></p><p>[5] Basil Hosmer, <em>mm — 3D Matmul (Matrix Multiplication) Visualizer.</em> <a href="https://bhosmer.github.io/mm/ref.html">https://bhosmer.github.io/mm/ref.html</a></p><p>[6] Prof. Tom Yeh, AI by Hand, <em>Special: DeepSeek | CSCI 5722: Computer Vision | Spring 25.</em> <a href="https://www.youtube.com/watch?v=idF6TiTGYsE">https://www.youtube.com/watch?v=idF6TiTGYsE</a></p><p>[7] DeepSeek-V2, <em>A Strong, Economical, and Efficient Mixture-of-Experts Language Model</em>. <a href="https://arxiv.org/pdf/2405.04434">https://arxiv.org/pdf/2405.04434</a></p><p>[8] Polo Club of Data Science, Georgia Tech. <em>Transformer Explained Visually</em>. <a href="https://poloclub.github.io/transformer-explainer/">https://poloclub.github.io/transformer-explainer/</a></p><p>[9] Machine Learning University (MLU). <em>Visual explanations of core machine learning concepts.</em> <a href="https://mlu-explain.github.io/">mlu-explain.github.io</a></p><p>[10] Seonghyeon Kim, <em>MNIST</em> <em>Digit Recognition Visualization.</em> <a href="https://okdalto.github.io/VisualizeMNIST_web/">https://okdalto.github.io/VisualizeMNIST_web/</a></p><p>[11] Allison George. <em>Interactive Cartoon Neural Networks</em>. <a href="https://aegeorge42.github.io/">aegeorge42.github.io</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=397139ddd706" width="1" height="1" alt=""><hr><p><a href="https://levelup.gitconnected.com/beyond-the-math-visual-resources-to-help-you-build-an-intuitive-understanding-of-llms-397139ddd706">Beyond the Math: Visual Resources to Help You Build an Intuitive Understanding of LLMs</a> was originally published in <a href="https://levelup.gitconnected.com/">Level Up Coding</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The “Collapsing” Talent Stack]]></title>
            <link>https://levelup.gitconnected.com/the-collapsing-talent-stack-2fb34b69b54b?source=rss-6205971bc42------2</link>
            <guid isPermaLink="false">https://medium.com/p/2fb34b69b54b</guid>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[future-of-work]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[talent]]></category>
            <category><![CDATA[workplace]]></category>
            <dc:creator><![CDATA[Jian Ruan]]></dc:creator>
            <pubDate>Tue, 21 Jan 2025 21:07:58 GMT</pubDate>
            <atom:updated>2025-01-21T21:07:58.856Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>“AI will collapse the talent stack,”</strong> declared Claire Vo, Chief Product Officer at LaunchDarkly and creator of ChatPRD [1]. This bold statement captures a profound shift that is reshaping the modern workforce. Roles once seen as distinct — product management, engineering, and design — are starting to merge into unified, multifaceted responsibilities, driven by advancements in AI and automation.</p><p>I see this happening firsthand in my own startup. As a CTO, my role has naturally evolved into what I now think of as CPTO (Chief Product &amp; Technology Officer), navigating both the strategic vision of <em>what</em> to build and the technical execution of <em>how</em> to build it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*m_-uGbt3zvklDn7b.jpg" /><figcaption>Visualizing the “collapsed talent stack”. Source: Claire Vo [2].</figcaption></figure><h3>The Controversy Around “Collapsed” Roles</h3><p>This “collapse” of roles has sparked debates in online forums like Reddit [3]. Critics argue that tools like ChatPRD, which automate product management tasks, strip away the human touch and undervalue the depth of expertise required for specialized roles. Others see the idea of one person mastering PM, engineering, and design as overly idealistic — especially in traditional corporate environments, where specialization is often the norm.</p><p>These criticisms aren’t without merit. In structured, hierarchical organizations, specialization is key to operational efficiency. However, in startups, where agility and adaptability are critical, this consolidation of the talent stack is already happening. As markets evolve, individuals with these “collapsed” skill sets — those who can adapt and excel across multiple disciplines — are becoming some of the most valuable contributors, not just in startups but also in larger organizations striving to stay competitive.</p><p>Andrew Ng touches on this evolving dynamic in a recent post, highlighting how advancements in AI are reshaping software development. As coding and prototyping become faster and more efficient, there’s a growing demand for roles focused on defining <em>what</em> to build — like product management and design [4]. In many cases, engineers are stepping into these responsibilities, reflecting the shifting composition of development teams. This transition is often more achievable for technical engineers adding PM skills than for PMs acquiring deep technical expertise. While tools like Cursor and Bolt simplify prototyping, deploying production-ready solutions still requires comprehensive technical knowledge, from understanding data pipelines to managing scalable architectures.</p><h3>Why “Collapsed” Skill Sets Are Attainable — Through Inductive Reasoning</h3><p>Critics often question whether mastering PM, engineering, and design is realistic. I believe it is — and the key lies in the growing use of <strong>inductive reasoning</strong> in learning and problem-solving.</p><p>Many times, when I had no clear idea how to implement a desired functionality, I just asked AI to generate the code for me. At first, I didn’t understand the code, but by asking the AI to explain its logic and refining it through experimentation, I gradually developed a deeper understanding. This process mirrors <strong>inductive reasoning</strong>: starting with an answer and working backward to uncover the “why.” It reminds me of my school days when I’d get stuck on a math problem; I would often look up the answer first and then work backward to understand how it was solved — a much more efficient way of learning than struggling to solve it entirely on my own.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*L0gKwsAjqHisP6XQFZ75KQ.png" /><figcaption>Using Cursor to generate and explain test code — an inductive approach to learning through experimentation. Source: Created by the author.</figcaption></figure><h3>Inductive vs. Deductive Learning</h3><p><strong>Inductive Reasoning</strong>: Start with a specific result and work backward to understand the broader principles.</p><ul><li>Example: Use AI to generate functional code for a task. Analyze the generated code, understand its structure, and infer the underlying concepts or patterns, such as how loops or API calls are handled.</li></ul><p><strong>Deductive Reasoning</strong>: Start with general principles and apply them to derive specific solutions.</p><ul><li>Example: Study coding fundamentals, such as data structures and algorithms, then use that knowledge to write code for specific tasks.</li></ul><p>The key difference lies in the approach:</p><ul><li><strong>Deductive reasoning</strong> is like the structured learning in university computer science programs, where students master theory first before applying it to real-world problems.</li><li><strong>Inductive reasoning</strong>, on the other hand, is often the method of self-taught developers, who begin by building projects, learning patterns through experimentation, and refining their understanding as they go.</li></ul><p>Both approaches have their place, but the rise of AI tools has made inductive reasoning a more accessible and efficient pathway for learning to code and delivering real-world results quickly.