Paweł Huryn’s Post

34% of U.S. adults use LLMs daily. But very few know how to write good prompts. Great share by: Paweł Huryn Original post below: ⬇️ ⬇️ ⬇️ 34% of U.S. adults use LLMs daily. But very few know how to write good prompts. I just found a free, interactive AI prompting course by Anthropic. It’s the best material I’ve seen so far. The course is broken up into 9 chapters with step-by-step exercises. Each lesson has a “playground” where you can experiment with examples and see how changing prompts changes the model’s response. After completing this course, you will be able to: • Understand the structure of a great prompt • Recognize common failure modes • Learn the 80/20 techniques to address them • Understand LLMs strengths and weaknesses • Build a strong prompts library • Automate your work like never before 𝐇𝐞𝐫𝐞'𝐬 𝐡𝐨𝐰 𝐭𝐨 𝐚𝐜𝐜𝐞𝐬𝐬 𝐢𝐭 𝐟𝐨𝐫 𝐟𝐫𝐞𝐞: Step 1: Make a copy of this Google Sheet: https://lnkd.in/dCJRdSME Step 2: Install the Claude for Sheets extension: https://lnkd.in/da6aj-mR Step 3: Enable use of the Claude for Sheets extension on your document Step 4: Add your API key (the course uses the smallest, fastest, and cheapest model) You’ll find more detailed instructions in the Google Sheet. Hope that helps! --- P.S. Interested in AI PM? Good prompts help you tackle one of the three challenges in LLM pipeline development: the Gulf of Specification. But finding the perfect prompt is rare. You need to log traces, analyze errors, measure performance, and improve in iterations. To dive deeper, I recommend the AI Evals For Engineers & PMs course. I’m participating in the current cohort with 400+ AI engineers and PMs. The next cohort starts on July 21.  I got a special $800 discount for my community: https://bit.ly/aievals35 _____________________ DM us to be featured on our next post. Master the language of AI. One prompt at a time. How to Talk to AI

Prompt Engineers are in high demand — and these 4 methods are the reason. Master them to stand out in the AI job market. #1 — RAG (Retrieval Augmented Generation) Don’t rely on guesswork — feed the AI with your trusted knowledge base. That’s RAG. #2 — Chain of Thought (CoT) Guide the model to reason step-by-step — like explaining something to an 8-year-old. #3 — ReAct (Reason + Act) Need outside info? ReAct lets the AI think, fetch, and combine data from multiple sources. 4 — DSP (Directional Stimulus Prompting) Drop a keyword hint like ‘software’ or ‘consulting’ to get focused answers fast. https://lnkd.in/gD9f-Aiv Learn these skills by doing. Enroll in AI courses at www.eduarn.com — and build your future.

This would usually take weeks to build — but thanks to Lovable, I developed it in less than 3 hours. Introducing AI Calendar for Duetoday — a Notion-style calendar designed specifically for students. During my uni days, I struggled to manage assignment deadlines, exams, and study sessions. So this weekend, I built the feature I wish I had back then. Unlike typical planners, Duetoday’s calendar comes with an AI Assistant built right in. Just ask: - “When’s my next exam?” - “Create a study block for calculus,” and it handles the rest. This is a free side feature for all Duetoday users as we continue building a true StudyOS — an intelligent studyspace designed to help every learner stay organized, learn faster, and achieve more.

🚀 Day 2: Understanding Convergence in Machine Learning Algorithms Excited to share insights from my Day 2 presentation – "Convergence in ML Algorithms" — a deep dive into how optimization processes stabilize during model training and why it’s a critical aspect of achieving accurate, efficient, and reliable AI systems. 🔍 What is Convergence? Convergence refers to the point during training when further iterations no longer significantly improve model performance. It marks stability in optimization — ensuring the model has learned efficiently without overfitting or underfitting. ⚙️ Core Concepts Covered Convergence Essentials – Explained the importance of convergence, and the difference between premature (local minima) and global (true minima) convergence, highlighting their impact on model performance. Core Optimizers – Discussed the gradient descent family (Batch, Stochastic, and Mini-Batch), and how adaptive methods like Adam and RMSprop accelerate and stabilize convergence. Convergence Proofs – Covered theoretical and empirical guarantees in convex and non-convex optimization problems, including techniques for escaping saddle points. Acceleration & Control – Explored how Momentum, Nesterov Accelerated Gradient, and Learning Rate Scheduling (Step, Exponential, Cosine Annealing) improve convergence speed and stability. Case Studies – • ResNet-50 on ImageNet: Showed how SGD with momentum and cosine scheduling achieves 76% top-1 accuracy in 90 epochs. • TD(0) in Reinforcement Learning: Demonstrated faster convergence compared to Monte Carlo methods, thanks to its iterative update mechanism. 💡 Key Takeaways Choose optimizers wisely: SGD for large, clean datasets; Adam for noisy or sparse gradients. Apply early stopping and adaptive learning rates to ensure efficient convergence. Proper tuning of hyperparameters is essential for stability and generalization. 🎯 Final Thoughts Convergence isn’t just a mathematical concept — it’s the heart of model training efficiency. Understanding it allows us to design smarter, faster, and more reliable machine learning systems.

