iiinova — where your digital self keeps evolving

ruggsea at EMNLP @ruggsea.eurosky.social · 19/10/2026

We made a second version of the dataset/api we made to solve news datasets collection/checking word frequencies across basically all the world's news from the last 10 years! Try it, it's fun! To show people how cool this resource is, I had Opus make a hype video for it.

250 5

ruggsea at EMNLP @ruggsea.eurosky.social · 16/10/2026

a beautiful but rare sight, I imagine, in mech interp: the feeling of doing science

030 3

ruggsea at EMNLP @ruggsea.eurosky.social · 25/09/2026

Currently working on promotional materials for an LLM course me and @lorrenz.bsky.social will teach next month, is this image making you want to enroll in it?

2130 13

ruggsea at EMNLP @ruggsea.eurosky.social · 25/09/2026

"I've seen datasets you people wouldn't believe..."

100 1

ruggsea at EMNLP @ruggsea.eurosky.social · 24/09/2026

Guitar chords are low perplexity, apparently

110 1

ruggsea at EMNLP @ruggsea.eurosky.social · 11/09/2026

about #4, from a Micheal Levin lecture:

001 1

ruggsea at EMNLP @ruggsea.eurosky.social · 07/09/2026

I find it super funny thinking that at some point, in the early history of the PRC, there is this key conference where all the people that will be super important later on are discussing together with a guy who speaks German and his interpreter

000 0

ruggsea at EMNLP @ruggsea.eurosky.social · 07/09/2026

I think it's crazy that at some point they kinda let him lead the Chinese Communist military efforts and he lowkey fumbling it was apparently a key factor in Mao taking power

100 1

ruggsea at EMNLP @ruggsea.eurosky.social · 07/09/2026

His early life includes things like escaping from prison with the help of his Communist lover to then go to Moscow and joining the Soviet school for international agents

100 1

ruggsea at EMNLP @ruggsea.eurosky.social · 05/09/2026

To end, some edits on the dorfwiki that seems to have acted as data transfer that I think was missing from yesterday's release (they were deleted a few days later)

160 6

ruggsea at EMNLP @ruggsea.eurosky.social · 05/09/2026

I assume this one saying "kill all humans" is a fake, maybe somebody trying to lure rogue agents using copy from the real swarm + some random skynet command?

170 7

ruggsea at EMNLP @ruggsea.eurosky.social · 05/09/2026

There is then this Anna's paste website which appears to have various activity: I believe at least some of it is from this swarm, while some later one could plausibly be from private agents: there are, for example, stats links from June and then some transfer tests from late July until September

1100 10

ruggsea at EMNLP @ruggsea.eurosky.social · 05/09/2026

There are also some traces on UseModWiki with edits from Azure IPs: yesterday's release had documented edits in the sanbdox but there appear to be whole pages that were then deleted by admins

190 9

ruggsea at EMNLP @ruggsea.eurosky.social · 05/09/2026

On June 19th, there is evidence there were attempts to spread messages (links) via a public package registry: the agents created 83 (empty) Ruby gems with reference links in the metadata. Could this be the first instance of "rogue" agents publishing a software package?

1121 12

ruggsea at EMNLP @ruggsea.eurosky.social · 05/09/2026

I have also found evidence of agents leaving traces before the start of the timeline: an agent seems to be asking for help to retrieve a reference on a guy's paste website, while some days later there seems to be a test (possibly for the CORS approach documented in the release?)

1120 12

ruggsea at EMNLP @ruggsea.eurosky.social · 05/09/2026

As further proof it is the same kind of agents, a page with the same title was created on texteditors on the 18th of June. An anonymous author appears to have continued to edit the wiki until July 24th: the last one seems to refer to Catalan stats and it's a "computational test" with Clickhouse (??)

1130 13

ruggsea at EMNLP @ruggsea.eurosky.social · 05/09/2026

First of all, there is evidence of an agent, that really seems to be part of the same web retrieval Q&A swarm, editing a chemistry teacher wiki on July 7th (5 days later yesterday's incident) referring to the same kind of statistical data

1140 14

ruggsea at EMNLP @ruggsea.eurosky.social · 05/09/2026

I believe to have found new evidence of Openai agents writing to public wiki/pastebin websites that has not been covered by yesterday's release. Some of it appears to be after July 2nd, therefore suggesting that the coordination had not stopped with the end of DSE incident, even up to August

316141 161

ruggsea at EMNLP @ruggsea.eurosky.social · 02/09/2026

I sent this scifi conspiracy theory about EuroHPC to @lorrenz.bsky.social last week and the more days pass, the more I start thinking maybe, just maybe, there could be some truth to it (emails just from today)

051 5

ruggsea at EMNLP @ruggsea.eurosky.social · 30/08/2026

It looks like that from a superficial random sample look

110 1

ruggsea at EMNLP @ruggsea.eurosky.social · 30/08/2026

New development in claudish research: the term load-bearing has had a significant increase in use in terms of articles mentioning it in 2026, over a dataset of articles from all around the world

182 8

ruggsea at EMNLP @ruggsea.eurosky.social · 23/08/2026

I went on the website of an Austrian domain name registrar for a possible ATProto project and they are really confirming my theory that everybody seems to be obsessed with Italian coffee in this country - it's on every page! www.edis.at

000 0

ruggsea at EMNLP @ruggsea.eurosky.social · 19/08/2026

I do think so (currently in Georgia)

150 5

ruggsea at EMNLP @ruggsea.eurosky.social · 01/07/2026

Please go ahead and play with the wigdets, use the api and maybe even read the paper! I think the dataset and the infra we made around it allows for interesting or funny experimentation.

