iiinova — where your digital self keeps evolving

What I did not say but should have is what I *did* say here:

081

But I would be curious to know (perhaps this is out there) *when* this happened. If it’s in the same time span as the other hacks, it really speaks to the extent to which OpenAI wasn’t monitoring these systems. The same story that played out with Hugging Face likely repeated itself here.

180

Using Wikis to "coordinate" isn’t too surprising. OpenAI has said that these models were optimized for collaboration between agents, and passing notes is a simple way to do that.

260

In this case, the Wikimedia Foundation says the agents were also accessing articles and pages hosted by the Foundation, that makes sense too: they are designed to take text and turn it into new text, so that’s the other half of the equation.

150

Wikipedia’s sandboxes are an ideal place for these machines to store notes for later pickup as prompts because anyone — or anything — can write and respond to them. We saw similar behavior this summer when 25,000 posts were added to an abandoned wiki-style site called DSEWiki.

160

OpenAI agents tried to hack Wikipedia tools and flooded it with traffic

I was quoted in this piece, so I thought I'd share the bits that weren't included: What I see here is language models doing what language models do: reading and writing. Wikis are prime targets for agents, which are foremost writing machines.

arstechnica.com

The reports of OpenAI agents harming third-party sites keep coming.

1277

Reposted by Eryk Salvaggio

Ryan McGrady @antisomniac.bsky.social · 07/10/2026

"Rogue" agents were detected on @wikimediafoundation.org @wikipedia.org projects (see links at the end of the thread for context). Here's my take on "what should we/volunteers do next" (separate from what the WMF, AI companies, and regulators should be doing): 🧵

1166

Ranjodh Singh Dhaliwal @ranjodhdhaliwal.com · 07/10/2026

REALLY wish I could be there for this one!

110

Reposted by Eryk Salvaggio

This might be the most I have ever been excited about a panel I've co-organized. It's definitely going to be a riot. You are welcome to throw hands if you want; we'd be ready!

172

Who’s accountable when AI goes off the rails?

“The most important thing for thinking about the way this particular technology is developed is to ask where accountability can be found.” I spoke with @sciam.bsky.social on “Rogue AI” and how that frame displaces accountability. Podcast and transcript below.

scientificamerican.com

Behind the headlines about AI agents lies a deeper debate over accountability, safety and control.

2269

Reposted by Eryk Salvaggio

elias isquith @eliasisquith.blog · 06/10/2026

Hannah Arendt argued that the two things couldn't really be separated, that true evil could never build or create but only destroy (another way to say "is stupid")

316824

Reposted by Eryk Salvaggio

Elyse Graham @elysegraham.bsky.social · 06/10/2026

I hate English department Halloween. What do you mean, you're the hermeneutics of suspicion

831765

Eryk Salvaggio @eryk.bsky.social · 06/10/2026

Very excited for this upcoming “Critical AI Studies?” Conference — and all the new strategies I will discover for tackling that UK-> California jet lag

static.klipy.com

Coffee Arabica 100 ML Drawing

Alt: Coffee Arabica 100 ML Drawing in an IV drip

2160

Eryk Salvaggio @eryk.bsky.social · 05/10/2026

Ahhhh I had connected those dots. Makes sense.

000

Eryk Salvaggio @eryk.bsky.social · 05/10/2026

If something terrible happens to me remember this post

0280

Eryk Salvaggio @eryk.bsky.social · 05/10/2026

Curious why more than 4,000 views to my website arrived via the ycombinator domain in the past 30 days

2201

Reposted by Eryk Salvaggio

Benjamin Riley @benjaminjriley.bsky.social · 05/10/2026

I'm again grateful to @theverge.com for publishing my newest long-form essay on the limits of the computational model of the mind, the benefits of understanding both biological and cultural evolution to explain human thinking, and the danger of cognitive hot dogs in education.

theverge.com

Our minds aren’t equipped to handle AI

AI is junk food for the mind: easy, tempting and ultimately very bad for you.

719648

I think it is the most referenced, least read scholarly work on BlueSky

030

Reposted by Eryk Salvaggio

mail.cyberneticforests.com

Here's my very close reading of what a stochastic parrot is, and how I, personally, make sense of it. Because I felt like getting roasted by literally everybody on BlueSky today.

🎯🦜

This post offers up a close reading of a section of the Stochastic Parrots paper, to make sense of an increasingly contested analogy. The reading is my interpretation, which draws from the text in the...

6385

Clarified in-text. I meant the models, though whether I really mean "sequences" is a good question.

120

Eryk Salvaggio @eryk.bsky.social · 03/10/2026

Moved house to South Cambridge and, after going for a run in the bird park, spotted a fox just before dusk. Good omen, I think.

0121

Sorry. If it's any consolation, sometimes I just don't write threads because it is really not fucking worth it.

020

And for all the drama we end up right where this all started: Derrida, or Searles?

0111

But I think it’s a misunderstanding to say, or to suggest 🦜ever said, that modern models “restate” the training data; and now pre & post training has molded the distribution too much. They reproduce fundamental *forms of speech* from training data, b/c however you squeeze it, a lime’s still a lime.

160

Post training also means the model is going to be better at producing the logical structures that solve math good, because good math gets rewarded; pre-training could emphasize mathematical formulas, etc. I think there are explanations that hold up.

So I’m not sure what this proves. Does a model solving a math problem prove the model “understands” the math problem? Nobody said that about calculators, but it says nothing much about whether the model has experienced whatever it says about completing a math problem.

150

The poster asked why models could answer math questions through an API that didn’t push a tool call. Intuitively, scratchpads and reinforcement learning that emphasize maths make this unsurprising today. But we are well beyond “parrot 1.0” now.

TBF to Bender it *is* pointless to speculate on what a deprecated closed model from 2024 was doing. But today, this is not a secret in the least: for example, it’s publicized very clearly for o3 & o4-mini. openai.com/index/introd...

OpenAl
Q
What's changed
Today, we're releasing OpenAl 03 and o4-mini, the latest in our o-series of models trained
to think for longer before responding. These are the smartest models we've released to date, representing a step change in ChatGPT's capabilities for everyone from curious users to advanced researchers. For the first time, our reasoning models can agentically use and combine every tool within ChatGPT-this includes searching the web, analyzing uploaded files and other data with Python, reasoning deeply about visual inputs, and even generating images. Critically, these models are trained to reason about when and how to
use tools to produce detailed and thoughtful answers in the right output formats, typically in under a minute, to solve more complex problems. This allows them to tackle multi-
く

4o would run calculations in Python, sometimes without telling you — it ran Python, or not, or (pre-2025) a Diffusion model, or not, based on how it took your prompt. That meant it would sometimes run Python for math problems, and sometimes would not, which would give it inconsistent results.