What if learning to reason requires unlearning first? We often imagine fine-tuning as a straight climb: add more data, get better results. But when we studied how small models learn to reason through code distillation, the curve bent the other way. We found what we call a valley of reasoning. As we scaled reasoning data from 1K → 10K → 30K examples, model performance on competitive coding first dropped by half before climbing back to surpass baseline by over 100%! Reasoning doesn’t emerge linearly. Small models need to unlearn surface-level patterning before internalizing structured reasoning. Understanding that valley might be key to designing data curricula that help small models think, not just copy. At Collinear, we’re studying these post-training dynamics to build smarter improvement loops, where every dataset, reward, and evaluation helps models climb out of their own valley. Paper accepted at NeurIPS ‘25 DL4C and linked in comments. #AI #PostTraining #AIResearch #LLM #Reasoning

I’m not a topper. Not a genius. Just a normal guy who found the shortest path to DSA. I used to overthink — “How do toppers learn DSA so fast?” I thought I wasn’t smart enough. But after talking for hours with AI (even that’s funny ), I found something interesting : DSA is not a course. It’s a pattern — and once you see the pattern, you can’t Unser it. That changed everything. ⸻ Here’s my 1-week plan that made me finally “get it”: Focused session 1 — Sum & Average Understand loops. Dry run with pen & paper. Focused session 2 — Max / Min One pass. Track values. Dry run edge cases. Focused session 3 — Reverse & Two Pointers Understand how swapping logic actually moves data. Day 4 — Prefix Sums & Subarray Basics The secret bridge to advanced patterns. Focused session 5 — Variations Second largest, rotate array, remove duplicates. Focused session 6 — Mini Test 4 small problems. Dry run + code. Focused session 7 — Teach Someone (or AI) If you can explain it, you’ve learned it. ⸻ My Rule: 30 minutes called as Focused session. No distractions. Explain → Dry Run → Code → Twist. That’s it. No fancy course. No 10-hour grinds. Just small wins — every day. ⸻ When you study like this, something magical happens — You stop memorizing, and start seeing patterns. That’s when DSA finally “clicks.” 🔥 #100DaysOfCode #DSA #LearningJourney #Arrays #StudyPlan #CodingMindset

Day 2 - #MLTuesdays📊 What Makes a Machine Learn? When I first stepped into Machine Learning, I didn’t see magic — I saw math, models, and a whole lot of confusion. Honestly, I was lost more often than not. But the more I read, experimented, and made mistakes, the more things started to click. Slowly, I began to see the logic beneath the layers — how algorithms think, how data speaks, and how everything connects. Machine Learning, at its heart, is simple in purpose but profound in design. It’s about finding patterns — turning data into decisions through experience. Imagine teaching a child to recognize fruits. You show apples and oranges, correct mistakes, and gradually, the child learns what makes each unique. That’s exactly how a machine learns: it takes data → makes predictions → compares them to reality → adjusts to do better next time. This cycle — data → model → prediction → feedback — is the heartbeat of Machine Learning. And here’s what I found fascinating: the intelligence doesn’t lie in memorising data, but in generalising — performing well on data it has never seen. That’s the real definition of learning — for machines, and maybe even for us. 💭 Reflection: What was the most confusing concept for you when you first started exploring Machine Learning? #MLTuesdays #MachineLearning #AIJourney #DataToDecisions #LearningThroughMistakes Link to medium post - https://lnkd.in/gpqtnqct [AI generated image]

ML Made Easy! 🚀 Imagine you’re in a new city and looking for a good restaurant. You don’t know much, so you ask people nearby. If most of them recommend Restaurant A, you’ll probably go there. That’s exactly how KNN works! It looks at the nearest neighbors to make a decision. 📌 Core Idea: For a new data point, KNN looks at the “K” closest points in the dataset. ▪️ For Classification → assigns the majority class among neighbors. ▪️ For Regression → takes the average value of neighbors. 📌 How “closeness” is measured: ▪️ Using Euclidean Distance: Think of it as the “straight-line distance” between two points. Formula between two points (x₁, y₁) and (x₂, y₂): d = √[(x₂ - x₁)² + (y₂ - y₁)²] ▪️ Example: Imagine two students → Student A: studied 5 hours, scored 80. Student B: studied 7 hours, scored 85. The Euclidean distance between them = √[(7-5)² + (85-80)²] = √[4 + 25] = √29. The smaller the distance, the more “similar” the points. 📌 Key Parameters: ▪️ K value: Small K → model is sensitive to noise. Large K → smoother predictions but may ignore local details. ▪️ Distance metric: Euclidean is common, but Manhattan and cosine similarity are also used. 📌 Why Beginners Love KNN: ▪️ Super intuitive, no heavy math at the start. ▪️ Non-parametric → doesn’t assume any fixed data distribution. ▪️ Works well on small, structured datasets. 🔷 Takeaway: KNN is like “asking your neighbors before making a choice.” Simple, practical, and surprisingly effective for many problems! Learning never stops, and every simple concept mastered builds the foundation for Advanced AI !!!📈

NotebookLM just got even better!!! Google’s AI-powered tool for working with your own sources has rolled out exciting new features to make learning and knowledge management more interactive. ✨ New in NotebookLM: Featured Notebooks Source Discovery Quizzes & Flashcards Mindmaps Audio & Video Overviews In my latest article, I walk through these improvements and how they can support your daily work. 👉 https://lnkd.in/d-Mkta3m #NotebookLM #AI #Productivity #Google #Learning #GenerativeAI #DataScience

Ever felt overwhelmed by technical concepts? Here's how I fixed it. For weeks, I struggled with understanding machine learning basics. Every article felt like it was written for PhD students. Then I tried asking AI to explain it like I'm 12 years old. It worked because it removed the intimidation factor and used simple analogies. This led to actually understanding concepts instead of just memorizing definitions. The key is: there's no shame in asking for simple explanations. What's your approach to learning complex topics? Share your tips below.