It was a map of Spain. Every block, 0.1% of the population.

The follower count moved too. From the outside, it looked like something had happened.

118,712 impressions. 124 new followers.

I sat with those two numbers for a while.

98% of that reach was out of network. Which is a polite way of saying the algorithm carried it to strangers, and the algorithm can stop carrying it whenever it decides to.

Last September, a linear regression cheatsheet on X did something similar. 128,000 impressions. 11,533 interactions.

349 people expanded the post to actually look at the diagram.

That’s the one that stayed with me.

It was a cheatsheet. Being read was the entire point of it. Eleven thousand people engaged with something fewer than four hundred opened.

Surely the reach is still worth something?

Something. Not what we think it is.

Roughly a quarter of a million impressions across two posts. 124 followers. 349 people who looked at the thing I actually made.

Reach is a crowd walking past a window. An audience is whoever came inside.

Impressions are rented. They arrive because an algorithm moved them, and they leave the same way. You don’t own a single one of them.

This list is around 9,500 people. Every one of them typed an email address and pressed a button. Nobody subscribes by accident.

The number that matters is not who saw it. It’s who came back without being asked.

So I stopped checking impressions. I check opens now. That number moves slowly; it looks small next to 118,712, and nobody screenshots it.

It’s also the only one that has ever turned into anything real.

Anyone can be seen. Almost nobody gets returned to.

— Josep

Are you still here? 🧐

👉🏻 I want this newsletter to be useful, so please let me know your feedback!

Before you go, tap the 💚 and the restack buttons at the bottom of this email to show your support, it really helps and means a lot!

What's the biggest number you've ever posted, and what did it actually change? 👇🏻

Feeling behind? You’re not.
Today’s issue compresses the essentials (what matters, what doesn’t) into a buildable path.

Minimal theory, maximum leverage.

Following my Transformers cheat sheets (architecture, encoder, decoder), today we go end-to-end:

1. What to learn

2. What to build first

3. How to avoid the rabbit holes.

⚠️ It’s a longer, denser issue, but it’s meant to be a keeper. Bookmark it, steal the prompts, and ship something this week.

Subscribe now

TL;DR (paste this in your notes)

• LLMs ≠ magic. Learn the Transformer + tokens + pretrain→post-train→inference pipeline.

• Start “outside in.” Ship value via APIs or open models first; fine-tune later.

• Leverage > novelty. Framing, evaluation, and alignment matter more than training a giant from scratch.

Why this, why now

Understanding LLMs and GenAI is crucial for everyone, from seasoned data professionals to beginners, as they are set to revolutionize text data processing and our future. With new models and applications constantly emerging, it’s essential to stay updated and maintain sharp skills in this rapidly evolving field.

#1 Understanding the Basics

What are LLMs?

Large Language Models are a type of artificial intelligence trained on extensive text datasets. These models can generate human-like text, understand context, and even carry on conversations. They’re used in various applications, from chatbots to content creation and beyond.

So… why are they so popular?

LMs are popular due to their ability to generate coherent, contextually relevant, and grammatically accurate text. Their exceptional performance on diverse language tasks and the accessibility of pre-trained models have democratized AI-powered natural language understanding and generation.

LLMs core components

Key concepts of LLMs include:

• Transformer Architecture: It is the backbone of LLMs, featuring self-attention mechanisms that enable the model to weigh the importance of different words in a sentence.

• Tokenization: Breaking down text into manageable pieces or tokens. This is performed by Tokenizers.

• Pre-training: Involves training the model on a large corpus of text to learn language patterns, grammar, and context.

• Fine-tuning: Adapts the pre-trained model to specific tasks using smaller, task-specific datasets.

• NLU (Natural Language Understanding): The ability to understand and interpret human language.

• NLG (Natural language Generation): The ability to generate coherent and contextually relevant text.

• Prompt Engineering: Crafting input prompts to guide the model towards generating desired outputs, essential for tasks performed via API access.