</p><h3>CICD Between Inductive and Deductive Reasoning in the AI Era</h3><p>It’s important to recognize that inductive and deductive reasoning are not mutually exclusive — they complement each other in the learning and problem-solving process.</p><p><strong>During Inductive Learning:</strong></p><p><em>Start with Specific Observations</em></p><ul><li>AI generates code for you, and you observe its behavior or outcome (e.g., whether it runs successfully or makes logical sense).</li><li>You then generalize your understanding based on these specific examples.</li></ul><p><em>Form Broader Understanding</em></p><ul><li>By asking the AI to explain the code it generated and verifying its logic, you build a conceptual framework about how the code works.</li><li>For instance, after seeing a few examples of code that manipulate APIs, you might infer general patterns about how APIs are handled in programming.</li></ul><p><strong>During Deductive Learning:</strong></p><p><em>Apply Learned Concepts</em></p><ul><li>Once you understand the AI’s explanations or patterns in the code, you apply this general knowledge to evaluate or modify other pieces of code.</li><li>For instance, after learning about loops or conditionals in one example, you can use that understanding to analyze new code generated by the AI.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tAd3kYBdio2nFk1fghuriQ.png" /><figcaption>Continuous cycle of reasoning: Comparing inductive reasoning (specific to broad) and deductive reasoning (broad to specific) in learning and problem-solving. Source: Created by the author.</figcaption></figure><h3>How LLMs Leverage Inductive Reasoning</h3><p>LLMs primarily function through inductive reasoning. They generate outputs based on patterns and examples in their training data, rather than applying pre-defined logical frameworks. <strong>As humans interact with these tools, we, too, can adopt more inductive approaches: experimenting with generated results, learning from them, and gradually building expertise.</strong></p><p>This process parallels scientific exploration. A compelling example is the Nobel Prize-winning work of Demis Hassabis and his team at DeepMind, which revolutionized protein structure prediction [5]. Traditionally, deductive reasoning — working from established theories — could only take research so far. Their breakthrough came from an inductive approach: generating potential protein structures using AI, then reasoning backward to explain their functionality. This process exemplifies how inductive reasoning can drive innovation, uncovering new insights that would have been difficult to achieve through deduction alone.</p><p>Similarly, leveraging inductive reasoning in AI allows us to move faster, iterate more freely, and learn through experimentation. Whether it’s building software, designing workflows, or tackling complex challenges, this approach empowers us to <strong>explore possibilities first and refine our understanding later</strong>. By embracing this mindset, we can harness the full potential of AI to accelerate discovery and innovation across disciplines.</p><h3>Applying Inductive Reasoning Beyond Product Building</h3><p>This approach isn’t limited to product building. Startups are increasingly leveraging “collapsed” roles across other functions, such as marketing and HR. As discussed in the YC Lightcone podcast [6], companies are using engineers to automate tasks traditionally handled by specialized teams. For example, rather than hiring a Chief Marketing Officer to build a marketing team, some startups rely on engineers to manage user acquisition pipelines, optimize paid advertising, and analyze campaign performance — all through data-driven, automated systems.</p><p><strong>This trend isn’t just about cutting costs; it’s about fundamentally</strong> <strong>rethinking how we solve problems</strong>. By leveraging versatile individuals and AI tools, companies can approach challenges across domains with unprecedented efficiency. As the lines between roles continue to blur, adaptability will be the defining skill of the future workforce.</p><h4>Reference</h4><p>[1] Lenny &amp; Friends Summit 2024, <em>Product Management Is Dead, So What Are We Doing Instead?</em> <a href="https://www.youtube.com/watch?v=93fCvFkY1Lg">https://www.youtube.com/watch?v=93fCvFkY1Lg</a></p><p>[2] Claire Vo, <em>Product Management Is Dead</em>, <a href="https://www.chatprd.ai/blog/product-management-is-dead">https://www.chatprd.ai/blog/product-management-is-dead</a></p><p>[3] Reddit, <em>AI Is Adding to the Stress of Being a PM…</em>, <a href="https://www.reddit.com/r/ProductManagement/comments/1i50tx7/ai_is_adding_to_the_stress_of_being_a_pm/">https://www.reddit.com/r/ProductManagement/comments/1i50tx7/ai_is_adding_to_the_stress_of_being_a_pm/</a></p><p>[4] Andrew Ng Newsletter, <em>Jan 15, 2025 | Deeplearning AI</em>, <a href="https://www.deeplearning.ai/the-batch/issue-284/">https://www.deeplearning.ai/the-batch/issue-284/</a></p><p>[5] Nobel Prize Organization, <em>Nobel Prize in Chemistry 2024: Demis Hassabis and John Jumper | Press Release</em>, <a href="https://www.nobelprize.org/prizes/chemistry/2024/press-release/#:~:text=Demis%20Hassabis%20and%20John%20Jumper,amazing%20capacity%20as%20chemical%20tools">https://www.nobelprize.org/prizes/chemistry/2024/press-release/#:~:text=Demis%20Hassabis%20and%20John%20Jumper,amazing%20capacity%20as%20chemical%20tools</a></p><p>[6] YCombinator, <em>Vertical AI Agents Could Be 10X Bigger Than SaaS</em>, <a href="https://www.youtube.com/watch?v=ASABxNenD_U&amp;list=PLQ-uHSnFig5Ob4XXhgSK26Smb4oRhzFmK&amp;index=4">https://www.youtube.com/watch?v=ASABxNenD_U&amp;list=PLQ-uHSnFig5Ob4XXhgSK26Smb4oRhzFmK&amp;index=4</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2fb34b69b54b" width="1" height="1" alt=""><hr><p><a href="https://levelup.gitconnected.com/the-collapsing-talent-stack-2fb34b69b54b">The “Collapsing” Talent Stack</a> was originally published in <a href="https://levelup.gitconnected.com/">Level Up Coding</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[My 2024 in a Spreadsheet: Chat with an LLM About My Daily Logs]]></title>
            <link>https://levelup.gitconnected.com/my-2024-in-a-spreadsheet-chat-with-an-llm-about-my-daily-logs-0531d69bd080?source=rss-6205971bc42------2</link>
            <guid isPermaLink="false">https://medium.com/p/0531d69bd080</guid>
            <category><![CDATA[personal-development]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[data-visualization]]></category>
            <category><![CDATA[journaling]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Jian Ruan]]></dc:creator>
            <pubDate>Fri, 27 Dec 2024 02:25:30 GMT</pubDate>
