llamasearch.dev

My notes on AI — rewritten from first principles, in code, with interactive demos.

Abhishek Bisht
Abhishek Bisht
Applied AI Research Engineer. Building and writing about the things that matter.
Deep Learning

Neurons All the Way Down

From counting words to training a neural language model.

7 chapters · n-grams → optimizers → BPE
NLP

The Transformer Deep Cuts

The transformer, one concept at a time.

10 chapters · attention → MoE → VLMs
Architectures

The Evolving Transformer

One codebase, model by model. Each stop shows only what changed.

9 chapters · 2017 → DeepSeek → PaliGemma
Pre-Training

The Training Ground

Pre-training and fine-tuning at scale.

6 chapters · data → scaling → LoRA
Inference

Beyond the Forward Pass

Decoding, inference, and serving.

4 chapters · sampling → paging → speculation
Deep Learning

Decoding DeepSeek

What DeepSeek invented: MLA, MoE, multi-token prediction, FP8.

7 chapters · V3 case study
RL + Reasoning

Reasoning from Scratch

RL, alignment, and reasoning — starting with no extra training.

12 chapters · CoT → PPO → GRPO → R1
NLP

Trust, but Verify

Using and trusting LLMs: evaluation, RAG, agents, and safety.

5 chapters
Search

Finding Needles

From the inverted index to agents that search for themselves.

20 chapters · 6 arcs
Architectures

Architecture Deltas

The nine architectures from The Evolving Transformer, each as a change from the stop before it.

9 stops · 6 maps · single page
Architectures

The Architecture Atlas

The five diagrams we have for The Evolving Transformer.

5 diagrams · gallery + compare
Research

The Paper Trail

Every paper cited here — searchable, tagged, one click away.

152 papers · 9 series · searchable & paginated
Interview

The Gauntlet

Interview question sets and Python practice. Reveal, track, quiz yourself.

6 sets · Python Under the Hood inside · self-test
Python

The Sandbox

Write and run Python in your browser. No install.

Pyodide · runs locally
Archives

The Archives

Early notes, 2018–2024. The path that led here.

76 posts · 2018 – 2024 · year-wise
Course

AI Researcher Course

Six months. Ten phases. Applied AI research, end to end.

10 phases · 6 months · syllabus + tracker