Hacker News: the_origami_fox

New comment by the_origami_fox in "Moldova broke our data pipeline"

the_origami_fox — Tue, 03 Mar 2026 07:09:09 +0000

That's a cool map.

New comment by the_origami_fox in "IDF killed Gaza aid workers at point blank range in 2025 massacre: Report"

the_origami_fox — Fri, 27 Feb 2026 07:10:10 +0000

Unfortunately this article and the conversation here is taken in the context of the entirety of Israel is bad, the whole military war against Hamas and its patrons was bad, and that there is a genocide in Gaza. It leaves little room to have meaningful conversations about difficult conditions in the war and whether soldiers acted incorrectly or not and if they should be tried or not with regards to this single incident.

New comment by the_origami_fox in "Top UN legal investigators conclude Israel is guilty of genocide in Gaza"

the_origami_fox — Wed, 17 Sep 2025 08:22:45 +0000

There are many issues raised with the report, including it omits the invasion by the government of Gaza, Hamas, on 7 October 2023 entirely, and it omits that the Israeli army is fighting the army of Gaza, the Qasam brigades, who had 40,000 salaried fighters (pre-war), have fired thousands of missiles, developed hundreds of kilometres of tunnels specifically for urban warfare, and subverted public and private infrastructure for urban warfare. For such a serious allegation, it is important to consider and address all aspects and not simply omit them.

I would like to add, I don't think this topic is appropriate for Hacker News.

DeepSeek's Multi-Head Latent Attention

the_origami_fox — Mon, 03 Mar 2025 07:42:05 +0000

Article URL: https://liorsinai.github.io/machine-learning/2025/02/22/mla.html

Comments URL: https://news.ycombinator.com/item?id=43239248

Points: 4

# Comments: 1

New comment by the_origami_fox in "One Genius' Lonely Crusade to Teach a Computer Common Sense (2016)"

the_origami_fox — Wed, 09 Oct 2024 07:35:45 +0000

2016

Micrograd.jl

the_origami_fox — Wed, 28 Aug 2024 07:24:42 +0000

Article URL: https://liorsinai.github.io/machine-learning/2024/07/27/micrograd-1-chainrules.html

Comments URL: https://news.ycombinator.com/item?id=41376893

Points: 154

# Comments: 49

Covering All Birthdays

the_origami_fox — Sat, 27 Jul 2024 09:44:32 +0000

Article URL: https://liorsinai.github.io/mathematics/2024/07/09/birthday-covering.html

Comments URL: https://news.ycombinator.com/item?id=41085471

Points: 42

# Comments: 22

Covering All Birthdays

the_origami_fox — Sun, 14 Jul 2024 11:29:41 +0000

Article URL: https://liorsinai.github.io/mathematics/2024/07/09/birthday-covering.html

Comments URL: https://news.ycombinator.com/item?id=40960290

Points: 1

# Comments: 0

Generative transformer from first principles in Julia

the_origami_fox — Fri, 05 Apr 2024 07:26:01 +0000

Article URL: https://liorsinai.github.io/coding/2024/03/23/transformers-gpt.html

Comments URL: https://news.ycombinator.com/item?id=39939592

Points: 6

# Comments: 0

New comment by the_origami_fox in "Radix Tree in Julia"

the_origami_fox — Fri, 22 Mar 2024 08:37:42 +0000

Author here. While researching this topic I came across this article https://www.geeksforgeeks.org/compressed-tries/ by GeeksForGeeks (compressed trie is another name for radix tree). Currently it is one of the top results on Google for this topic. Generally I've found GeeksForGeeks to be a good resource, but this article is especially poor. The code is bizarre and makes no sense:

- It defines properties like 'bitNumber', 'leftChild' and 'rightChild' which have nothing to do with this type of tree.

- The search function search(k) calls an overloaded version of itself search(root, k) that is (1) illegal in these languages and (2) not defined.

- The insert function does not account for all scenarios.

And more.

I'm fairly convinced this was written with the assistance of a LLM. That would be a logical explanation to all those illogical choices.

Radix Tree in Julia

the_origami_fox — Fri, 22 Mar 2024 08:34:58 +0000

Article URL: https://liorsinai.github.io/coding/2024/03/21/radix-tree.html

Comments URL: https://news.ycombinator.com/item?id=39788540

Points: 2

# Comments: 2

New comment by the_origami_fox in "Show HN: Name That Nation"

the_origami_fox — Tue, 06 Feb 2024 08:14:58 +0000

The pictures come straight from Wikipedia. From the picture for North Korea (https://en.wikipedia.org/wiki/North_Korea): Territory controlled by North Korea shown in dark green; territory claimed but not controlled shown in light green.