            <atom:updated>2024-12-27T18:56:33.881Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*UoA4TljLM4aVvT_M" /><figcaption>Photo by <a href="https://unsplash.com/@mattragland?utm_source=medium&amp;utm_medium=referral">Matt Ragland</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h4>Why I Started Daily Logging</h4><p>Inspiration often comes from the most unexpected places. For me, it started with the movie <em>Ready Player One</em>, particularly the “Museum of Life” — a digital archive of memories created by Halliday, the creator of Oasis. The idea of a personal museum — <strong>a ledger of my own existence</strong> — intrigued me. It inspired me to create my own archive to capture and reflect on the small but meaningful moments of daily life.</p><p>Like many others, the COVID pandemic gave me a rare opportunity to pause and reflect. With life slowing down while I was stuck at home, I decided to give journaling a try. At first, I attempted to write thoughtful reflections in Google Docs, but the unstructured format quickly became overwhelming. Most days, I found myself staring at a blank page, unsure of where to begin. Realizing this approach wasn’t working, I decided to pivot. Instead of crafting detailed prose, I opted for a <strong>simple spreadsheet to log the essentials: what I did each day — straightforward, no frills, and no overthinking</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QcrL8NiPlhgQnxq_8JJrMA.png" /><figcaption>Sample spreadsheet-based logging (mock data). Source: Created by the author.</figcaption></figure><h4>The Power of Logging</h4><p>Switching from traditional journaling to logging was a revelation for me. It wasn’t about crafting eloquent reflections — it was about facing reality. Logging gave me a straightforward and honest account of “<em>what the heck I actually did today</em>.” By focusing on capturing simple, factual entries, I could declutter my thoughts and gain clarity about how I was spending my time.</p><p>To make sense of the data, I created a productivity dashboard based on my daily logs. The dashboard visualizes key metrics, including total records, productivity-related records, and patterns over time. The productivity time series chart remains stable — after all, there are only 24 hours in a day, and I can only log a handful of meaningful entries. <strong>It’s a powerful reminder to focus on what truly matters.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kUXUKEQxkJs5PZKrm-gsZA.png" /><figcaption>Productivity dashboard visualizing key metrics. Source: Created by the author.</figcaption></figure><h4>My Year in Review with Vector Embeddings &amp; LLM</h4><p>In 2024, I logged a total of <strong>3,336 </strong>records, with <strong>1,703 </strong>of them<strong> </strong>specifically related to productivity (excluding routine activities like sleeping or eating). While my dashboard provided a helpful high-level overview of these records, I wanted to dive deeper into the patterns within my life and work. Even more, <strong>I wanted AI to help me summarize and interpret the data in ways that could provide actionable insights.</strong></p><p>Here’s how I analyzed my daily logs:</p><ol><li><strong>Export Data</strong>: Save daily logs in a structured .csv format for easy processing.</li><li><strong>Create Embeddings</strong>: Convert meaningful fields (such as “Name” and “Type”) into vector embeddings to capture semantic relationships.</li><li><strong>Store Embeddings in a Vectorstore</strong>: Enable efficient and scalable retrieval of relevant entries based on semantic similarity.</li><li><strong>Visualize Data</strong>: Apply dimensionality reduction techniques (like PCA) and clustering algorithms (K-Means) to uncover patterns.</li><li><strong>Chat with LLM</strong>: Use a locally hosted LLM (Llama 3.2) to interact with the data and derive meaningful insights.</li></ol><p>I ran Llama 3.2 on my local computer using <a href="https://ollama.com/">Ollama</a>, a tool that simplifies LLM model management. With Ollama, you can start and stop models effortlessly with a simple command, and it automatically downloads all the necessary files. This streamlined setup made it easy to analyze my daily logs while keeping it secure and private.</p><h4>Vector Embedding Visualization</h4><p>Vector embeddings transform each journal entry into a high-dimensional numeric representation, capturing the semantic essence of the text. These embeddings enable advanced analytical techniques, such as clustering and dimensionality reduction, to uncover hidden patterns.</p><p>To visualize the data, I used <strong>PCA</strong> (Principal Component Analysis) to reduce the high-dimensional embeddings to three dimensions. I then applied <strong>K-Means</strong> clustering to group the daily logs into <strong>10 distinct clusters</strong>, with each cluster representing similar themes or activities. The resulting 3D scatterplot revealed intriguing patterns within the data:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/768/1*MAqCPI8kU-FDFkkgl_JoPg.gif" /><figcaption>3D visualization of vector embeddings with clusters. If GIF can’t be loaded in mobile app, please open with web browser. Source: Created by the author.</figcaption></figure><p>The 10 clusters can be broadly categorized into <strong>five main groups: work, entertainment, health, food, and routine</strong>. Within the work category, there are clear subcategories such as strategy, coding, design, and communication, reflecting the diversity of tasks I engage in professionally. Entertainment stands out as a significant cluster — no surprise there! And, of course, food activities make their presence known 😋. In contrast, routine-related data points, such as daily habits, are notably distinct and show little diversity, underscoring their structured and repetitive nature.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OHOYOOVkSU8Dy_uyPbWUEg.png" /><figcaption>Clusters of vector embeddings grouped into work, entertainment, health, food, and routine. Source: Created by the author.</figcaption></figure><p><strong>As I explored the scatterplot, dot by dot, it felt like taking a walk through my year. </strong>Each point was a snapshot of a moment I had lived, a task I had completed. In the work-related clusters, I saw traces of intense coding sprints, brainstorming sessions, and design projects — moments where I was deeply in the zone, pushing ideas forward. The life-related clusters, though, brought a different energy. They reminded me of quieter times — restful nights, delicious meals, and hobbies like gaming or music nights that kept me balanced. It was like watching the rhythm of my life unfold, with each cluster telling its own story about how I spent my time.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*flg7tMGPlS6Ec-G4_MxJTA.png" /><figcaption>Examples of work-related entries within clusters. Source: Created by the author.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xDzwNZ3FzZb73yyqFpszoA.png" /><figcaption>Examples of life-related entries within clusters. Source: Created by the author.</figcaption></figure><h4>Chat with LLM</h4><p>Interacting with an LLM brought a new dimension to my analysis. By leveraging my journal embeddings stored in a vector database, I could ask the AI to uncover patterns, highlight recurring themes, or even inject a bit of humor into my year in review.</p><p><strong>2024 Work Themes</strong></p><p>When I asked the LLM to analyze my work-related entries, it accurately captured key aspects of my daily efforts. It identified that much of my work focused on coding, design, and communication. The AI picked up on specific phrases like “debug” and “reverse engineering,” highlighting my emphasis on problem-solving. It also recognized themes tied to Orbbit, the startup I’m building, and acknowledged the discipline required to stay consistent. <strong>The LLM’s analysis closely mirrored reality, making it a useful tool for meaningful self-reflection.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xkTAcX49pF23I8UYTyMw3w.png" /><figcaption>LLM-generated summary of recurring themes in work-related entries. Source: Created by the author.