Main Differences between LLMs and Deep Learning Models

LLMs differ from other deep learning models primarily due to their size and use of self-attention mechanisms. Key differentiators include:

• Transformer Architecture: This revolutionary design underpins LLMs and has transformed natural language processing.

• Contextual Understanding: LLMs capture long-range dependencies in text, enhancing their contextual comprehension.

• Versatility: They excel in various language tasks, including text generation, translation, summarization, and question-answering.

Subscribe now

#2 How to get started with LLMs?

1. Understanding the Transformer Architecture in LLMs

Now that you’re familiar with LLMs, let’s delve into the Transformer architecture that powers these models. The original Transformer, introduced in the paper Attention Is All You Need, revolutionized natural language processing.

Key Features:

• Self-Attention Layers: Allow the model to focus on different parts of the input sequence.

• Multi-Head Attention: Enables the model to attend to information from different representation subspaces.

• Feed-Forward Neural Networks: Process the output from the attention mechanism.

• Encoder-Decoder Architecture: Facilitates tasks like translation.

Remember, you can learn more about it in the following article about the Transformers Architecture.

2. Pre-training LLMs

Now that you understand the fundamentals of LLMs and the transformer architecture, it’s time to explore pre-training LLMs. Pre-training is crucial for enabling LLMs to grasp human language by exposing them to huge amounts of text.

This part is (usually) performed by companies like OpenAI, Google, DeepSeek, Meta, or Anthropic.

Key Concepts:

• Objectives of Pre-training: LLMs learn language patterns, grammar, and context through exposure to extensive text corpora. Key tasks include masked language modeling and next sentence prediction.

• Text Corpus for Pre-training: LLMs are trained on diverse and massive datasets, including web articles, books, and more, with billions to trillions of text tokens. Common datasets are C4, BookCorpus, Pile, OpenWebText, etc.

• Training Procedure: Understand the technical aspects such as optimization algorithms, batch sizes, and training epochs, and learn about challenges like mitigating data biases.

For further learning, check out the module on LLM training from CS324: Large Language Models.

As training an LLM from scratch requires a lot of resources, we can access pre-trained models directly via API (OpenAI, Google…) or using open-source models in HuggingFace.

3. Accessing LLMs and using them

In today’s landscape, accessing and utilizing LLMs has become easier than ever, thanks to both commercial APIs and open-source platforms.

Using Commercial APIs

The most common one is OpenAI and their GPT models, but others like Anthropic can be used as well.

• API Access: OpenAI provides robust API access to its models, such as GPT-4 and ChatGPT, allowing developers to integrate powerful language capabilities into their applications.

• Ease of Use: With simple HTTP requests, you can send text prompts to the API and receive generated responses. The API supports various parameters to fine-tune the behavior of the model, such as temperature, max tokens, and more.

• Applications: This API is versatile and can be used for chatbots, content generation, summarization, translation, and other NLP tasks.

Using Open-Source Models (Hugging Face)

• Model Hub: Hugging Face offers a vast repository of open-source models, including versions of GPT, BERT, T5, Mistral, Meta’s Llama and many more, which can be accessed for specific tasks.

• Transformers Library: The Transformers library by Hugging Face provides a comprehensive toolkit for using and fine-tuning these models. It supports multiple frameworks, including TensorFlow and PyTorch.

• Ease of Use: With Hugging Face, you can load pre-trained models with just a few lines of code and fine-tune them on your dataset. The library also offers utilities for tokenization, training, and deploying models.

4. Fine-Tuning LLMs

Once we know how to access and use pre-trained LLMs, the next step is understanding the process of fine-tuning and how to train them for specific tasks. Fine-tuning tailors pre-trained models to perform tasks like sentiment analysis, question answering, or translation with greater accuracy and efficiency.

Why Fine-Tune LLMs?

• Task-Specific Performance: While pre-trained LLMs have a general understanding of language, fine-tuning is essential to excel in specific tasks by learning their unique nuances.