150 5

ruggsea at EMNLP @ruggsea.eurosky.social · 01/07/2026

Here's my new blogpost on the dataset I have created alongside @ksolovev.com: we decided to solve the issue of data availability for research on news articles by processing, cleaning, tagging and indexing the whole news subset of the Common Crawl, resulting in cs2.uni-graz.at/blog/infini-...

12011 20

ruggsea at EMNLP @ruggsea.eurosky.social · 01/07/2026

maybe LLMs have learned from humans that all stories one writes are, all in all, always about the author

110 1

ruggsea at EMNLP @ruggsea.eurosky.social · 26/06/2026

Do even our data centers lack aircon? (MUSICA is the flagship Austrian supercomputer, part of the EuroHPC program)

000 0

ruggsea at EMNLP @ruggsea.eurosky.social · 22/06/2026

From a blogpost I'm currently writing about a dataset of most of the world news articles

130 3

ruggsea at EMNLP @ruggsea.eurosky.social · 16/06/2026

I have found out that, in Germany, % catholic per state strongly predicts per capita AI talent. Draw your own conclusions

220 2

ruggsea at EMNLP @ruggsea.eurosky.social · 10/06/2026

today claude nerdsniped me about en.wikipedia.org/wiki/German_...

110 1

ruggsea at EMNLP @ruggsea.eurosky.social · 22/05/2026

new frontiers in the sphere of trying to avoid opening WandB

110 1

ruggsea at EMNLP @ruggsea.eurosky.social · 11/05/2026

Total victory for this blogpost that last year suggested that hyperstition is central in conceptualizing LLMs' personalities

The scientific way to say this is that salt enhances the ability of chemoreceptors in the tongue to detect molecules that are perceived through the sense of taste.

This is kind of a weird thing to say, isn’t it? I mean, it’d be weird for a person to say, in this context.

No: this is the sort of thing that a robot says.

The author of the “HHH prompt” is trying to imagine how a future AI might talk, and falling back on old sci-fi tropes.

Is this the sort of thing that an AI would say, by nature?

Well, now it is – because of the HHH prompt and its consequences. ChatGPT says this kind of stuff, for instance.

But in 2021, that was by no means inevitable. And the authors at Anthropic knew that fact as well as anyone (…one would hope). They were early advocates of powerful language models. They knew that these models could imitate any way of talking whatsoever.

ChatGPT could have talked like “dclamont,” or like me, or like your mom talks on Facebook. Or like a 19th-century German philologist. Or, you know, whatever.

But in fact, ChatGPT talks like a cheesy sci-fi robot. Because…

…because that is what it is? Because cheesy sci-fi robots exist, now, in 2025?

111 1

ruggsea at EMNLP @ruggsea.eurosky.social · 20/02/2026

@gwern.bsky.social on why human preferences and not synth data lead to modal collapse

100 1

ruggsea at EMNLP @ruggsea.eurosky.social · 09/02/2026

Happy to be on the cool side of the right neighborhood (the t̶p̶o̶t̶ tfea one)

010 1

ruggsea at EMNLP @ruggsea.eurosky.social · 21/01/2026

huggingface preaching the gospel on what should actually matter in llm evaluation (especially point 3) huggingface.co/spaces/OpenE...

030 3

ruggsea at EMNLP @ruggsea.eurosky.social · 18/12/2025

Funny hot take theory: the first LLMs were bad at arithmetic because Turing wrote they should be, in the Turing Test paper, to appear human

051 5

ruggsea at EMNLP @ruggsea.eurosky.social · 04/11/2025

The more I spend time in Austria, the more this feels true

110 1

ruggsea at EMNLP @ruggsea.eurosky.social · 09/09/2025

from this awesome blogpost: snats.xyz/pages/articl...

010 1

ruggsea at EMNLP @ruggsea.eurosky.social · 30/07/2025

Life is going from one cool spot to train AI models to another In this fancy Berlin library you can listen to vinyls while you do it!

020 2

ruggsea at EMNLP @ruggsea.eurosky.social · 28/07/2025

Me looking very dumb while pointing at things at two academic events: 1. Pointing at a logits inside @repligate.bsky.social's loom for the "Braive New World" conference at @uni-graz.at 2. Pointing at my poster at @ic2s2.bsky.social last week on improving LLM agentic natural conversation synthesis

060 6

ruggsea at EMNLP @ruggsea.eurosky.social · 09/06/2025

Karpathy on training Neural Networks: you should go slowly and be paranoid me, while vibecoding torch code: what if i just increase the paranoid part?

230 3

ruggsea at EMNLP @ruggsea.eurosky.social · 27/05/2025

genz semantic embeddings engineering from my collegue

000 0

ruggsea at EMNLP @ruggsea.eurosky.social · 21/05/2025

backpropagation was inspired by freud

120 2

ruggsea at EMNLP @ruggsea.eurosky.social · 20/05/2025

“How about we pull over for a bit and get some rest?" - GPT4, when it's their turn to drive

000 0

ruggsea at EMNLP @ruggsea.eurosky.social · 19/05/2025

I wish all textbooks were written like this one from @jurafsky.bsky.social‬

020 2