</figcaption></figure><p><strong>2024 Life Themes</strong></p><p>When analyzing life-related entries, the LLM captured key themes like workouts for health, hobbies, and networking events. While networking technically overlaps with work — since it involves talking to users and gathering insights — the LLM categorized it as a life activity. The insights for life-related entries were less detailed compared to work-related ones, likely because many life activities, like daily routines, are simpler and more repetitive.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hGaHlXAQ3qqU6SxXSK4NZw.png" /><figcaption>LLM-generated summary of recurring themes in life-related entries. Source: Created by the author.</figcaption></figure><p><strong>A Joke Based on My Journal </strong>😄</p><p>For a bit of fun, I asked the LLM to create a joke based on my 2024 journey, and it incorporated references to Polytopia (a game I’ve been playing) and polyhedrons.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*idc_D-eKPPcdCmK5QLgLuQ.png" /><figcaption>LLM-generated joke based on my 2024 journey. Source: Created by the author.</figcaption></figure><h4>Final Thoughts</h4><p>One of the most valuable lessons I’ve learned through logging is that <strong>anxiety often arises from the gap between expectation and reality</strong>. We’re notoriously bad at predicting how long things will take, and life rarely goes exactly as planned. Socrates famously said, “<em>Know yourself,</em>” but how can we truly know ourselves if we forget our own past?</p><p>Daily logging became my tool for self-awareness. Each entry was a small, simple act of <strong>reconciling what I wanted to do with what I actually accomplished</strong>. Over time, patterns began to emerge, and I started to understand my own rhythm. I learned to accept that meaningful progress is rarely immediate — it’s slow, steady, and incremental.</p><p>My spreadsheet evolved into a <strong>time series dataset of my existence</strong>. Day by day, week by week, I could track my growth. And growth, as I’ve learned, is rarely exponential. It’s often flat, subtle, and imperceptible. Yet logging helped me trust the process. On days when progress felt invisible, simply recording what I had done reminded me that small, consistent actions add up over time.</p><p>Consider this:</p><ul><li>Improving by <strong>0.01%</strong> daily for 10 years makes you <strong>1.44</strong> times better.</li><li>At <strong>0.1%</strong> daily, you become <strong>38.4</strong> times better.</li><li>And at <strong>1%</strong> daily, the improvement is an astonishing <strong>5.93 × 10¹⁵</strong>.</li></ul><p>Of course, life isn’t linear. There will always be setbacks, plateaus, and unexpected detours. But the principle remains: consistency compounds. Even the smallest steps, taken consistently, can lead to remarkable outcomes over time.</p><p>As I reflect on my journey, I realize that the true value of this practice lies not in what I’ve already logged, but in <strong>what the next decade of logging could reveal</strong>. The magic is in the slow build, the <strong>compounding effect of discipline and self-awareness</strong>. After all, as the saying goes: <em>wherever you go, there you are</em>.</p><h4>Resources</h4><p>[1] Journal-LLM Analysis, Source Code on GitHub, <a href="https://github.com/Aphanmiz/journal-llm-analysis/tree/main">https://github.com/Aphanmiz/journal-llm-analysis/tree/main</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0531d69bd080" width="1" height="1" alt=""><hr><p><a href="https://levelup.gitconnected.com/my-2024-in-a-spreadsheet-chat-with-an-llm-about-my-daily-logs-0531d69bd080">My 2024 in a Spreadsheet: Chat with an LLM About My Daily Logs</a> was originally published in <a href="https://levelup.gitconnected.com/">Level Up Coding</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Building a Robust Ledger: An Engineer’s Guide to Double Entry Accounting]]></title>
            <link>https://levelup.gitconnected.com/building-a-robust-ledger-an-engineers-guide-to-double-entry-accounting-79804045abb0?source=rss-6205971bc42------2</link>
            <guid isPermaLink="false">https://medium.com/p/79804045abb0</guid>
            <category><![CDATA[accounting]]></category>
            <category><![CDATA[fintech]]></category>
            <category><![CDATA[ledger]]></category>
            <category><![CDATA[database]]></category>
            <category><![CDATA[system-design-concepts]]></category>
            <dc:creator><![CDATA[Jian Ruan]]></dc:creator>
            <pubDate>Tue, 17 Dec 2024 21:08:05 GMT</pubDate>
            <atom:updated>2024-12-17T21:08:05.777Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*oMC_N39ViSQ3aSCT" /><figcaption>Photo by <a href="https://unsplash.com/@kellysikkema?utm_source=medium&amp;utm_medium=referral">Kelly Sikkema</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h3>Introduction</h3><p>As a technical founder in the fintech space, one of the first lessons I learned was how crucial it is to build a robust and scalable ledger. It is the backbone of any system that handles money — whether you’re building a fintech startup, launching a financial service, or a traditional business expanding into embedded finance. Without a reliable ledger that serves as the <strong>single source of truth</strong> for all money movement, managing complex financial transactions can quickly turn into a nightmare.</p><p>However, as an engineer, I first struggled with the core concept in building a robust ledger: <strong>double-entry accounting</strong>. While many tutorials explain accounting scenarios using tables, the concept often didn’t feel intuitive to me. It made logical sense, but I couldn’t quite connect it to the technical systems I was designing.</p><p>Everything finally clicked when I started <strong>visualizing</strong> these accounting principles as <strong>vectors</strong> —<strong> with both magnitude and direction</strong>. This approach gave me a clearer mental model, and in this post, I’ll share how I came to understand these concepts — and how you can too.</p><h3>What Is Double Entry Accounting?</h3><p>Double-entry accounting means that <strong>every transaction affects at least two accounts</strong>, with the <strong>total debits always equal the total credits</strong>. In simple words, for every transaction, you track both where the money came from and how it was used, making sure that money doesn’t just appear or disappear out of nowhere.</p><h3>Understanding Debits and Credits Through the Expanded Accounting Equation</h3><h4>Debits</h4><p>Debits increase the value of <strong>Assets</strong> or <strong>Expenses</strong>.</p><p>Example: When your business receives cash, the cash account (an asset) is debited because its value increases.</p><h4><strong>Credits</strong></h4><p>Credits increase the value of <strong>Liabilities</strong> or <strong>Revenues</strong>.</p><p>Example: If your business takes out a loan, the loan account (a liability) is credited to reflect the additional obligation.</p><p><strong>Assets = Liabilities + Equity</strong> is the standard accounting equation, which can be expanded into a more detailed version as shown in the following chart.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*oW5wmSLiLwsEKU_8JUpECQ.png" /><figcaption>The accounting equation: Assets, liabilities, and equity explained. Source: Created by the author.</figcaption></figure><h3>Debit Normal Accounts &amp; Credit Normal Accounts</h3><h4>Debit Normal Accounts</h4><p>These accounts hold a <strong>debit balance</strong>, meaning their value<em> increases with debits</em>.</p><ul><li><strong>Assets:</strong> cash in a bank account, equipment, or inventory, which represent investments or holdings.