LLM Internals

LLM Internals is a free, open-source educational resource maintained by Amit Shekhar, founder of Outcome School and IIT alumnus (2010–14). Published under the Apache License 2.0 on GitHub, it provides a growing series of in-depth blogs and videos that walk through how large language models actually work under the hood — from tokenization to advanced inference optimization. The project links directly to Outcome School's 9-month AI and Machine Learning Program, serving as a public companion resource for developers who want to go beyond using AI tools and understand the internals.

What It Is

LLM Internals is a curated, open-source reference guide structured as a series of standalone deep-dives into the core concepts powering modern LLMs. Each entry in the repository links to a detailed blog post or video that explains one specific topic — such as KV Cache, Paged Attention, Flash Attention, or LoRA — with step-by-step numeric examples and clear analogies. The project is not a software library or framework; it is a learning resource designed to build deep conceptual understanding for software engineers transitioning into AI roles.

Curriculum Coverage

The repository covers a broad and growing set of LLM internals topics, organized into logical clusters:

• Tokenization: Byte Pair Encoding (BPE), tokenization mechanics
• Attention: Q/K/V matrices, scaling factor math (√dₖ), causal masking, multi-head attention, grouped query attention, Flash Attention, RoPE (Rotary Position Embedding)
• Transformer Architecture: Feed-forward networks, residual connections, layer normalization, RMSNorm, encoder-decoder design, Mixture of Experts (MoE)
• Inference Optimization: KV Cache, Paged Attention, Speculative Decoding, Medusa, EAGLE, Continuous Batching, Prompt Caching, Prefill vs Decode phases
• Training Techniques: Backpropagation math, cross-entropy loss, LoRA fine-tuning, RLHF, Large Reasoning Models (LRMs)
• AI Security: Prompt injection, LLM watermarking, the Lost in the Middle problem
• Emerging Topics: Harness Engineering, DeepSeek-V4 architecture decoding

Connection to Outcome School's AI Program

The GitHub repository serves as the public-facing, open-source layer of Outcome School's broader AI and Machine Learning Program — a 9-month live online program targeting working tech professionals who want to transition into roles such as AI Engineer, LLM Engineer, MLOps Engineer, or Agentic AI Engineer. The program's curriculum extends the repository's content into hands-on system design, building projects from scratch (including LLMs, AI agents, and neural networks), mock interviews, and career placement support. The repository itself is maintained by the program's instructor and is explicitly described as a growing series that will expand as new blogs and videos are published.

Audience and Prerequisites

The resource targets software engineers who already have programming experience in any language and want to develop a deep, internals-level understanding of AI and ML. No prior Python experience is required. The stated goal is to move learners from being AI consumers — using tools like Claude or ChatGPT — to AI builders who understand transformer architectures, attention mechanisms, and inference pipelines at a code and math level.

Update: Active and Growing Repository

The repository was created in April 2026 and was last updated in September 2026, with 1,687 GitHub stars and 186 forks as of that date. The README explicitly notes: "This series will continue to grow as I write more blogs and create more videos on new topics." Recent additions include coverage of DeepSeek-V4, EAGLE speculative decoding, LLM watermarking, and Harness Engineering in AI — indicating active, ongoing content development aligned with the latest developments in the LLM space.