Felix (aka @PewDiePie) is low key at the frontier of AI research building himself a "council" of competing LLMs that have a genetic-algorithm-like process for replacement. His "tech" phase is honestly amazing. From Linux -> GPUs -> AI, so good 🔥 piped.video/watch?v=qw4fDU18…

New video out! 🥳 piped.video/YxNh26RQC4w

Tacit Talk Episode 28: Co-dfns & AI with Aaron Hsu @arcfide is out now! 🥳 piped.video/watch?v=-PMCkAB_…

Tried out Co-dfns APL compiler for the first time today. For a simple softmax calculation, it generated 3.7M CUDA kernels... that is... pathologically bad 😳 @arcfide 🤔🤔🤔

I'll be seeing this live in a month @cppunderthesea but I couldn't help myself - 20 minutes in and it is great 🥳 (Are We There Yet? - @seanparent) piped.video/watch?v=RK3CEJRa…

This is a great talk by @mattgodbolt ... but his conclusion on C++20 modules: "Build times went from 10 sec to 40 sec... I was expecting segfaulting compilers ... [but] I got it to work pretty quickly... I think that is a great place" 🤔🤦‍♂️🤔 piped.video/watch?v=jlt_fScV…

Python: The Documentary, for those interested piped.video/watch?v=GfH4QL4V…

Maybe other array folks can do better. The funny part is the initial C++ / Rust code is my code from a YouTube video I made 2 years ago. And the C code is cheating - it is using a constant time formula. piped.video/watch?v=wGCWlI4A…

Plugged into the AI matrix, fixing @Github search 🤩 (this is just the mth iteration of the nth iteration)

Honestly, @OpenAI's gpt-5 / gpt-5-fast work so terribly with Cursor. @AnthropicAI's Claude 4 Sonnet it vastly superior. The gpt models are always doing odd things and trying to use tools that aren't installed and it floods the tracing and makes it useless.