</li><li><strong>Expenses:</strong> salaries, rent, or utilities, which represent costs incurred.</li></ul><h4>Credit Normal Accounts</h4><p>These accounts hold a <strong>credit balance</strong>, meaning their value <em>increases with credits</em>.</p><ul><li><strong>Liabilities:</strong> loans payable or accounts payable, which represent obligations or borrowed funds.</li><li><strong>Revenue:</strong> sales or service income, which represent money earned.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*q3BiL_qtoIXQWrtRDgD76A.png" /><figcaption>Normal accounts: Debit (uses of money) and credit (sources of money). Source: Created by the author.</figcaption></figure><h3>Example: Lending Co.</h3><p>Let’s break this down using <strong>Lending Co.</strong> as an example. Lending Co. is a platform that connects investors like Alice with borrowers like Bob. Here’s the scenario:</p><ul><li><strong>Alice (Investor):</strong> She deposits $10,000 into Lending Co., expecting a 4.8% return on her investment ($480).</li><li><strong>Bob (Borrower):</strong> He borrows $5,000 from Lending Co. at 12% interest, meaning he owes $600 in interest.</li></ul><h3>Accounting Ledger — Separating Debits and Credits into Two Columns</h3><p>Here’s how these transactions are recorded in an accounting ledger, where debits and credits are clearly separated into two distinct columns. Each transaction is recorded as a row, ensuring that the<strong> total debits always equal the total credits</strong>. You can determine the balance for each account by adding up the entries in corresponding columns.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TcLZFzp0P0_fgoNiSTLs0Q.png" /><figcaption>Accounting ledger example: Debits and credits separated into two columns for human readability. Source: Created by the author.</figcaption></figure><h3>The Challenge — Translating Accounting Ledgers into Database Logic</h3><p>Now, here’s where things get more complex. While reading an accounting ledger like the above is intuitive for humans, handling balances in a database becomes more challenging. To calculate a balance:</p><ol><li>You first need to first filter all debits and credits for an account.</li><li>Then, based on the account type (debit normal or credit normal), decide whether to increase or decrease the balance. For example:</li></ol><ul><li>For <strong>debit normal accounts</strong>, a debit increases the balance, and a credit decreases it.</li><li>For <strong>credit normal accounts</strong>, it’s the reverse — a credit increases the balance, and a debit decreases it.</li></ul><h4>Why This Can Feel Confusing 🤯</h4><p>This process can feel like juggling two sets of rules. You’re constantly switching logic depending on the account type — when to increase, when to decrease — which can make simple calculations feel unnecessarily complicated.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ewSBv5tuv1HeFvAcTx48tg.png" /><figcaption>Reference table for when to increase or decrease different account types. Source: Created by the author.</figcaption></figure><h3>Database Ledger with amount_signed: A Streamlined Approach to Debits and Credits</h3><p>Let’s explore a simpler and more efficient way to track balance in a ledger — using <strong>signed amounts</strong>. Instead of maintaining separate columns for debits and credits, you assign negative values to debits and positive values to credits. (Of course, you could reverse this<strong> </strong>and assign negatives to credits instead, as I’ll explain later through the lens of vectors — these values simply represent direction.)</p><p>Here’s how the amount_signed version makes life easier:</p><ul><li><strong>Every transaction naturally balances to zero:</strong> Debits and credits offset each other, ensuring accuracy.</li><li><strong>Calculating balances becomes straightforward:</strong> With a single column, you can sum the signed amounts to determine an account’s balance in one step.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0e6CRpBFr9nPhcWCxvitqQ.png" /><figcaption>Database ledger example using amount_signed to streamline tracking of debits and credits. Source: Created by the author.</figcaption></figure><h3>The Challenge: Interpreting Negative Numbers</h3><p>While this system is efficient, interpreting <strong>negative values</strong> in real-life contexts can feel <strong>counterintuitive</strong>. For example, in the table above, Lending Co.’s cash account shows <strong>-$120</strong>, even though the company clearly has <strong>$120</strong> in the bank. <em>What does this negative value mean?</em> It seems weird to think of cash amount in the bank as “negative.” This disconnect can make the concept feel awkward at first glance.</p><h3>Understanding Debits and Credits as Vectors</h3><p>Here’s a helpful way to conceptualize this: debits and credits are not just scalar numbers — they’re <strong>vectors</strong> <strong>with</strong> <strong>both magnitude (the dollar amount) and direction (debit or credit)</strong>.</p><p>For consistency, I assign debits to the negative direction and credits to the positive direction. However, this is arbitrary — you can choose the reverse, as long as you apply it consistently throughout your system.</p><p>In this framework, <strong>-$120</strong> doesn’t mean a loss. It simply indicates <em>$120 in the debit direction</em>. It’s a directional marker, not a value judgment.</p><h3>Converting for User-Friendly Display</h3><p>Okay, but here’s the practical question: how do you show this to end users who don’t want to think about vectors or directions? Well — you adjust the sign based on the account type:</p><h4>For Debit Normal Accounts (e.g., cash):</h4><ul><li>If amount_signed = -$120:<br>Adjust the sign: amount_signed * (-1) = $120<br>Display: <em>&quot;You have $120 in the bank!&quot;</em></li><li>If amount_signed = $120:<br>Adjust the sign: amount_signed * (-1) = -$120<br>Display: <em>&quot;Oops, your bank account is overdrawn by $120!&quot;</em></li></ul><h4>For Credit Normal Accounts (e.g., loans):</h4><ul><li>If amount_signed = -$120:<br>Adjust the sign: amount_signed * (+1) = -$120<br>Display: <em>&quot;You overpaid $120 for your credit card!&quot;</em></li><li>If amount_signed = $120:<br>Adjust the sign: amount_signed * (+1) = $120<br>Display: <em>&quot;You still owe $120 on your credit card!&quot;</em></li></ul><h4>Why This Works</h4><p>By viewing debits and credits as vectors, this approach simplifies calculations while maintaining accuracy. Each transaction balances to zero, with signs adjusted to provide clear and intuitive displays for end users.</p><p>At first, seeing something like <strong>-$120</strong> in a cash account may feel odd, but once you grasp the concept of accounting directions, it becomes intuitive. Plus, no more tedious mental gymnastics trying to figure out when to increase or decrease balances!</p><h3>Visualizing Debit and Credit as Vectors: The Lending Co. Story</h3><p>To make accounting concepts more intuitive, let’s visualize how debits and credits behave as vectors within Lending Co.’s example. In the following visualizations, each account is depicted as a plane, where the <strong>vertical axis</strong> indicates <strong>direction</strong> — <strong>credit (+) </strong>or <strong>debit (-)</strong> — and the <strong>horizontal axis</strong> represents <strong>time</strong>. Every entry is a vector, combining <strong>magnitude</strong> (the amount) with <strong>direction</strong> (credit or debit).</p><h4>Transaction #1: Alice Invests $10,000</h4><p>Alice invests $10,000 into Lending Co., expecting a return on her investment.