Why Is the Sea Salty but Rivers Are Fresh?
Rain slowly dissolves minerals from rocks, rivers carry them to the sea, and evaporation leaves the salt behind — for hundreds of millions of…
Read article →Atul Vij
Python: Rename Hundreds of Files From an Excel List
Python: Rename Hundreds of Files From an Excel List
Rename scanned invoices, photos or reports in bulk using an Excel sheet of old and new names — with a dry run, a log,…
Read article →The Complete Guide to Excel Lookups: VLOOKUP, XLOOKUP, INDEX-MATCH and Beyond
The Complete Guide to Excel Lookups: VLOOKUP, XLOOKUP, INDEX-MATCH and Beyond
Everything about looking up data in Excel in one place: exact and approximate matches, left lookups, two-way lookups, multiple criteria, last match, returning all…
Read article →Build Your Own LLM, Lesson 12: Quantize and Run Your Model Locally
Build Your Own LLM, Lesson 12: Quantize and Run Your Model Locally
Shrink models with quantization: fp32 to int8 with PyTorch dynamic quantization, size and speed trade-offs, GGUF and 4-bit formats, and running open models locally…
Read article →Build Your Own LLM, Lesson 11: Instruction Tuning and Chat Templates
Build Your Own LLM, Lesson 11: Instruction Tuning and Chat Templates
Turn a text predictor into an assistant: instruction-response data, chat templates with role tokens, masking the loss on prompts, and why each model needs…
Read article →Build Your Own LLM, Lesson 10: Fine-Tuning a Pre-Trained Model With LoRA
Build Your Own LLM, Lesson 10: Fine-Tuning a Pre-Trained Model With LoRA
Fine-tune efficiently with LoRA: freeze the base model, add low-rank adapter matrices to attention layers, train under 2% of the parameters, and how the…
Read article →Build Your Own LLM, Lesson 9: Evaluating a Model — Loss, Perplexity and Test Prompts
Build Your Own LLM, Lesson 9: Evaluating a Model — Loss, Perplexity and Test Prompts
Measure a language model: cross-entropy loss, perplexity and what it means, validation sets, fixed test prompts, and why benchmarks and human checks are both…
Read article →Build Your Own LLM, Lesson 8: Sampling — Temperature, Top-k and Top-p
Build Your Own LLM, Lesson 8: Sampling — Temperature, Top-k and Top-p
How a language model picks the next token: greedy decoding, temperature, top-k and nucleus (top-p) sampling, implemented in a dozen lines and compared side…
Read article →Public Wi-Fi Safety: What Is Actually Risky (and What Isn’t)
Public Wi-Fi Safety: What Is Actually Risky (and What Isn’t)
Airport, café and hotel Wi-Fi explained honestly: why HTTPS protects most browsing, the real risks (fake hotspots, login-page phishing, file sharing) and simple habits…
Read article →Build Your Own LLM, Lesson 7: Train a Tiny GPT on Your Own Text
Build Your Own LLM, Lesson 7: Train a Tiny GPT on Your Own Text
Stack transformer blocks into a GPT, add token and position embeddings, train it on a text file with AdamW on a laptop CPU, watch…
Read article →Build Your Own LLM, Lesson 6: Multi-Head Attention and the Transformer Block
Build Your Own LLM, Lesson 6: Multi-Head Attention and the Transformer Block
Extend single-head self-attention to multi-head attention, add the feed-forward network, layer norm and residual connections, and test the causal mask — the full transformer…
Read article →Build Your Own LLM, Lesson 5: Self-Attention From Scratch
Build Your Own LLM, Lesson 5: Self-Attention From Scratch
Understand and code the attention mechanism behind transformers: queries, keys, values, the causal mask and softmax — first with NumPy on four tokens, then…
Read article →