8/5/2025

I finally understand how large language models actually work After reading the 2025 textbook “Foundations of LLMs” It blew my mind and cleared up years of confusion Here’s every...

20 tweets

4 min read

Sehaj Singh

@heysehajsingh

I finally understand how large language models actually work After reading the 2025 textbook “Foundations of LLMs” It blew my mind and cleared up years of confusion Here’s everything i learned (in plain english):

🖼️ 1 image loading...

Sehaj Singh

@heysehajsingh

To understand LLMs, start with pre-training. We don’t teach them specific tasks. We flood them with raw text and let them discover patterns on their own. This technique is called self-supervised learning and it’s the foundation of everything.

Sehaj Singh

@heysehajsingh

There are 3 ways to pre-train: → Unsupervised: No labels at all → Supervised: Classic labeled data → Self-supervised: Model creates its own labels (e.g., “guess the missing word”) LLMs use #3 it scales like crazy and teaches them language from scratch.

Sehaj Singh

@heysehajsingh

Example of self-supervised learning: “The early bird catches the worm.” Mask some words: → “The [MASK] bird catches the [MASK]” The model’s job? Fill in the blanks. No human labels. The text is the supervision.

🖼️ 1 image loading...

Sehaj Singh

@heysehajsingh

This leads to 3 main model types: → Encoder-only (BERT): Understands text → Decoder-only (GPT): Generates next word → Encoder-decoder (T5): Translates input to output Each has strengths. Think of them as different tools for different jobs.

Sehaj Singh

@heysehajsingh

Let’s break it down further. Decoder-only (GPT-style): Trained to guess the next word: “The cat sat on the ___” → “mat” This is called causal language modeling. Loss is measured by how wrong the guesses are (cross-entropy).

🖼️ 1 image loading...

Sehaj Singh

@heysehajsingh

Encoder-only (BERT-style): Takes the whole sentence. Randomly hides some words and tries to reconstruct them. This is masked language modeling uses left and right context. Great for understanding, not generation.

Sehaj Singh

@heysehajsingh

Example: Original: → “The early bird catches the worm” Masked: → “The [MASK] bird catches the [MASK]” The model predicts “early” and “worm” by understanding the whole sentence. It’s learning language by solving puzzles.

Sehaj Singh

@heysehajsingh

Encoder-decoder (T5, BART): Treats everything as a text-to-text task. Examples: “Translate English to German: hello” → “hallo” “Sentiment: I hate this” → “negative” This setup lets one model do it all: QA, summarization, translation, etc.

🖼️ 1 image loading...

Sehaj Singh

@heysehajsingh

Once pre-trained, we have two options: → Fine-tune it on a labeled dataset → Prompt it cleverly to do new tasks Fine-tuning adjusts weights. Prompting? Just tweaks the input text. Let’s dive into the magic of prompts.

Sehaj Singh

@heysehajsingh

Prompting = carefully phrasing input so the model does what you want. Example: “I love this movie. Sentiment:” It’ll likely respond: “positive” Add a few examples before it? That’s in-context learning no fine-tuning needed.

Sehaj Singh

@heysehajsingh

Prompting gets deep. Advanced strategies: • Chain of thought → “Let’s think step by step...” • Decomposition → Break complex tasks into parts • Self-refinement → Ask the model to critique itself • RAG → Let it fetch real-time data from external sources

🖼️ 1 image loading...

Sehaj Singh

@heysehajsingh

This is all possible because of the way these models are trained: predict the next word over and over until they internalize language structure, reasoning patterns, and world knowledge. It's not magic. it's scale.

Sehaj Singh

@heysehajsingh

But raw intelligence isn’t enough. We need models to align with human goals. That’s where alignment comes in. It happens in two major phases after pretraining 👇

Sehaj Singh

@heysehajsingh

Supervised Fine-Tuning (SFT) Feed the model good human responses. Let it learn how we want it to reply. RLHF (Reinforcement Learning w/ Human Feedback) Train a reward model to prefer helpful answers. Use it to steer the LLM. This is how ChatGPT was aligned. RLHF is powerful but tricky. Newer methods like Direct Preference Optimization (DPO) are rising fast. Why? They skip the unstable reward modeling of RL and go straight to optimizing for preferences. More stable. More scalable.

Sehaj Singh

@heysehajsingh

Inference how the model runs is just as important as training. To serve real-time outputs, we use tricks like: → Top-k / nucleus sampling → Caching past tokens → Batching requests → Memory-efficient attention These make LLMs usable at scale.

Sehaj Singh

@heysehajsingh

So how do LLMs really work? Trained on massive text Predict the next word millions of times Use Transformers to encode dependencies Adapt via prompting or fine-tuning Aligned to human preferences Served with smart inference

Sehaj Singh

@heysehajsingh

this was based on the brilliant textbook: "Foundations of Large Language Models" by Tong Xiao and Jingbo Zhu (NiuTrans Research Lab) arxiv: arxiv.org/abs/2501.09223… highly recommend it if you're serious about understanding LLMs deeply.

🔗 Link preview loading...

Sehaj Singh

@heysehajsingh

Want to learn how AI works and automate your work? Subscribe to our free newsletter here to learn AI: theprohuman.ai/subscribe

🔗 Link preview loading...

Sehaj Singh

@heysehajsingh

I hope you've found this thread helpful. Follow me @heysehajsingh for more. Like/Repost the quote below if you can:

https://twitter.com/heysehajsingh/status/1952663453765906504

Rate this thread

Help others discover quality content

Ready to create your own threads?

Get Started Free