Data science/ML/AI

📚 What I’m learning for 2027 I’ve been working in software and data science for over 8 years, but lately I’d be lying if I said I wasn’t a little worried about where our jobs are heading. 😅 The future feels more uncertain than ever, so I’ve been thinking seriously about what’s actually worth learning to stay relevant in 2027 and beyond. I searched around for resources I’d personally want to invest my time in, and i figured why not sharing with you guys as well. This is my shortlist 👇 🧠 1. Let’s Build GPT from Scratch, Andrej Karpathy Build a GPT yourself and finally understand what’s happening behind the API. ⏱️ ~2h 🔗 https://www.youtube.com/playlist?list=PLAqhIrjkxbuWI23v9cThsA9GvCAUhRvKZ 🔥 2. Neural Networks: Zero to Hero, Andrej Karpathy A deeper dive into neural networks, backpropagation, language models, GPT and tokenization. ⏱️ ~19h 🔗 https://karpathy.ai/zero-to-hero.html 🤖 3. Hugging Face AI Agents Course Learn how AI agents actually work: tools, actions, reasoning and agentic workflows. 💰 Free 🔗 https://huggingface.co/learn/agents-course/unit0/introduction 🏗 4. Designing Data-Intensive Applications, Martin Kleppmann The classic for understanding databases, distributed systems, replication, partitioning, streams and designing systems that scale. 📖 ~600 pages 🔗 https://github.com/aasthas2022/SDE-Interview-and-Prep-Roadmap/blob/main/System%20Design/Resources/Designing%20Data%20Intensive%20Applications%20by%20Martin%20Kleppmann.pdf ⚙️ 5. Made With ML The production side of ML: deployment, testing, monitoring, data pipelines and MLOps. 💰 Free 🔗 https://madewithml.com/#course 🎯 Why these? My bet for 2027 is that writing code itself will become easier, while understanding AI + production systems + architecture will become even more valuable. So that’s what I’m focusing on. If you know a resource that belongs on this list please share it so everybody can find it valuable. Hope this helps ❤️

✍️ SQL JOIN Explained Visually #SQL

Machine Learning Visualized This is an interactive curriculum with animations and exercises that show how machine learning algorithms actually work. You can watch gradient descent, decision boundaries, neural networks, clustering, and more evolve step by step. It is great for building intuition instead of treating models as black boxes. 🎬 Free Interactive + Animation Course ⏰ Duration: Self-paced 🏃‍♂️ Self Paced 👨‍🏫 Created by: Daniel Sobrado / community project 🔗 Link #MachineLearning #Interactive #Visualization #Course ➖➖➖➖➖➖➖➖➖➖➖➖➖➖ 👉 Join @bigdataspecialist for more 👈

5 LLM quantization techniques, clearly explained: 1. RTN: ignores them. Rounds every weight to the nearest grid level with no calibration data. Cheapest option, weakest at low bit widths. 2. GPTQ: repairs after rounding. Quantizes a layer column by column and adjusts the remaining weights to absorb the error before moving on. 3. AWQ: protects before rounding. Finds the ~1% of weight channels that matter most and scales them up so they survive quantization. Everything still ends up in plain INT4. 4. LLM. int8(): isolates at inference. Outlier dimensions run in FP16, the other 99.9% run in INT8, and the results are merged. 5. QAT: solves it during training. The model is fine-tuned with rounding baked into every forward pass, so it adapts to the damage before quantization is actually applied. All five produce the same artifact, a model at a fraction of its trained precision. They differ only in where the outlier problem gets addressed. The visual above nicely summarises these techniques. #LLM

🗂 10 Websites for Finding Real-World Datasets Finding good datasets is often harder than building the model. These websites cover almost every domain imaginable. 1. Kaggle Datasets 2. Hugging Face Datasets 3. Google Dataset Search 4. UCI Machine Learning Repository 5. OpenML 6. Our World in Data 7. World Bank Open Data 8. data.gov 9. FiveThirtyEight Data 10. AWS Registry of Open Data You'll rarely run out of project ideas with these bookmarked. #Datasets

🎲 What Makes Random Forest "Random"? A Random Forest isn't just "many decision trees." Each tree sees a different random sample of the data. Then... At every split... It only considers a random subset of features. So instead of producing 100 identical trees... You get 100 different opinions. The final prediction is the majority vote (classification) or average (regression). The randomness is exactly what makes the forest stronger.

📦 Your CSV Might Be Using Twice the Memory It Needs Open a CSV in Pandas. Run:

df.info()

You'll often notice many text columns have the type:

object

If a column contains repeated values like:

London
London
London
Paris
Paris
Berlin

convert it to:

category

Instead of storing the full text every time, Pandas stores each unique value once and references it internally. On large datasets, memory usage can drop dramatically.

From Zero to Data Scientist This is a free, open-source curriculum from Microsoft's Azure Cloud Advocates team that breaks data science down into 20 digestible lessons spread across 10 weeks. 👉 Free curriculum with quizzes and assignments 👉 No prior experience needed to start 👉 It's Project-based so you're building a portfolio as you learn 👉 Created by Microsoft experts and students 👉 Has a strong discord community to back you up Explore it here: https://github.com/microsoft/data-science-for-beginners

🧩 Why One-Hot Encoding Exists Machine learning models don't understand words. They understand numbers. So how do you feed a value like:

Color = Red

You can't simply write:

Red = 1
Blue = 2
Green = 3

The model might think Green > Blue > Red, even though colors have no natural order. Instead, we create separate columns:

Red    1 0 0
Blue   0 1 0
Green  0 0 1

This is called One-Hot Encoding. It represents categories without introducing fake relationships.

🚨 UC Berkeley just open-sourced FreeToken. It claims 2–4× faster local LLM inference than Ollama, and the wild part is the models it can run: • Qwen3.6-35B on 8GB VRAM → 39.3 tok/s • DeepSeek-V4-Flash 284B on 32GB VRAM → 22 tok/s • GLM-5.2 753B on 96GB VRAM → 14.9 tok/s How? These are Mixture-of-Experts models. A 35B model doesn't actually use all 35B parameters for every token. FreeToken keeps the experts in system RAM and intelligently decides whether a missing expert should be sent to the GPU or computed on the CPU. The good part is the best strategy depends on your exact machine. A 5090 desktop and an 8GB laptop may want completely opposite approaches. It also checkpoints agent context, so coding agents don't repeatedly prefill thousands of unchanged tokens. Open weights don't mean much if nobody can afford the hardware to run them. FreeToken is attacking that gap. 📄 Paper: https://arxiv.org/pdf/2608.16157 💻 Repo: https://github.com/FlashML-org/FreeToken

Topic: SQL 🔍 Quick look before the question:

SELECT e.name, e.salary
FROM employees e
WHERE e.salary > (
  SELECT AVG(salary)
  FROM employees
  WHERE department = e.department
);

🐼 One Pandas Function That Can Save You From Ugly if/else Suppose you want to classify customers:

spending >= 1000 → VIP
spending >= 500  → Regular
otherwise        → Low

You could write a complicated function. Or:

import numpy as np

df["segment"] = np.select(
    [
        df["spending"] >= 1000,
        df["spending"] >= 500
    ],
    [
        "VIP",
        "Regular"
    ],
    default="Low"
)

Now the rules are visible directly in the code. This becomes especially useful when you have several conditions.