Disnocen
Cryptographer Engineer at Blockstream
Intro
This is my website. It is simply updated through a google docs file!
I also have a previous, sparingly used blog here, but I may not really believe everything I say there. Consider this website the most updated one.
My X account: sparingly used, maybe I should tweet more. You can write me on telegram to start a convo
Through Ctrl +f you get the full experience of searching a website!
2026-08-04
Opencode is awesome
Is codex good or is the model?
We need harness benchmarks
Nvim is fantastic too
Like the real IDE. With LazyVim it just works. Sure it's bloated or whatever the cool kids say nowadays. Then they use VScode. Still it works, it's fast and has everything
Even rust language server works awesomely, better than VScode from a keyword and UX stand point
Constraining the search space and choose tradeoffs
I love programming and creating sketches. I love to see things happening. LLMs are awesome for it. Just never let them create something human-readable. That is so wrong it will make you lose faith in the whole AI thing
LLMs are for coding, and basic search engines. Some kind of knowledge tourism, that’s it. You can not use them to make decisions, real brainstorming etc. you either know what you talk about, or you’ll get confused. Google Docs, pieces of paper or some kind of e-ink tablet to write are the tools for thinking. You need to know your tradeoffs (kyt). You explore the new field, can’t ask LLMs for “summaries”
Did you really read what an LLM produces in plain english? Have you tried to read something well written like The History of the Decline and Fall of the Roman Empire by Gibbon? I am not an expert, but the latter seems to me real writing.
2026-08-03v2
Gpt 5.6. Sol works very well I must say. Working on consensus right now, simulating basic BFT in case of network problems + byzantine behaviour. Fun weekend project I must say.
Just never let the LLM plan, really. It’s not good. Too much details and too little thinking/overview and understanding of consequences at the same time.
I wonder now if my job is just being a good planner. Writing code is obsolete. Fortunately I am a researcher by training, but can I be an expert in all parts of a project?
---
Codex is my sysadmin now. Moved me from dwm to hyprland in minutes, and everything works. Created a repo for my dotfiles and it works there too. All that while working. A couple of tmux sessions had everything done for me. Controlled from the phone, of course.
It helps on a lot of things, and I use it at like 10% of their potential. It feels like re-learning how to use my Linux PC from scratch.
I have been very bearish on LLM last months, mostly for a combination of bad models, bad use on my part and the fact that LLMs simply can not write. I fell for the “knowledge base” thing and produced documents writing meaning less things. Man, you have to really guard that llm, you know? You leave it alone writing a 10 pages “research paper” and you find it sprouting “scaffolding” and “security boundary” and |invariant” all over the place.
----
Today another AI-based problem with a piece of Bitcoin infrastructure, after the ColdCard debacle
2026-08-03
First entry
Wow! Unexpectedly the thing is working! What a wonderful thing!
It also has autoupdate. I can not believe how easy it was
So easy to update. Never thought about that easiness
The difficulty is in having a VPS now it seems