• Efficiency: Fine-tuning leverages the pre-trained model’s knowledge, reducing the data and computation needed compared to training from scratch. This process requires a much smaller dataset.

Fine-Tuning LLMs with access to their weights

1. Choose the Pre-trained LLM: Select a pre-trained model that suits your task. For instance, for question-answering, choose a model designed for natural language understanding.

2. Data Preparation: Prepare a labeled dataset for your specific task, ensuring it is properly formatted.

3. Fine-Tuning Process:

   • Use parameter-efficient techniques to fine-tune the model, considering LLMs have tens of billions of parameters.

   • If you don’t have access to the weights, explore alternative approaches or frameworks that facilitate fine-tuning without direct weight manipulation.

By following these steps, you can adapt pre-trained LLMs to achieve optimal performance on your desired tasks. You can read more about it here.

Fine-Tuning LLMs Without Access to Model Weights

When you don’t have access to an LLM’s weights and must use an API, you can still fine-tune the model using in-context learning and prompt tuning.

1. In-Context Learning: Leverage the LLM’s ability to learn from provided examples. By giving input-output examples within the prompt, the model can perform tasks without explicit fine-tuning.

2. Prompt Tuning:

   • Hard Prompt Tuning: Modify the input tokens directly in the prompt to guide the model’s output.

   • Soft Prompt Tuning: Concatenate the input embedding with a learnable tensor. Prefix tuning is a related approach where learnable tensors are used with each Transformer block, not just the input embeddings.

3. Parameter-Efficient Fine-Tuning Techniques (PEFT):

   • LoRA and QLoRA: These techniques allow fine-tuning by introducing a small set of learnable parameters, called adapters, instead of updating the entire weight matrix. QLoRA, for instance, enables fine-tuning a 4-bit quantized LLM on a single consumer GPU without performance loss.

By using these methods, you can adapt LLMs for specific tasks efficiently, even without direct access to the model’s weights. Here are some resources to explore further:

• Quantization for LLMs: Reduce AI Model Sizes Efficiently

• Prompt Engineering Course by OpenAI — Inferring, Transforming, and Expanding with ChatGPT

And don’t forget to check my webinar about fine-tuning Disilbert and Mistral 7B!

5. Alignment and Post-Training in LLMs

LLMs can sometimes generate content that is harmful, biased, or misaligned with user expectations. Alignment involves adjusting an LLM’s behavior to align with human preferences and ethical standards, aiming to reduce the risks of biased, controversial, or harmful content.

Techniques to Explore:

• Reinforcement Learning from Human Feedback (RLHF): This method uses human annotations on LLM outputs to train a reward model, guiding the model to produce more desirable outputs.

• Contrastive Post-Training: This technique leverages contrastive methods to automatically create preference pairs, refining the model’s responses to better match user expectations.

By employing these techniques, you can enhance the alignment of LLMs, ensuring they produce content that is safe, ethical, and aligned with human values.

Subscribe now

6. Evaluating LLMs

Evaluating the performance of LLMs is crucial to assess their effectiveness and identify areas for improvement. Key aspects of LLM evaluation include:

1. Task-Specific Metrics: Select appropriate metrics for your specific task. For example:

   • Text Classification: Use metrics like accuracy, precision, recall, and F1 score.

   • Language Generation: Metrics such as perplexity and BLEU scores are commonly used.

2. Human Evaluation: Have experts or crowdsourced annotators assess the quality of generated content or model responses in real-world scenarios.

3. Bias and Fairness: Evaluate LLMs for biases and fairness, especially when deploying them in real-world applications. Analyze performance across different demographic groups and address any disparities.

4. Robustness and Adversarial Testing: Test the LLM’s robustness by subjecting it to adversarial attacks or challenging inputs to uncover vulnerabilities and enhance model security.

