gum

Mythos is good but not infallible

Sun, 07 Jun 2026 00:00:00 +0000

A note on a mistake in Proposition 7.1 of Mythos’s Erdős unit-distance writeup.

Thoughts on the Mythos system card with a specific eye on cyber capabilities

Fri, 10 Apr 2026 00:00:00 +0000

Crosspost: thoughts on Mythos’s claimed cybersecurity implications.

A mildly cursed 3.5× triton `tl.reduce` optimization

Fri, 05 Sep 2025 00:00:00 +0000

TL;DR: For small, compile-time K, manually unrolling a 3D→2D bitwise-OR reduction can beat tl.reduce by ~3.5×.

I’ll take you on a small adventure of some weird Triton compiler behaviour.

We’ll look at a reduction used inside an attention variant that reduces a 3D tensor along the last axis with bitwise OR to produce a 2D tensor and the weird stuff I encountered while doing that.

Concretely given an integer tensor

llama2.c running on original iPhone (240k)

Thu, 10 Jul 2025 22:00:00 +0000

240k LLM runs on the very first iPhone.

How I made PrimeIntellect's toploc 70x faster WIP

Sun, 23 Feb 2025 19:38:00 +0000

tbd. U can read up much in the comments of the code and pull-request here https://github.com/PrimeIntellect-ai/toploc/pull/3