</p><ul><li><strong>Lending Cash (Debit Normal Account):</strong><br>A debit vector of $10,000 moves downward, representing an increase in Lending Co.’s cash.</li><li><strong>Investor Principal (Credit Normal Account):</strong><br>A credit vector of $10,000 moves upward, reflecting the liability Lending Co. owes to Alice.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hTsuy-BZiZt6O8DBrxbXvQ.png" /><figcaption>Transaction #1 vector visualization. Source: Created by the author.</figcaption></figure><h4>Transaction #2: Bob Borrows $5,000</h4><p>Bob then takes out a loan for <strong>$5,000</strong>.</p><ul><li><strong>Lending Cash (Debit Normal Account):</strong><br>A credit vector of $5,000 moves upward, decreasing Lending Co.’s cash balance.</li><li><strong>Borrower Principal (Debit Normal Account):</strong><br>A debit vector of $5,000 moves downward, recording the loan balance Bob owes to Lending Co.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mQCd0GK9sX6V2WUFCVLSQQ.png" /><figcaption>Transaction #2 vector visualization. Source: Created by the author.</figcaption></figure><h4>Transaction #3: Lending Co. Accrues Interest from Bob</h4><p>Now, the loan starts generating income. Bob owes <strong>$600 </strong>in interest, and Lending Co. keeps <strong>$120</strong> of that as a fee and paying the rest <strong>$480</strong> interest to investor Alice.</p><ul><li><strong>Borrower Interest (Debit Normal Account):</strong><br>A debit vector of $600 moves downward, reflecting the interest Bob owes to Lending Co.</li><li><strong>Investor Interest (Credit Normal Account):</strong><br>A credit vector of $480 moves upward, representing the portion of interest Lending Co. owes to Alice.</li><li><strong>Lending Fee Revenue (Credit Normal Account):</strong><br>A credit vector of $120 moves upward, representing Lending Co.’s earnings from the transaction.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gSyizXG7Eo1NXeZoJC8S0A.png" /><figcaption>Transaction #3 vector visualization. Source: Created by the author.</figcaption></figure><h4>Transaction #4: Bob Pays Back Principal + Interest</h4><p>Bob repays the $5,000 principal and $600 interest, clearing his obligations.</p><ul><li><strong>Lending Cash (Debit Normal Account):</strong><br>A debit vector of $5,600 moves downward, reflecting the payment Lending Co. receives.</li><li><strong>Borrower Principal (Debit Normal Account):</strong><br>A credit vector of $5,000 moves upward, clearing Bob’s loan balance.</li><li><strong>Borrower Interest (Debit Normal Account):</strong><br>A credit vector of $600 moves upward, clearing Bob’s interest obligations.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KtKvDtwxCiIJxwE5jWV2Zw.png" /><figcaption>Transaction #4 vector visualization. Source: Created by the author.</figcaption></figure><h4>Transaction #5: Lending Co. Pays Back Alice</h4><p>Finally, Lending Co. pays Alice her $10,000 principal and $480 in interest.</p><ul><li><strong>Lending Cash (Debit Normal Account):</strong><br>A credit vector of $10,480 moves upward, decreasing Lending Co.’s cash balance.</li><li><strong>Investor Principal (Credit Normal Account):</strong><br>A debit vector of $10,000 moves downward, reflecting the repayment of Alice’s investment.</li><li><strong>Investor Interest (Credit Normal Account):</strong><br>A debit vector of $480 moves downward, representing the payment of interest to Alice.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eOMl0DuB2qYI99WnyuYNBw.png" /><figcaption>Transaction #5 vector visualization. Source: Created by the author.</figcaption></figure><p>After completing all the transactions, Lending Co. has <strong>$120</strong> remaining in the bank, representing the revenue earned from lending fees.</p><p>It’s important to note that in this example, <strong>Transaction #3: Lending Co. Accrues Interest from Bob</strong> is recorded <em>before</em> Bob repays Lending Co. This approach reflects the use of <strong>accrual accounting</strong> rather than <strong>cash accounting</strong>, where revenue and expenses are recognized when they are earned or incurred, not when cash is exchanged.</p><h3>Key Components of a Ledger</h3><p>To better understand how the double-entry accounting principles are applied, let’s break down the key components of a ledger.</p><h4><strong>Ledger</strong></h4><p>A ledger is an <strong>immutable</strong> record of all financial transactions, ensuring accuracy and transparency.</p><h4>Accounts</h4><p>Accounts act as “buckets” of value to track specific balances, such as cash in the bank (a debit normal account) or investor principal (a credit normal account).</p><h4>Transactions</h4><p>Transactions are financial events (e.g., loans, repayments) that generate <strong>at least two entries</strong> — a debit and a credit — to maintain zero-sum balance.</p><h4><strong>Entries</strong></h4><p>Entries are the individual components of a transaction, categorized as either debit or credit.</p><h4><strong>Balances</strong></h4><p>Each entry directly impacts the balance of the associated account.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KTLo7KOOb64YBJTsubfW7A.png" /><figcaption>Key components of a ledger. Source: Created by the author.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aSl90bIfmtNT6tN7YOBkvw.png" /><figcaption>Relationship between key components of a ledger. Source: Created by the author.</figcaption></figure><h3>How Leading Tech Companies Approach Ledger Design</h3><p>To gain a deeper understanding of ledger system design, I delved into technical blogs from companies like Uber, Square, Airbnb, and Stripe — leaders in moving money at scale. While their ledger implementations differ, they all adhere to the fundamental principles of double-entry accounting. Here are the key takeaways:</p><p><strong>Uber (Gulfstream Payment Platform): Scalability [1]</strong><br>Uber handles millions of transactions daily across business lines such as ridesharing and food delivery. What stands out is their ability to operate two ledger systems simultaneously — a legacy system and a modern one — during migration, ensuring data consistency without downtime.</p><p><strong>Square (“Books” Ledger): Immutability [2]</strong><br>Square transitioned from single-entry to double-entry accounting to enhance accuracy and scalability — a move that enables greater product flexibility, such as offering SMB loans. (It’s hard to imagine how Square’s support team troubleshot money issues in the early days without double-entry accounting!)</p><p><strong>Airbnb (Event-Based Ledger): Decoupling [3]</strong><br>Airbnb transitioned from a MySQL-based data pipeline to an event-driven approach using Apache Spark, significantly speeding up financial reporting. They decoupled financial logic from product logic by separating platform events (product interactions) from payment events (money movement). Both types of events generate accounting entries that ultimately impact the ledger, creating a more efficient and scalable system.</p><p><strong>Stripe (Global “Ledger” System): Tracking and Tracing [4]</strong><br>As Stripe’s primary offering is its payment API, it must seamlessly handle billions of transactions at any given time. To achieve this, Stripe places a strong emphasis on tracking and tracing money movement. They use state-machine modeling to manage fund flows and have developed a data quality dashboard to ensure timely data, simplify troubleshooting, and uphold high data integrity.</p><h3>Conclusion</h3><p>As a16z partner Angela Strange famously said, <em>“Every company will become a fintech company” </em>[5]<em>.