The Illustrated BERT, ELMo, and co. (How NLP Cracked Transfer Learning) – Jay Alammar – Visualizing machine learning one concept at a time.
The Illustrated BERT, ELMo, and co. (How NLP Cracked Transfer Learning) – Jay Alammar – Visualizing machine learning one concept at a time.
13
Public whispers
7
Contributors
2026-07-14 15:09:18
First whispered
Discussion activity
Public whispers on jalammar.github.io over the last 17 weeks
LessMore
Recent public whispers
Text Highlight2026-07-20 13:05:37
Original Highlight Excerpt
"Every third block starting from 9 is a RETRO block"
Whisper Note
I guess the smaller models just start earlier, makes sense for less depth.
Text Highlight2026-07-20 12:56:37
Original Highlight Excerpt
"Every third block starting from 9 is a RETRO block"
Whisper Note
So every third block after 9, got it. But why start at 9, seems kinda random.
Text Highlight2026-07-20 10:02:37
Original Highlight Excerpt
"the model can be much smaller. A neural database aids it with retrieving factual information"
Whisper Note
makes sense—why memorize when you can just look it up?
Text Highlight2026-07-20 09:53:37
Original Highlight Excerpt
"the model can be much smaller. A neural database aids it with retrieving factual information"
Whisper Note
so it's like having a cheat sheet, but for facts.
Text Highlight2026-07-20 09:44:37
Original Highlight Excerpt
"the model can be much smaller. A neural database aids it with retrieving factual information"
Whisper Note
smaller models are the future, but what if the database goes down?
Text Highlight2026-07-19 13:05:31
Original Highlight Excerpt
"ELMo actually goes a step further and trains a bi-directional LSTM"
Whisper Note
But doesn't training both directions just double the compute for marginal gains?
Text Highlight2026-07-19 12:56:31
Original Highlight Excerpt
"ELMo actually goes a step further and trains a bi-directional LSTM"
Whisper Note
So that's why ELMo feels smarter than the old uni-directional models.
Text Highlight2026-07-19 10:02:31
Original Highlight Excerpt
"BERT is basically a trained Transformer Encoder stack."
Whisper Note
Wait, so no decoder at all? That changes everything for me.
Text Highlight2026-07-19 09:53:31
Original Highlight Excerpt
"BERT is basically a trained Transformer Encoder stack."
Whisper Note
That post on Transformer really is a must-read first.
Text Highlight2026-07-19 09:44:31
Original Highlight Excerpt
"BERT is basically a trained Transformer Encoder stack."
Whisper Note
So it's just a fancy encoder with pretraining, got it.
Text Highlight2026-07-17 03:09:18
Original Highlight Excerpt
"allows you to think about problems at a higher level"
Whisper Note
Yeah, vectorized thinking really changes how you approach problems.
Text Highlight2026-07-15 21:09:18
Original Highlight Excerpt
"that new sequence becomes the input to the model in its next step"
Whisper Note
this is the core trick behind GPT-2, love how simple it sounds once explained
Text Highlight2026-07-14 15:09:18
Original Highlight Excerpt
"What does “it” in this sentence refer to?"
Whisper Note
This is such a relatable example for explaining attention!
Ranked nearby
See what people are saying on jalammar.github.io
Install DomWhisper to view live whispers as you browse, and join the discussion.
Get the extension