Library — Karthik’s Picks | Gruhesh Sri Sai Karthik

Book

Mathematics for Machine Learning

Deisenroth, Faisal & Ong — Cambridge (free PDF)

Linear algebra, calculus, and probability with ML context. Still the best gap-filler before rigorous grad-level ML courses.

Video

Essence of Linear Algebra

3Blue1Brown — YouTube

Visual intuition for vectors, matrices, and transforms. Watch before any LA course — pairs with my linear algebra post.

Video

Neural Networks (3Blue1Brown)

3Blue1Brown — YouTube

Gradient descent and backprop as geometry, not symbol pushing. The visual complement to coding-first courses.

Video

Neural Networks: Zero to Hero

Andrej Karpathy — YouTube

micrograd → makemore → GPT from scratch in PyTorch. Still the highest-ROI free playlist in ML (2026 lists agree).

Course

Practical Deep Learning for Coders

Jeremy Howard — fast.ai (free)

Top-down: train a model first, derive the theory after. Fast path to usable skills while Karpathy builds depth.

Course

MIT Introduction to Deep Learning (6.S191)

MIT — annual lectures & labs (free)

Compact university survey updated each year. Good bridge between intuition playlists and Stanford-scale LM courses.

Guide

The Illustrated Transformer

Jay Alammar

The canonical visual walkthrough of self-attention. Read before the original paper — still cited in every 2026 LLM roadmap.

Video

But what is a GPT?

3Blue1Brown — YouTube

Tokens, embeddings, and next-token prediction without drowning in matrix notation. Matches how I introduce GPT on the blog.

Course

How Transformer LLMs Work

DeepLearning.AI — Alammar & Grootendorst (free)

Short course on tokenization, attention, KV cache, GQA, and Hugging Face implementations. Current through 2025–26 architecture tweaks.

Course

Hugging Face LLM Course

Hugging Face (free)

Hands-on transformers, fine-tuning, and deployment on the Hub. The standard free path after you understand attention.

Course

Stanford CS336 — Language Modeling from Scratch

Stanford — Spring 2026

Build tokenizer, transformer, distributed trainer, and data pipeline from zero. The course the industry points to for serious LM engineering.

Course

Hugging Face AI Agents Course

Hugging Face (free, certified)

Agents, tool use, smolagents, LangGraph, LlamaIndex — the 2025–26 skill layer on top of base LLMs.

Guide

How to Scale Your Model

Google DeepMind — jax-ml (free)

Blog-style textbook on scaling LLMs across TPUs/GPUs — sharding, parallelism, and roofline math for training and inference. The systems-side companion to CS336.

Video

Let's Build GPT: from scratch

Andrej Karpathy — YouTube

~2 hours implementing a GPT-style decoder. Watch after Illustrated Transformer; code-along with my GPT-2 build post.

Book

Deep Learning

Goodfellow, Bengio & Courville — MIT Press (free online)

Reference text for optimisation, regularisation, and sequence models. Part II remains the practical core.

Book

Hands-On Large Language Models

Alammar & Grootendorst — O'Reilly

Code-first tour of modern LLMs — same authors as the DeepLearning.AI transformer short course. Best book-shaped companion to HF.

Book

AI Engineering

Chip Huyen — O'Reilly

Production LLM systems: data, eval, deployment, and agents. The shift in 2025–26 from 'train a model' to 'ship a product'.

Book

Designing Machine Learning Systems

Chip Huyen — O'Reilly

ML in production — data loops, monitoring, and iteration. Pairs with AI Engineering for the full systems picture.

Book

Reinforcement Learning: An Introduction

Sutton & Barto — MIT Press (free PDF)

RL foundations. Chapters 3–6 before DeepSeek-R1-style reasoning papers make sense.

Paper

Attention Is All You Need

Vaswani et al. — NeurIPS 2017

The transformer. Section 3 is the whole architecture — read after Illustrated Transformer + a code walkthrough.

Paper

Training Compute-Optimal Large Language Models (Chinchilla)

Hoffmann et al. — DeepMind 2022

~20 tokens per parameter for compute-optimal training. Still the scaling law every 2026 efficiency paper argues with.

Paper

Language Models are Unsupervised Multitask Learners (GPT-2)

Radford et al. — OpenAI 2019

Decoder-only LM at scale. Short and readable — pair with Karpathy's GPT video and my GPT-2 build.

Paper

DeepSeek-R1: Reasoning via Reinforcement Learning

DeepSeek-AI — 2025

Reasoning emerges from RL (GRPO) without heavy SFT. Defines the 2025–26 'test-time compute / reasoning model' wave.

Paper

LoRA: Low-Rank Adaptation of Large Language Models

Hu et al. — ICLR 2022

Parameter-efficient fine-tuning everyone uses in practice. Short paper, huge impact on how we actually train.

Paper

DeepSeek-V3 Technical Report

DeepSeek-AI — 2024

671B-param MoE, 37B active per token, trained at a fraction of typical frontier-model cost. The paper that reset expectations on training efficiency going into 2025–26.

Paper

Kimi K2: Open Agentic Intelligence

Kimi Team — Moonshot AI, 2025

1T-param open-weight MoE tuned for agentic/tool-use tasks, with the MuonClip optimizer for zero-loss-spike training at scale.

Course

CS224N: NLP with Deep Learning

Stanford — lecture slides (free)

Stanford NLP with transformers, BERT, and GPT-era design. Lecture 8 on attention is a classic.

Course

CS231n: CNNs for Visual Recognition

Stanford — notes & assignments (free)

Vision gold standard. Assignment 2 (backprop from scratch) is still one of the best ML exercises.

Course

Machine Learning Specialization

DeepLearning.AI & Andrew Ng (Coursera audit free)

Structured ML foundations if you want certificates and paced weeks before diving into LLM-specific material.

Tool

Hugging Face Hub

Hugging Face — free tier

Models, datasets, Spaces, and the courses above. The open default for sharing and running LLM experiments.

Tool

PyTorch

PyTorch Foundation — open source

What Karpathy, CS336, and most research code use. Learn autograd here before trusting high-level trainers.

Tool

Ollama

Ollama — free

Run open-weight LLMs locally with one command. Best way to test quantised models on a laptop.

Tool

vLLM

vLLM — open source

PagedAttention and high-throughput serving. What people reach for when batching and KV-cache efficiency matter.

Tool

smolagents

Hugging Face — open source

Minimal agent framework from HF. Good first library after the Agents Course — less surface area than LangChain.

Tool

Weights & Biases

W&B — free for individuals

Experiment tracking and hyperparameter sweeps. Three lines to log a training run.

Tool

Hugging Face Papers

Hugging Face

Papers With Code was discontinued and now redirects here. Daily trending papers with linked code, discussion, and SOTA context.

Tool

Cursor

Anysphere — freemium

Editor with LLM in the loop. Useful for implementing coursework and debugging training scripts.