7. Continuous Learning and Adaptation

To keep LLMs updated with new data and tasks, consider these strategies:

1. Data Augmentation: Continuously augment your dataset to prevent performance degradation due to outdated information.

2. Retraining: Periodically retrain the LLM with new data and fine-tune it for evolving tasks to ensure the model stays current.

3. Active Learning: Implement active learning techniques to identify instances where the model is uncertain or likely to make errors. Collect annotations for these instances to refine the model.

Additionally, to mitigate common issues like hallucinations, explore techniques such as retrieval augmentation.

Subscribe now

#3 Building and Deploying LLM Applications

Once you’ve developed and fine-tuned an LLM for specific tasks, the next step is to build and deploy applications that harness the LLM’s capabilities. This involves creating practical, real-world solutions that make the most of your LLM’s potential.

Building LLM Applications

When developing applications that leverage Large Language Models (LLMs), consider the following:

1. Task-Specific Application Development:

   Tailor your applications to meet specific use cases, such as web interfaces, mobile apps, chatbots, or integrations into existing software systems.

2. User Experience (UX) Design:

   Prioritize user-centered design to ensure your LLM application is intuitive, user-friendly, and meets the needs of your target audience.

3. API Integration:

   If your LLM acts as a language model backend, create RESTful APIs or GraphQL endpoints to facilitate seamless interaction with other software components.

4. Scalability and Performance:

   Design your applications to handle varying levels of traffic and demand. Optimize for performance and scalability to provide a smooth and reliable user experience.

Deploying LLM Applications

Now that you’ve developed your LLM application, it’s time to deploy it to production. Here are key considerations for a successful deployment:

1. Cloud Deployment:

   Deploy your LLM applications on cloud platforms like AWS, Google Cloud, or Azure. These platforms offer scalability, reliability, and easy management of resources.

2. Containerization:

   Use containerization technologies such as Docker and Kubernetes to package your applications. This ensures consistent deployment across various environments and simplifies scaling and management.

3. Monitoring:

   Implement robust monitoring solutions to track the performance of your deployed LLM applications. This allows you to detect and address issues in real time, ensuring optimal performance and reliability.

Practical experience is crucial. Here’s how you can get hands-on:

• Welcome to the Hands-on LLM Course by

• Hugging Face and PyTorch Lightning by Jon Krohn.

Subscribe now

A final note

If you’ve made it this far, you’ve already taken the first step: understanding that this isn’t about knowing everything, it’s about moving forward bit by bit.

With patience, curiosity, and consistency.

No one starts out knowing.

But we all start in the same place: by taking the first step.

Are you in?

Hope to see you in the community soon!

Sincerely,

— Josep

Your turn

Some final resources to check:

• Attention Is All You Need (must read)

• My illustrated Transformers saga (architecture, encoder, decoder)

• Module on Modeling from Stanford CS324: Large Language Models

• HuggingFace Transformers Course

Are you still here? 🧐

👉🏻 I want this newsletter to be useful, so please let me know your feedback!

Before you go, tap the 💚 and the restack buttons at the bottom of this email to show your support; it really helps and means a lot!

P.S. Share with the coworker who thinks self-attention is a personality trait.

Any doubt? Let’s start a conversation! 👇🏻

Almost every reply picked out the same word.

Is SQL still worth learning in 2026?

I get where the question comes from. SQL is what you learn in week two of a bootcamp. No release cycle, no conference, no version number anyone argues about. It looks finished.

So p…

Read more

She could build an end-to-end scikit-learn Pipeline in her sleep.

Six months into her first job, she’d used it exactly once.

“I feel like I forgot everything.”

She hadn’t forgotten anything. The stack just changed the moment she stopped interviewing.

The Python libraries you learn for intervie…

Read more

Today we go one layer up. Not stitching tables together, but organizing the whole query so it reads like structured code instead of a wall of nested logic.

Here’s this week’s cheatsheet 👇🏻

Ever inherited a SQL query from someone who left three months ago?

