How 3Blue1Brown fits a message between zero and one
The third part of Grant Sanderson's information theory series passed a million views, and the U.S. Constitution's preamble serves as a test case.
This algorithm changed how I think about compression | Information theory, part 3
Up next
Why this story? The channel's weekly summary: what went up this week, what is drawing views now, and where the channel stands among similar channels.
Things we can learn from the video
Press to jump
- 5:08A language model estimates which character is likely to come next based on the text it has already seen.
- 8:40Each letter of "math" narrows an interval according to its predicted probability.
- 11:08A number inside the final interval and the message length let a decoder reconstruct the text.
- 16:57The bits needed to specify that number closely match the message's information value under the model.
- 22:50NanoGPT and Qwen are tested as compressors on samples including the U.S. Constitution's preamble.
- 35:17Sanderson argues that compression isn't the same as intelligence, but improving it draws on prediction and adaptation.
Grant Sanderson starts with four letters: "math." In "This algorithm changed how I think about compression | Information theory, part 3," each letter narrows a stretch of the number line. How much room a letter gets depends on how likely it is to come next. By the end, a number inside the final sliver, sent along with the message length, is enough for a decoder using the same model to recover the word.
That's arithmetic coding, and it gives Sanderson a way to show what a language model's predictions are worth: the better it predicts the text, the fewer bits it takes to send it. The comments took that idea outside computing. One viewer thought of familiar proverbs you can leave half unsaid. Another brought up chess masters remembering positions from real games more easily than randomly arranged boards.
The two uploads went up on the same day and are now drawing the most daily views on 3Blue1Brown. The compression lesson reached a million first, while "Encoding the encyclopedia on a rod" is approaching half a million. The older "Reinventing Entropy | Information theory, part 1" is also among the channel's fastest-moving videos, so there's a natural place to start if part three sent you looking for the groundwork.
3Blue1Brown added subscribers this week, putting it among the education channels in its group that grew, alongside Simplilearn and MIT OpenCourseWare. Johnny Harris and Tom Scott held steady on subscribers. Sanderson's channel is drawing about five times its usual daily views, with both new uploads near the top of its own viewing list.
Published on Trari
Where next?
A professional in this field? Want to write about Science & learning on Trari? Tell us ›
What viewers say
It's like how, in natural speech, we only say half of many proverbs, implying its ubiquitous ending.
@234930See it on YouTubeReminds me of how, when grandmasters look at positions in chess, they can memorize positions far faster than random people, but only if the positions come from real positions rather than randomly set up.
@ursumajor1See it on YouTubeI can't tell you how small all of Physics 1 and 2 feels to me after I've taught them for 10 years, 7 years as a TA and 2.5 as a professor. They feel absolutely small to me because of the compression and decompression I can do by derivation from a few rules.
@IsYitzachSee it on YouTube
2 more comments
Incredible reflections on intelligence toward the end. The idea that you understand something better when it becomes smaller in your mind really matches what the "click" feels like. A pile of lonely facts suddenly refactoring into a few generating ideas.
Schmidhuber has this theory that the good feeling is the compression progress. So how fast things are getting smaller. Also kind of explains curiosity. Already compressed things are boring, pure noise is boring, and the interesting stuff is in between.
This video did a lot of compressing for me, thank you.@EeroAlvarSee it on YouTubeI felt so much joy at 25:00 when he said "hold on to that" and i realised that there are 10+ more minutes in the video
@FlutterstormSee it on YouTube
Public comments from the video's page on YouTube
Want to know when somebody answers you?
The videos drawing views on 3Blue1Brown right now
- Channel
- 3Blue1Brown
- Subscribers
- 8,770,000 ▲ 20,000 this week
- Channel views, last 7 days
- 2,637,922
- Videos published this week
- 2
- Video published
- November 17, 2026
- Video views so far
- 1,027,502
- Views on its first day
- 570,632
The numbers
Here's how 3Blue1Brown's views and subscribers moved this week.
💡 Test your knowledge
Which word does Sanderson use to demonstrate arithmetic coding letter by letter?
Sanderson uses the four letters of "math" to show how each new character narrows the interval.




