16 Jul 2026
NLP
Deep Dive
The Byte Latent Transformer: Inside Meta's Vocabulary-Free LLM
How Meta's BLT matches tokenized LLMs at 8B/4T bytes using entropy-based dynamic patching — no vocabulary, no merge rules, and 8–15% better robustness on noisy text.
→
16 Jul 2026
NLP
Deep Dive
How Google Fixed Arithmetic: Inside Gemini's 256K Unigram Tokenizer
256K vocabulary, Unigram LM, split-digit tokenization, byte fallback — the four engineering decisions in Gemini's tokenizer and why each one exists to fix a specific failure mode in BPE.
→
16 Jul 2026
NLP
Visual Guide
Tokenizers, Explained Visually
Animations and interactive demos: watch BPE merge tokens live, see the token tax across six languages, understand entropy-based dynamic patching, and see why arithmetic fails.
→
16 Jul 2026
NLP
Research Review
The State of Tokenizers in 2026 — From BPE to Byte Latent Transformers
Every major lab doubled their vocabulary this year. Here's why, what each lab changed, where the token tax still bites, and why Meta's byte-level model may signal the end of subword tokenization.
→
14 Jul 2026
ERA v5 · S2
NLP
Interactive
Building a Multilingual BPE Tokenizer That Treats Four Scripts Fairly
One 10 000-token vocabulary for English, Hindi, Telugu, and Bengali — with a fertility spread of 0.034 and a score of 28 280. Five design choices that made the difference.
→
04 Jul 2026
ERA v5 · S1
Deep Learning
Interactive
Seeing Why Neural Networks Work — Four Live Proofs
Activations, depth, embeddings, and generalization — proved with live in-browser experiments. Watch a model learn to separate rings, watch embeddings cluster by meaning, watch memorization close as data grows.
→