Nothing lines up. The same metric is calculated three different…

Read more

I have a bunch of projects on GitHub. Is that my portfolio?

Short answer: no.

A GitHub is a repository. A portfolio is a story.

Two different objects. Two different audiences. Two different jobs.

GitHub is for the person who wants to read your code. That’s a tiny audience. Usually another engineer. Usually al…

Read more

Today we go one layer deeper: the joins themselves.

Merging tables is where SQL gets real. And where most people get stuck.

LEFT, RIGHT, INNER, FULL.

Which one when?

What happens to unmatched rows?

Here’s the full cheatsheet 👇🏻

The Magic of SQL Joins

In the world of databases, data often resides in sepa…

Read more

Eight weeks ago, I hit publish on the first post of the databites reset.

I didn’t know if anyone would read it.
I didn’t know if I’d keep going.
I didn’t know if the format would hold up.

Eight Tuesdays later, we’re still here.

You opened the emails. You replied. You clicked. You stayed.

That means more than you probably realise. Eight weeks was the priva…

Read more

In this issue, we’ll break down the essentials of using APIs for data…

Read more

Not writer’s block. Something worse.

The feeling that whatever I wrote wouldn’t be worth anyone’s time.

I published anyway. You’re reading it now.

That’s what seven weeks of showing up actually looks like from the inside. Not a streak. Not momentum. A decision you make again every week, usually under so…

Read more

Transformers.

Here’s what’s coming your way:

✅ Week 1: Understanding the Transformers architecture → Link
✅ Week 2: The Encoder → Link
✅ Week 3: The Decoder → Link

Subscribe now

Understanding the Decoder - Part III

The decoder’s primary role …

Read more

My instinct was to separate everything cleanly. One client on Monday and Tuesday. Another on Wednesday and Thursday. Fridays for admin, proposals, the business side of things.

Three clients. Five days. Neat boxes.

It worked for about a month.

Then a client needed something on a Thursday. An…

Read more

Transformers

Here’s what’s coming your way:

✅ Week 1: Understanding the Transformers architecture → Link
✅ Week 2: The Encoder → Link
✅ Week 3: The Decoder → Link

Understanding the Encoder - Part II

The encoder is a fundamental component …

Read more

I had never taught at a university before.

No formal teaching experience. No academic track record.

Just the diagrams.

They’d seen the work, decided I understood something worth teaching, and reached out.

I said yes. Went all in. And it became one of the…

Read more

Transformers.

Here’s what’s coming your way:

✅ Week 1: Understanding the Transformers architecture
✅ Week 2: The Encoder → Link
✅ Week 3: The Decoder → Link

Subscribe now

The Transformers Architecture - Part I

With GPT-3.5 gaining 1 million users in a w…

Read more

Nobody asked about the model.

They asked: “What does this mean for us next quarter?”

I didn’t have that answer ready.

That’s not a modeling problem.

That’s a communication problem.

Most data education teaches you to build correctly.

Almost none of it teaches you to explain what you bui…

Read more

With one table, that’s easy to hold in your head. Add a JOIN, and one question trips up almost everyone: when you combine two tables and then filter them, what runs first?

The JOIN. It builds the combined table before any filter tou…

Read more

Watched the videos. Did the math. Understood the intuition behind every update step.

Then I sat in front of a real dataset. Messy, incomplete, and half-documented.

I had no idea what to do first.

That’s the gap nobody talks about.

Learning data science is something you can measure.

Courses completed. Concep…

Read more

The one you wrote. And the one your database actually uses to run it.

That’s it. Strip away the syntax, and that’s what’s happening underneath: the clauses you typed top to bottom get quietly reshuffled into a completely different sequence before a single row is touched.

Most people learn to write SQL. Almost n…

Read more

A new research project.

A new side concept I never shipped.

A new approach to content I mapped out in a notebook and never opened again.

Each one felt like momentum. Each one was actually a reset.

This is what happens without a system: you don’t stop working.

You just keep star…

Read more