</em> In today’s world, even businesses without prior financial expertise can leverage fintech and embedded finance to expand their offerings. However, at the core of any financial operation lies a robust ledger system, and building it correctly from the start is crucial.</p><p>For developers and founders working on ledger systems, here’s some key advice to keep in mind:</p><ul><li><strong>Prioritize Ledger Systems Early:</strong> Plan and implement ledger systems from the start. While they might seem secondary initially, ignoring them can lead to complex technical and operational issues later.</li><li><strong>Adopt Double-Entry Accounting from the Start:</strong> Use double-entry accounting early to avoid the pain of migrating from single-entry systems. It ensures balanced transactions and establishes a solid foundation for growth.</li><li><strong>Focus on Immutability:</strong> Design your ledger as an immutable, single source of truth. This simplifies audits, increases data reliability, and ensures consistency as your business scales.</li></ul><p>I hope this article has made the concepts of double-entry accounting clearer, especially through the lens of vectors — balancing magnitude and direction. A well-designed ledger is more than just a back-office necessity; it’s a foundation for building scalability, trust, and long-term growth.</p><p>Thank you for reading, and happy ledgering!</p><h4><strong>Reference</strong></h4><p>[1] Uber, Revolutionizing Money Movements at Scale with Strong Data Consistency, <a href="https://www.uber.com/blog/money-scale-strong-data/">https://www.uber.com/blog/money-scale-strong-data/</a></p><p>[2] Square, Books: An Immutable Double-Entry Accounting Database Service, <a href="https://developer.squareup.com/blog/books-an-immutable-double-entry-accounting-database-service/">https://developer.squareup.com/blog/books-an-immutable-double-entry-accounting-database-service/</a></p><p>[3] Airbnb, Tracking the Money — Scaling Financial Reporting at Airbnb, <a href="https://medium.com/airbnb-engineering/tracking-the-money-scaling-financial-reporting-at-airbnb-6d742b80f040">https://medium.com/airbnb-engineering/tracking-the-money-scaling-financial-reporting-at-airbnb-6d742b80f040</a></p><p>[4] Stripe, Ledger: Stripe’s System for Tracking and Validating Money Movement, <a href="https://stripe.com/blog/ledger-stripe-system-for-tracking-and-validating-money-movement">https://stripe.com/blog/ledger-stripe-system-for-tracking-and-validating-money-movement</a></p><p>[5] A16z, Every Company Will Become a Fintech Company, <a href="https://a16z.com/every-company-will-be-a-fintech-company/">https://a16z.com/every-company-will-be-a-fintech-company/</a></p><h4>Additional Resources</h4><p>[6] Hacker News, Discussion on Double Entry Accounting for Developers, <a href="https://news.ycombinator.com/item?id=23964513">https://news.ycombinator.com/item?id=23964513</a></p><p>[7] Anvil, Double Entry Accounting for Engineers, <a href="https://anvil.works/blog/double-entry-accounting-for-engineers">https://anvil.works/blog/double-entry-accounting-for-engineers</a></p><p>[8] Beancount Documentation, Double-Entry Bookkeeping, <a href="https://beancount.github.io/docs/the_double_entry_counting_method.html#double-entry-bookkeeping">https://beancount.github.io/docs/the_double_entry_counting_method.html#double-entry-bookkeeping</a><br><br>[9] Ledger-CLI, Double-Entry Accounting Software, <a href="https://ledger-cli.org/">https://ledger-cli.org/</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=79804045abb0" width="1" height="1" alt=""><hr><p><a href="https://levelup.gitconnected.com/building-a-robust-ledger-an-engineers-guide-to-double-entry-accounting-79804045abb0">Building a Robust Ledger: An Engineer’s Guide to Double Entry Accounting</a> was originally published in <a href="https://levelup.gitconnected.com/">Level Up Coding</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How I Used AI and the DiliTree Approach to Master Multiple Skills as a Startup CTO]]></title>
            <link>https://levelup.gitconnected.com/how-i-used-ai-and-the-dilitree-approach-to-master-multiple-skills-as-a-startup-cto-eeb7b7b437d5?source=rss-6205971bc42------2</link>
            <guid isPermaLink="false">https://medium.com/p/eeb7b7b437d5</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[skills]]></category>
            <category><![CDATA[learning]]></category>
            <category><![CDATA[startup]]></category>
            <dc:creator><![CDATA[Jian Ruan]]></dc:creator>
            <pubDate>Tue, 17 Dec 2024 21:07:45 GMT</pubDate>
            <atom:updated>2024-12-17T21:07:45.036Z</atom:updated>
            <content:encoded><![CDATA[<h3>Building Skills at Scale: How AI and DiliTree Transformed My Learning as Startup CTO</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*1E6_-KklR1_-nw4m" /><figcaption>Photo by <a href="https://unsplash.com/@cgower?utm_source=medium&amp;utm_medium=referral">Christopher Gower</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h3>Introduction: A Journey of Building, Pivoting, and Iterating</h3><p>In June 2024, I co-founded <a href="https://www.orbbit.co/"><strong>Orbbit</strong></a><strong> </strong>with a mission to solve key challenges in the real estate industry. Over time, our product evolved — starting as a networking platform, pivoting to an AI-powered financial copilot, and eventually becoming a flexible B2B payments solution.</p><p>As CTO, I was at the heart of every iteration, navigating the need to quickly design, build, and launch new solutions. Each pivot required me to learn new technical and strategic skills on the fly — often in areas where I had little prior experience.</p><p>To meet these challenges, I developed what I call the <strong>DiliTree (Due Diligence Tree) Approach</strong>, a structured learning method enhanced with AI. This approach enabled me to break down complex topics, master new skills quickly, and grow into what I call a<strong> “pillar-shaped” leader</strong>: someone with not just broad knowledge but also deep expertise in multiple areas— an essential quality for navigating the fast-paced and ever-changing world of startups.</p><h3>My Background</h3><p>I studied Statistics in college, which often surprises people. Why not Computer Science, the go-to major for tech jobs? My reasoning was simple: coding skills were easier to pick up through hands-on projects, but math, statistics, and algorithms were much harder to master independently. That decision has paid off. That decision has paid off — understanding data and the math behind machine learning has become foundational in today’s AI-driven world.</p><p>Early in my career, I dabbled in full-stack development but quickly realized it wasn’t my strength. Instead, I gravitated toward data science, drawn by its combination of analytical rigor and problem-solving. However, founding a startup introduced an entirely new challenge: the need to wear multiple hats. I found myself diving into everything — frontend, backend, AI/ML/data, and even areas like design — whether I liked it or not.</p><h3>The DiliTree Approach: Learning Almost Everything</h3><p>As co-founder and CTO, the sheer breadth of responsibilities required me to learn and adapt quickly to new challenges without getting overwhelmed. That’s where AI became a game-changer — from versatile LLMs to specialized tools like <a href="https://www.cursor.com/">Cursor</a> and <a href="https://v0.dev/">V0.dev</a>, there seems to be an AI tool (or soon will be) for almost everything.