How to live a happy life, according to the most elegant equation in humanity. (Hint: it's the math version of "omit needless words.") mikexcohen.substack.com/p/eulers-ide...

In my latest substack post, I show how to measure gender bias in open-source LLMs. The post comes with detailed explanations and a Python notebook file that you can easily adapt to any other biases. Curious? Come check it out :) mikexcohen.substack.com/p/gender-bia...

No comment.

Statistician: Someone who has their head in an oven, their feet in an ice bucket, and says "on average I feel fine."

If I had a million euros (that's 1.170686 million USD), I would do Daily Data Challenges to master machine-learning. Like this one, which involves running a t-test in Python to compare two random samples. Act like a millionaire. Solve Daily Data Challenges. mikexcohen.substack.com/p/ddc-15-two...

“King – man + woman = queen” sounds smart until you actually check the math. Spoiler: it’s fake news. mikexcohen.substack.com/p/king-man-woman-queen-is-fake-news

Do straight lines exist? open.substack.com/pub/mikexcoh...

I don't know what's going on in this AI image. But I do know this: Daily Data Challenge #25 is short and very practical for data science. You'll get to create a symmetric matrix and extract its unique elements. Can you solve it in less than 6 minutes? mikexcohen.substack.com/p/ddc-25-uni...

Here’s a picture of Thalorion The Destroyer reading my calculus textbook. I hope he likes it! Otherwise he might unleash doom on humanity. www.amazon.com/dp/B0DX6BPPZ8

Karl Gauss (yeah, *that* Gauss) sent me a fax this morning to say that he really enjoyed working through today's Daily Data Challenge. mikexcohen.substack.com/p/ddc-23-dis...

Why I left academia and neuroscience. This post on Substack has gained a lot of traction. I think many people identify with it. (Most of my posts are technical tutorials on machine-learning and LLM-mechanisms.) mikexcohen.substack.com/p/why-i-left...

Daily Data Challenge 8: Test for Pythagorean triplets A data challenge a day helps you master machine learning. More exercises and solutions at mikexcohen.substack.com youtube.com/shorts/LlqS2...

Bootstrapping: because sometimes asking “what if?” 10,000 times is a valid math strategy. open.substack.com/pub/mikexcoh...

genAI never fails to deliver the laughs.

Daily Data Challenges are going strong. Today's challenge is to create this beautiful graph using numpy and matplotlib. Can you do it? mikexcohen.substack.com/p/ddc-18-err...

This is not a skeet. mikexcohen.substack.com/p/this-is-no...

Have you done your Daily Data Challenge? mikexcohen.substack.com/p/ddc-8-test...

People talk a lot about GPT5 and hallucinations and trust and so on. But let's not forget that GPT2 is the ultimate arbiter of truth and science. Proof of my claim? This zinger from GPT2. Wanna explore GPT2's wisdom in your own private Python session? mikexcohen.substack.com/p/lessons-on...

GPT2 is hilarious.

Dear bluesky neuroscientists: You should pay attention to the field of LLM mechanistic interpretability. Not because LLMs are like real brains (they are absolutely nothing alike), but because mech interp analysis methods would be useful in neuroscience. Start learning here: tinyurl.com/48pum4pb

One of my fav quotes: "The way you do anything is the way you do everything." The problem with these pithy faux-philosophical quotes is that they don't stand up to scrutiny. For example, I definitely do not pick my nose with the same rigor, depth, and intensity with which I write 600-page textbooks

Can you code the Fourier transform from scratch? I bet you can. Spend 20 minutes on the Substack post, and you'll learn everything you need to understand and code the FT in Python. mikexcohen.substack.com/p/the-fourie...

Do you have enough confidence? I am, of course, talking about confidence intervals, the statistical measure of uncertainty around a sample characteristic like the average or correlation. Here's my latest series on Substack about confidence intervals: mikexcohen.substack.com/p/confidence...