The Weiler-Atherton polygon clipping algorithm

the_origami_fox — Fri, 13 Oct 2023 11:55:25 +0000

Article URL: https://liorsinai.github.io/mathematics/2023/09/30/polygon-clipping.html

Comments URL: https://news.ycombinator.com/item?id=37869579

Points: 28

# Comments: 2

New comment by the_origami_fox in "Pythagorean Theorem found on clay tablet 1k years older than Pythagoras (2009)"

the_origami_fox — Fri, 06 Oct 2023 06:54:48 +0000

And if you choose b=x/10, x/100, x/1000 ... then each iteration will add the exact decimal point.

New comment by the_origami_fox in "Pythagorean Theorem found on clay tablet 1k years older than Pythagoras (2009)"

the_origami_fox — Thu, 05 Oct 2023 10:17:13 +0000

For anyone wondering how they got the approximation sqrt(2)=1+24/60+51/60^2+10/60^3.

It's based on the simple idea that:

     Z = (a + b)^2 = (a^2 + (2a+b)*b)
  => (2a+b)* b < Z-a^2

Given an initial estimate "a", we need to find the largest "b" such that the term on the left is less than the term on the right. Therefore our estimate will always be slightly less than the actual answer and we can repeat the process to get slightly closer.

For the first iteration, Z=2 and a=1. We choose b=x/60:

  (2+x/60)*x/60 < 2-1^2
  120x + x^2 < 3600
  x = 24 ... 3456 < 3600
  x = 25 ... 3625 > 3600

So our first term is 24/60.

Repeat with a=1+24/60 and b=x/60^2:

  (2(1+24/60)+x/60^2)*x/60^2< 2-(1+24/60)^2
  10080x+x^2 < 518_400
  x = 51 ... 516_681 < 518_400
  x = 52 ... 526_864 > 518_400

Repeat multiple times.

Writing this in code I can easily get: 1;24,51,10,7,46,6,4,44,50,28 = 1.4142135623730951

This whole process can be codified into the long division algorithm for square roots which works quite neatly with base 10.

Edit: formatting

Implementing the Gzip-kNN Classification Paper

the_origami_fox — Thu, 17 Aug 2023 13:28:04 +0000

Article URL: https://liorsinai.github.io/mathematics/2023/08/13/gzip-knn.html

Comments URL: https://news.ycombinator.com/item?id=37161235

Points: 1

# Comments: 0

New comment by the_origami_fox in "Ask HN: Could you share your personal blog here?"

the_origami_fox — Thu, 06 Jul 2023 09:13:42 +0000

https://liorsinai.github.io/ long detailed posts on mathematics, machine learning and algorithms. Most of the code is in Julia but older posts also feature Python, C# and C++. I recently did posts on Transformers and Diffusion Denoising Probabilistic models.

The most popular post is a Sudoku solver in Python.

How (not) to compare 2D scatter plots

the_origami_fox — Mon, 19 Jun 2023 20:27:21 +0000

Article URL: https://liorsinai.github.io/mathematics/2023/05/13/similarities-2d.html

Comments URL: https://news.ycombinator.com/item?id=36396071

Points: 3

# Comments: 0

New comment by the_origami_fox in "Denoising Diffusion models from first principle in Julia"

the_origami_fox — Sun, 08 Jan 2023 07:28:30 +0000

Auuthor here. Will update.

New comment by the_origami_fox in "Denoising Diffusion models from first principle in Julia"

the_origami_fox — Sun, 08 Jan 2023 07:27:10 +0000

Author here. I have noticed similar behaviour. As part of this exercise I tried to train a model to generate Pokemon based on This Pokemon Does Not Exist by HuggingFace. However my models only converged to nosiy smudges after 50 iterations and so I excluded it from the posts (I do mention my experiments at the end of part 2).

My first assumption was that the mdoel I was training was too small: 13 million parameters as opposed to the 1.3 billion in ruDALL-E (not sure how much of this is only the diffusion model). So that's a 100x smaller. I want to experiment with upscaling it.

Reading this I'm wondering if there's more I need to do. For example, training a conditioned model - "cheat" by given it in the index of the Pokemon during training but then you sample without an index - or make the model predict the standard deviation (beta tilde). Or as you say, work with loss functions.

More work to be done here.