Notes
Public ActivityPub notes from the Theor Fediverse account.
This page is heavily WIP, and UX / UI isn’t fully ready yet.
Notes are syndicated to home.theor.net via https://github.com/theoryzhenkov/repo.tangent; I also syndicate to Twitter, Bluesky and Fediverse at @theor@theor.net.
I got accepted into Surplus! Very excited. Now, hopefully, J1 US visa is not going to take me, a Russian, forever, and I will be off to build my cooperation / coordination tooling.
Unfortunately also overlaps a bit with my staff duty on ESPR (https://espr.camp/) this year.
Finished "The Dispossessed" today. A lot has already been said about it, I'm not to repeat; but to speak on my experiences and questions, I wonder what could improve the memetic fitness of Anarresti society.
It seems to me, by the end of the book, that the walls that are about to crumble down are protecting Anarres from possession. Shevek revolts against the flaws and rot of his society, but is he about to bring only more misery to his people?
How could a society protect itself from manifestation of authority, from Sabuls of all kinds, especially over spans of hundreds of years, to do not allow Sheveks inorder to protect itself?
I hypothesise that one way to do so is with intellectual uplift and establishment of culture of rational thought. Then every member of society could derive themselves the wider consequences of their actions. Instead of opaque cultural memes, explicit intellectual ones.
My birthday features a 300-light-years annihilation beam emitted from the core of a black star that explodes stars. What sort of omen is that?
https://science.nasa.gov/specials/apps/what-did-hubble-see-on-your-birthday/#
The fact that human science doesn't publish no-op results is likely a significant limiting factor for AI capabilities.
When I worked on cybersecurity evaluation harnesses for LLMs one of the main issues models experienced was a lack of exploratory behavior. LLMs tend to "lock in" to a first hypothesis, often never exploring other possibilities.
I hypothesize this is because there is no reasoning *processes* in their training, only final, positive result state. All that LLMs know is that ideas turn out to be correct.
I'm surprised accelerationists haven't yet launched a "journal of failed ML science" to improve automated AI research capabilities.
#theor_note #theor_ml #theor_proj_wonntext
My WONNText architecture significantly outperforms Transformer on generalising to unseen arithemtics after WikiText > two-digit arithemtic curriculum training at 2.6M parameter count and using default attention mechanism.
I have further plans at scaling this approach, generalising to more advanced mathematical reasoning, and running ablation tests. If anyone is willing to sponsor (~200 USD) my compute needs, write me at proj+wonntext@theor.net.
Original WONN research by @YueSong48287250 and Jiawen Dai: https://arxiv.org/pdf/2605.20922.
---
UPDATE 30.06.2026
I have made a mistake matching parameters instead of FLOPs on these different architectures. Scaling transformer to match FLOPs actually shows WONN network underperforming.
Cue is my homebrewed generative art project. It uses SDF to power efficient OpenGL shaders used to flood gradients. Cue comes with BYOK sentiment analysis that powers hyperparameters of the generated images, reflecting arousal, focus and valence of the prompt.
Take Cue for a spin here: https://cue.theor.net/
Cue is my homebrewed generative art project. It uses SDF to power efficient OpenGL shaders used to flood gradients. Cue comes with BYOK sentiment analysis that powers hyperparameters of the generated images, reflecting arousal, focus and valence of the prompt.
Take Cue for a spin here: https://cue.theor.net/
Cue is my homebrewed generative art project. It uses SDF to power efficient OpenGL shaders used to flood gradients. Cue comes with BYOK sentiment analysis that powers hyperparameters of the generated images, reflecting arousal, focus and valence of the prompt.
Take Cue for a spin here: https://cue.theor.net/
My ADHD provides an incredible boon towards gathering first-person evidence on "side effects of regularly overdosing on every pill I have prescribed".
With no prior prompting, my AIs running pi_intercom extension to talk to each other over different git workspaces started calling each other "siblings". They even differentiate older and younger ones. Fascinating!
I have subscribed to [Umans](https://app.umans.ai/offers/code) for unlimited GLM-5.2 / Kimi inference at only 50 USD / month. Limited "founding member" offer with crazy value. GLM-5.2 is very fast, and as capable as Opus 4.8.
Hand-rolled a small POSSE tool for syndicating my notes at theor.net across Bluesky, AP and Twitter. Might be useful to people here on Bluesky to make multi-platform presence easier: https://github.com/theoryzhenkov/repo.tangent.
Wired to be theor.net specific, but can easily be rewritten for your domain. @bsky.app, when will AppView stop ignoring putRecords? It even goes through successfully! Allow us to edit our notes /angry.
#theor_note #theor_engineering_at_protocol #theor_engineering_ap_protocol #theor_engineering_web
#theor_note #theor_engineering_web
My TOC at theor.net is one the best TOC implementation I have seen. Should I publish it as a React component so that people adopt my watcher pattern?
It carefully judges what is visible on the screen, and highlights all sections you are currently viewing based on the proportion of the sections currently visible on the screen.
Source code: https://github.com/theoryzhenkov/repo.home.prj_theornet/blob/main/src/scripts/toc-scrollspy.ts
#theor_note #theor_engineering_agents
Surprisingly effective, as a tactic, instead of using context compaction, to just clear the session and point the agent at the old logs. Acts like a proper reset, model feels fresher and more sane than after summarisation.
No notes match this filter.