Daily data challenge. 10 minutes a day to improve your coding, data science, and visualization skills. open.substack.com/pub/mikexcoh...

Happy teacher's day (in India)!

Have you done your Daily Data Challenge? mikexcohen.substack.com/p/ddc-3-do-t...

Why do people keep claiming that LLMs are "just linear algebra"? Most of the magic happens at the nonlinearities between the linear transformations. If you're going to be angry at math, at least be angry at the "non" part of it 🤔

Here's a neat thing: You can use machine-learning analyses to understand how large language models work. Learn how here: mikexcohen.substack.com/s/ml-on-llms

Why do artificial neural network models like LLMs expand the data dimensionality? It's because nonlinearly separable features can become linearly separable in higher-dimensional space. mikexcohen.substack.com/p/llm-breakd...

So weird that we have to drink 8ish cups of water a day. What did humans do before cups were invented??

Dear LLM mechanistic interpretability teachers: If you think the course linked below might be helpful for your courses, let me know! Discounts or free coupons can be organized 👍 www.udemy.com/course/dullm...

These heatmaps show how the distributions of Q, K, and V vectors inside a GPT transformer change as the text travels deeper into the language model, and picks up new contextual information. Wanna learn more? Check my Substack post (comes with Python code!) mikexcohen.substack.com/p/llm-breakd...

Be who you are and say what you feel, because those who mind don't matter and those who matter don't mind. (You've earned my respect if you know the source without looking it up!)

Have you done your daily data challenge? substack.com/@mikexcohen/...

Joseph Fourier was a French physicist, mathematician, and inventor of Cirque du Soleil (ok, just kidding about that last one).

Dear scientists of BlueSky: Wouldn't it be great to learn how large language models work by running experiments on them and analyzing the data? Yeah, you can do that, and I'll show you how. mikexcohen.substack.com/s/ml-on-llms

Have you done your Daily Data Challenge? mikexcohen.substack.com/p/ddc-1-viol...

I read recently that I can use these posts to share my unfinished thoughts, so therefore I

I manipulated GPT2's internal representations to prefer "Berlin" vs. "Paris." You can do it too! Check out my Substack post to find out how. mikexcohen.substack.com/p/llm-breakd...

I like turmeric, but I dislike that it has two "r"s. (I should really stop posting about tuumeric.)

I think more people would eat turmeric if it were purple.

Have you seen those neat-looking text heatmaps that show how LLMs like Claude or GPT interpret text? They're surprisingly easy to create in Python. open.substack.com/pub/mikexcoh...

"Omit needless words." Strunk was trying to tell you how to live a good life, not just how to write a good sentence. mikexcohen.substack.com/p/strunk-and...

So you wanna manipulate a GPT to give the output you want? I show how to do that in part 2 of my LLM breakdown series. (note: this is for small open-source models, not ChatGPT5 🥲) mikexcohen.substack.com/p/llm-breakd...

FYI: LLMs don't think or reason. They recognize patterns in token ("tokens" are numbers that represent words), and then generate new text that follow those patterns. When LLMs seem logical, it's because they been carefully trained to identify and reproduce patterns in human-written logical texts.

I started a Substack :) mikexcohen.substack.com I'm experimenting with "long-form" writing (1-2k words per post). Honestly, after writing 600-page textbooks, 2k words feels very short. Anyway, feel free to stop by my Substack and check out what's there so far.

Hey look at this! My new 600-page calculus book has landed: www.amazon.com/dp/B0DX6BPPZ8 And the git repo for Python code to reproduce all book figures and exercise solutions: github.com/mikexcohen/C... YouTube video explanations of all 162 exercises: www.youtube.com/playlist?lis...

Somehow, owning a printer seems like a very adult thing to do. That's why I had to get rid of mine. Also, planned obsolescence.

There's nothing that can't be cured (or at least, a little bit ameliorated) by chocolate-covered raisins.

It turns out that adding sweet pomegranate cordial to coffee is not as good as you might think.