I found the following list of resources worthwile and often reread and rewatch. I often refer friends and colleagues to these sources for various reasons. This repository collects and categorizes these sources together with a small explanations why they're so great.
Black box that shit..
Bob Coecke on QNLP The content in the Machine learning street talk podcast is of exceptional quality, but this particular episode is beyond epic because of the character of Bob Coecke and the cross section of sciences which come together here: quantum computing, deep learning, natural language processing, category theory and diagrammetic proof theory. Essential material for people who understand that the real science nowadays cross polinates and has a multidisciplanary frontier with Category Theory as an important enabler.
As a bonus, the grin on the presenter's face during the whole introduction is priceless and worth a study in it's own right.
The overarcing theme here is choose your constraints explicitly and formalize them using the language of logic, universal algebra and category theory.
Liberties constrain, constrains liberate by Runar Bjarnasson
Axis of abstraction by Paul Philips. If I just keep promoting the word adhocicity it will become a word hopefully.
Simple made easy by Rich Hickey. I like his slogan but he seems to be drowning in his own analytical philosphic word soup without coming to anything substantial.
Univalent foundations of mathematics A mathematician talking about why maths needs compilation and continuous intergration. Vladimir Voevodsky is called the father of Homotopy type theory but he seems to be gotten there from the angle of reproducable sience and mathematics. His introduction is about an bug in a groundbraking paper 'Algabraic cycles and higher K theory' which discredited the hole field of motovic cohomology( from 12 min onwards. Unsurprisingly, later it turned out there were much more bugs 😄. The rest of the talk is about bugs in mathematics and what can be done about that.
The unvivalent foundations are merely there to provide an alternative foundations of mathematics on which the rest of mathematics can be expressed so it's suitable for verification by computers. The key is there that the language of univalent foundations / HOTT is closer to how mathematicians express proof and theorems than existing alternatives at that time.
The attitude which Voevodsky ascribes to fellow mathematicians that formalizing proofs is not the real mathematics and that it is on the one hand trivial but also very hard resonates with me. It's prevalent in software engineering(refactoring a system often means fixing critical bugs causing data losss and coruption) and machine learning (productionizing data science code 🤦♂️) too. The overconvidence in one's ability to write correct code is a real problem and hinders the progress we make as an civilization. Where is our library of Alexandria? A resuable, verified collection of software and algorithms and their properties with mapping to different programming languages available from the IDE?
These are a collection of my favorite sources of contrarian views. Very wrong, but deliciously so..
Rob Pike's Functional pearls in go (or maybe not) Not even wrong considering the Universe Rob Pike lives in
On ambition AKA My boss made me write this 😸
Machine learning street talk If ever in need of an advanced ML content firehose, look no further than Machine learning street talk. In addition the editing and introductions of the guest and the technical background are exquisitit, really wonderfully crafted and edited.
[Lex Fridman's podcast)(https://www.youtube.com/c/lexfridman) Lex has a wide array of interesting guests. The reason the podcast is so great is the preparation and genuine interest of the interviewer in the deep questions that occupy the mind the guests.