Thanks to visit codestin.com
Credit goes to dstekanov.substack.com
Dmytro Stekanov
Subscribe
Sign in
Home
Notes
Brain
Artificial Intelligence
Archive
About
Transformer Language Model from Scratch
What I learned implementing attention, RoPE, and SwiGLU from first principles
Jan 7
•
Dmytro Stekanov
Codestin Search App
Codestin Search App
Codestin Search App
[BPE] encode and decode
Implementing GPT-2 compatible encode/decode in pure Python
Dec 11, 2025
•
Dmytro Stekanov
Codestin Search App
Codestin Search App
Codestin Search App
[BPE] Optimizing BPE Tokenization: A Case Study in Algorithms and Systems
Part 3: From O(P) to O(log P) - the final frontier
Dec 11, 2025
•
Dmytro Stekanov
Codestin Search App
Codestin Search App
Codestin Search App
Most Popular
View all
[BPE] Scaling Up: Parallelization
Nov 11, 2025
•
Dmytro Stekanov
Codestin Search App
1
Codestin Search App
Codestin Search App
Building a Tokenizer From Scratch: A Journey Through Stanford’s CS336
Oct 29, 2025
•
Dmytro Stekanov
Codestin Search App
1
Codestin Search App
Codestin Search App
When your tokenizer meets reality: building BPE v2
Nov 10, 2025
•
Dmytro Stekanov
Codestin Search App
1
Codestin Search App
Codestin Search App
[BPE] Scaling Up: File Processing
Nov 11, 2025
•
Dmytro Stekanov
Codestin Search App
1
Codestin Search App
Codestin Search App
Why BPE tokenizers use UTF-8 instead of UTF-16 or UTF-32
Oct 27, 2025
•
Dmytro Stekanov
Codestin Search App
1
Codestin Search App
Codestin Search App
1
Dot Product: the simple operation that bends a complex world
Oct 2, 2025
•
Dmytro Stekanov
Codestin Search App
Codestin Search App
Codestin Search App
1
Latest
Top
[BPE] Scaling Up: Parallelization
Part 2: Parallelizing Pre-tokenization and Optimizing the Merge Step
Nov 11, 2025
•
Dmytro Stekanov
Codestin Search App
1
Codestin Search App
Codestin Search App
[BPE] Scaling Up: File Processing
Part 3 of the Tokenizer Series
Nov 11, 2025
•
Dmytro Stekanov
Codestin Search App
1
Codestin Search App
Codestin Search App
When your tokenizer meets reality: building BPE v2
Or: Why production tokenizers are nothing like the textbook version
Nov 10, 2025
•
Dmytro Stekanov
Codestin Search App
1
Codestin Search App
Codestin Search App
Building a Tokenizer From Scratch: A Journey Through Stanford’s CS336
Or: How I stopped memorizing and started understanding
Oct 29, 2025
•
Dmytro Stekanov
Codestin Search App
1
Codestin Search App
Codestin Search App
Why BPE tokenizers use UTF-8 instead of UTF-16 or UTF-32
Inside tokenizer training: why UTF-8 gives cleaner signals and faster learning than UTF-16 or UTF-32.
Oct 27, 2025
•
Dmytro Stekanov
Codestin Search App
1
Codestin Search App
Codestin Search App
1
Conditional probability - how to see what we usually don't see
We often evaluate the world directly.
Oct 12, 2025
•
Dmytro Stekanov
Codestin Search App
Codestin Search App
Codestin Search App
2
Dot Product: the simple operation that bends a complex world
Understanding the Dot Product: Math, Intuition, and AI Applications
Oct 2, 2025
•
Dmytro Stekanov
Codestin Search App
Codestin Search App
Codestin Search App
1
See all
Dmytro Stekanov
My personal Substack
Subscribe
Dmytro Stekanov
Subscribe
About
Archive
This site requires JavaScript to run correctly. Please
turn on JavaScript
or unblock scripts