What I’m playing with right now: sliders for controlling the look of a UI - except the sliders are actually **NUMBERS INSIDE THE MODEL** that fills out a form, based on which the UI is rendered.

There is a 0.00000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000001% chance that my shoe will kill us all in the next 10 years. But considering how improbable the emergence of life in this universe was… I think we should take the shoe seriously. 👟

I spent three months trying to shove mechanistic interpreatibility research into a production agent. I wrote up the whole thing on my Linkedin: www.linkedin.com/pulse/toward...

Towards shoving arXiv papers into a production system. At 3:47 a.m.

This is an English transcript of my keynote at CERNA.AI in Ostrava, 15 September 2026.

linkedin.com

Two random thoughts I cannot hold anymore: 1. AI can do tasks which takes a long time to senior dev, that famous METR plot. Great. Cgg. How many SUCH tasks do you have to solve during one year? Where are the customers who would use this amount of software?

Tried to write for LessWrong twice. Rejected by a moderator both times. Considering a third attempt. Maybe third time’s the charm. Or maybe I’m just too wrong for LessWrong.

People argue about what a language model is. A function. A next-token predictor. Something way more. Whatever it is, I still wanted to stand inside one. And walk through it. Literally. Now you can too. unt1l1f1nd-resi-doom.static.hf.space/index.html

I never stop being fascinated by the gap between academia and industry. “Look how useful it is.” “Look how interesting it is.” Both sentences are important. And yet one is heard in one field, and the other in the other. Sometimes I wonder: do these two worlds even know what the other one is saying?

I've wanted to have a conversation about AI with a philosopher for a long time. My philosopher friend isn't interested. He says there's no philosophy in AI. And philosophers who work in AI and claim otherwise? According to him, they're just sellouts. And I don't know what to make of it.

They kept saying activation steering can't run in production vLLM - hooks die in CUDA graphs. RhizoNymph's fork proved the fix: record the steering kernel into the graph, then steer by changing numbers in its buffers. I made it a pip plugin. Steered = vanilla speed. github.com/moudrkat/hotwire-vllm

GitHub - moudrkat/hotwire-vllm: Zero-overhead activation steering for vLLM - per-request steering vectors inside CUDA graphs. No fork, no enforce_eager, just pip install.

Zero-overhead activation steering for vLLM - per-request steering vectors inside CUDA graphs. No fork, no enforce_eager, just pip install. - moudrkat/hotwire-vllm

github.com

Anyone else here obsessed with LLM interpretability? I recently open-sourced a tool for observing model internals in your own AI application. It is hella cool. The model internals, not the tool. Okay, the tool as well. Give it a try! Contributors very welcome. github.com/moudrkat/bra...