</p><p>There are countless tutorials on how to use AI tools, most of which focus on technical aspects like commands or prompts. However, they often miss a key point: the value you get from AI depends on the quality of your own thinking. If you ask shallow questions, you’ll receive shallow answers. At its core, <strong>AI is simply generating the next token</strong> — whether text, image, or code — based on the highest probability.</p><p>As François Chollet, a renowned AI researcher, aptly said, <em>“Intelligence lies in broad or general-purpose abilities; it is marked by skill-acquisition and generalization, rather than skill itself”</em> [1]. This idea inspired me to create the DiliTree Approach — a framework for learning almost anything systematically. Rather than jumping straight into the specifics of “<strong>how to do X</strong>,” I start by building a clear mental map of the overall structure: <strong>what are the key things I need to learn</strong>?</p><h4>Here’s How It Works</h4><p><strong>Step 1: Start with a Framework</strong>: Begin by creating a high-level map of the topic. Identify the key concepts and organize them in a way that makes sense to you. Use any tool that works for you — whether it’s a mind map, bullet points, or platforms like Obsidian or Notion.</p><p><strong>Step 2: Ask Specific Questions</strong>: Use the framework to guide targeted questions about areas you don’t fully understand. AI tools can be particularly helpful here, enabling you to quickly fill in knowledge gaps and focus on practical applications.</p><p><strong>Step 3: Apply and Iterate</strong>: Strengthen your understanding by applying what you’ve learned to a real-world project or task. As you work, refine your framework by adding notes or updating it with new insights. Each iteration deepens your knowledge and helps you build a more comprehensive understanding of the topic.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ynOtyp7GP2Yza9OZKO8fRw.png" /><figcaption>DiliTree example — Next.js framework for frontend development. Source: Created by the author.</figcaption></figure><h4>Example: Using the DiliTree to Learn Next.js for Frontend Development</h4><p><strong>Step 1: Start with a Framework: </strong>To learn Next.js, I began by copying the table of contents from its official documentation into my DiliTree. This included topics like “Routing,” “Data Fetching,” and “Middleware.” Organizing these concepts into a clear structure allowed me to see how they fit together and identify the areas I needed to focus on first.</p><p><strong>Step 2: Ask Specific Questions:</strong> As I started building features, I used AI tools to answer specific questions. For instance, while working with the new App Router, I asked, “How does data fetching work in a server component?” or “What’s the difference between static and dynamic routing in the App Router?” These targeted queries helped me address knowledge gaps and apply the concepts more effectively.</p><p><strong>Step 3: Apply and Iterate:</strong> I then applied what I learned to a real project — building a frontend app with dynamic routes and server-side rendering. As challenges arose, I revisited my DiliTree to refine it, adding notes on best practices and insights gained through hands-on experience. Over time, this framework evolved, capturing both foundational concepts and practical nuances that enhanced my understanding.</p><h3>What About Design? Learning Through DiliTrees</h3><p>Design isn’t the first thing that comes to mind for a CTO’s role, but at a startup, you often have to do it all. I’m far from an artist (stick figures are my limit), but I knew that strong UI/UX design was critical for our product’s success.</p><p>To tackle this challenge, I applied the DiliTree approach to break design down into manageable components. I focused on key principles like typography, color theory, layout consistency, and user flow. Using AI tools like <a href="https://v0.dev/">V0</a>, I translated these concepts into functional, visually appealing designs. It wasn’t about innate design talent — it was about <strong>following a structured process that made the task approachable and achievable</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eMg9T9oitoYl-XSboqxifA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/574/1*r9wd5ZLEDdaguP-W31bSWw.png" /><figcaption>Left: DiliTree for UI design. Right: UI generated by V0 based on the design framework. Source: Created by the author.</figcaption></figure><h3>From T-Shaped to Pillar-Shaped: Building Depth Across Domains</h3><p>You’ve probably heard the advice to become a <strong>T-shaped person</strong> — someone with deep expertise in one area and broad knowledge across others. While this has been solid guidance for years, the rise of AI tools has shifted the paradigm.</p><p>With AI, it’s now feasible to go beyond the T-shape and become <strong>pillar-shaped</strong>: developing multiple areas of deep expertise faster and more efficiently than ever before. Using the DiliTree approach combined with AI tools, I’ve rapidly mastered a growing range of technical topics — system design, full-stack development, security, DevOps, AI/ML/data, and much more. But the approach isn’t limited to technical skills. I’ve also applied it to analyzing business models, structuring product strategies, and exploring entirely new domains.</p><p>Over time, I’ve created hundreds of DiliTrees, each tailored to a specific topic or skill, enabling me to learn faster and execute more efficiently. This approach leverages AI’s strengths — synthesizing vast information and providing instant feedback — while complementing our human ability to grasp concepts from just a few examples, without relying on massive datasets.</p><p>In a world where repetitive tasks are increasingly automated, <strong>relying on a single skillset is no longer enough</strong>. What truly sets you apart is your ability to combine diverse skills creatively and, more importantly, to acquire new ones quickly. AI accelerates this process by simplifying complex concepts, removing barriers to entry, and making it easier to thrive in new fields.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/820/1*mzOMaiHiMdQqdk6xmM1M0w.png" /><figcaption>Comparison: T-shaped vs. pillar-shaped skillsets. Source: Created by the author.</figcaption></figure><h3>Closing Thoughts: It’s Your Time to Scale Up</h3><p>With AI tools and agents at our disposal, <strong>individuals now have more power than ever to build, innovate, and create value at scale</strong>. By developing additional pillars of expertise, you can expand your capabilities and tackle challenges that once felt out of reach.</p><p>I’m just starting my writing journey and plan to share more lessons from my startup experience, along with technical insights on AI, machine learning, and data. Thanks for reading, and feel free to share your thoughts or questions in the comments!</p><h4>Reference</h4><p>[1] F. Chollet, <em>ARC: A self-contained benchmark for intelligence and generalization</em>, <a href="https://arcprize.org/arc">https://arcprize.org/arc</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=eeb7b7b437d5" width="1" height="1" alt=""><hr><p><a href="https://levelup.gitconnected.com/how-i-used-ai-and-the-dilitree-approach-to-master-multiple-skills-as-a-startup-cto-eeb7b7b437d5">How I Used AI and the DiliTree Approach to Master Multiple Skills as a Startup CTO</a> was originally published in <a href="https://levelup.gitconnected.com/">Level Up Coding</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>