OpenTube

The Algorithm That Made Modern AI Possible

StringsandTheory

161,370 views

Hi all, feeling good about this video as it is something I have been wanting to make for a long time. This can be seen as part 2 in my statistics series. Let me know what you want to see next!

Newsletter on Substack:
https://substack.com/@stringsandtheory?

🎥 Videos mentioned
Bayesian statistics: How to Predict ANYTHING
Discovery of GW250114: Why is everyone talking about this GW even...
Hawking radiation: Do Black Holes Have Temperature? | Hawking...

🔗 Resources for Bayesian statistics:
Bayes in the sky by R. Trotta: https://arxiv.org/pdf/0803.4089
Statistical methods of cosmology by Licia Verde: https://arxiv.org/pdf/0911.3105

Favorite books at the moment (amazon affiliate link):
⚫️Black holes: https://amzn.to/4hTfgdm
💻Machine learning: https://amzn.to/4gFRnVQ

Gear:
🎙️Microphone (it’s amazing!): https://amzn.to/3Z9wpsr

Ko-fi: https://ko-fi.com/stringsandtheory
If you want to support me and my work then consider buying me a "☕️" via Ko-fi and I'll be forever grateful :)

Business/sponsorship inquires✉️: [stringsandtheory@gmail.com](mailto:stringsandtheory@gmail.com)

Sub count📈: 3428 (Never thought I would reach a 100 subscribers...)

TIMESTAMPS:
0:00 -0:28 - Introduction
0:29 - 1:21 - Monte Carlo estimate of pi
1:22 - 2:32 - Markov Chains in action
2:33 - 4:20 - Bayesian inference
4:21 - 6:50 - Metropolis Hastings algorithm
6:51 - 7:52 - What do we do with the samples?
7:53 - END - Closing

20 Comments

@lipinglin1994 1 month ago

I love how the video just goes straight to the point.

@stevenstellman7035 5 months ago

I see that one of your commenters actually met Metropolis. I “met” him second-hand in 1968 as grad student at NYU doing Monte Carlo studies of polymer thermodynamics - he made my dissertation possible. Mal Kalos, the random numbers expert at Courant, showed me his famous paper (in the Journal of Chemical Physics by the way) which had only been out 15 years. It’s well known that Ulam “invented” MC which he applied to studies of neutron transmission, but few remember that its second important use was study of self-avoiding random walks as models of polymers – the first such paper was by the two Rosenbluths who were Metropolis’s co-authors but they gave it up when they learned of similar studies being done at Univ. of Illinois on the Illiac computer that had architecture based on von Neumann’s MANIAC computer. The early studies were plagued by high attrition. True to its nickname, Metropolis “importance sampling” focused on thermodynamically “important” regions of conformation space.

@guanxi99 6 months ago

Very well made! And many thanks for picking that super important and yet underrated topic. Out of curiosity: How is your job search going?

@kristm2010 6 months ago

Felt like I was watching a 3Blue1Brown or early days Vertiasium video! Amazing explanation and graphics, loved it:)

@Kandyfrom1986 5 months ago (edited)

This is a fantastic video. Thank you for the explanation. You connected all the necessary components, gave the intuition, and did not deviate from correctness (often simplification goes against it).
One small nitpick : During the metropolis hasting algorithm section, at the side, if you had drawn the markov chain during new theta exploration (similar to rainy, cloudy, sunny states) -- imho it would have improved that point by a slight margin.

@ratsxyz 6 months ago

As a researcher in AI myself I found the video to be excellent! Captured the essence of probabilistic modelling very well. Thank you!

@mine7777-e5y 6 months ago

really cool video! your storytelling and animations are top tier, it felt like a 3b1b video!!

@DxMedPhys 5 months ago

I've been watching you for a while and I think you've definitely improved in the quality of these educational videos! Understanding something is one factor but communicating it via video is a whole different task

@Nuclear_Mouse 5 months ago

Excellent! Very explanatory content. Can you clarify please what software you used to produce this animation?

@AminuRilwanu-t1m 5 months ago

Your videos are always so helpful, thank you!

@DuY-e8n 6 months ago

Loved the content, keep it coming!

@niktiktos 6 months ago

Nice, What matters most in this description is not the algorithm itself, although it is undeniably a brilliant tool, but the structure it moves through. Monte Carlo and Markov chains do not create the probabilistic reality, they merely reveal it. What truly guides the process is the underlying geometry of the space, just as in the circle example, where PI comes not from random sampling but from the relation between areas, or in parameter space, where the chain’s trajectory aligns with the density of the distribution. The algorithm is therefore only a way of interfacing with this structure, a local mechanism of movement, but it is the structure itself, often high-dimensional and compressible, that determines direction, stability, and meaning. Without it, the motion would be pure noise. With it, the process becomes discovery. And this is precisely the secret of AI, not the algorithm itself, but the structured reality it is able to approximate, compress, and navigate.

@johannesm8592 4 months ago

Can you make a video about how we went from MH to NUTS, and what the difference is between NUTS, Gibbs' and INLA

@GDOrtiz 6 months ago

I already knew everything you said, but it's always good to know that there are people creating informative videos like this. Keep up the good work.

@supergamer8030 6 months ago

thank you for this clear and beautiful video! <3

@adnanali7869 5 months ago

Trust the process 🙌

@Jakub-z6u 6 months ago

Excellent video, precise but not overwhelming and super clean graphics! Hope you make more videos like this (could you do information theory, the Bellman equation, RL which uses Markov chains too)?