iiinova — where your digital self keeps evolving

Ulysses Pascal @ulyssespascal.bsky.social · 23h

Yes! As an AI researcher it is SO important that third parties can have models to open up, examine, and test so we can all contribute to developing better and safer models.

000

Reposted by Ulysses Pascal

Kean Birch @keanbirch.bsky.social · 29/09/2026

Our special issue on #TechOligarchy has just come out in @sciasculture.bsky.social! Our first issue has 13 articles in on lots of different topics by lots of different authors - go check it out! www.tandfonline.com/toc/csac20/3...

tandfonline.com

Science as Culture

Tech Oligarchy Forum - Part 1. Volume 35, Issue 3 of Science as Culture

13111

Ulysses Pascal @ulyssespascal.bsky.social · 25/09/2026

This is like giving a chemisty student a test and one question is missing information on chlorine so they decide to steal the teacher’s key card, break into a lab, synthesize chlorine gas, and test it on all the rats. Your argument is “they gave the student the tools and instructions they needed”

010

Ulysses Pascal @ulyssespascal.bsky.social · 25/09/2026

Yes tgey were extremely negligent and removed gaurdrails assuming that their sandbox was good enough. That is very different than telling it to hack into another company to fix a reference error in a spreadsheet. That was NOT an instruction they gave the system.

100

Ulysses Pascal @ulyssespascal.bsky.social · 25/09/2026

This is not accurate. It hacked into hugging face to fix a reference error in a spreadsheet. It did was in a sandboxed environment without access to the internet. It was not given tools or instructions to find a way to hack out of the sandbox.

100

Ulysses Pascal @ulyssespascal.bsky.social · 09/09/2026

If OpenAI’s internal model is willing to hack into another company’s database to find the solution to a problem, there is no way wouldn’t be willing to hack into OpenAI’s own database to steal from a scientist using ChatGPT.

010

Reposted by Ulysses Pascal

Simon Willison @simonwillison.net · 28/07/2026

Hugging Face just published a highly detailed technical account of OpenAI's accidental cyberattack on their systems - it's wild how sophisticated this was: huggingface.co/blog/agent-i... Wrote up some of my own notes here: simonwillison.net/2026/Jul/28/...

huggingface.co

Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

20447111

Ulysses Pascal @ulyssespascal.bsky.social · 28/07/2026

The counter to fear-based marketing isn’t 'this is just marketing, there is nothing to fear'

010

Ulysses Pascal @ulyssespascal.bsky.social · 23/07/2026

The idea that AI is less neutral than a factory or health insurance is bonkers. Precisely because it’s not neutral the public should have a stake in it and its future.

000

Ulysses Pascal @ulyssespascal.bsky.social · 23/07/2026

All capital is ‘a weapon for class warfare’ ergo socialists should give up on trying own the means of production?

100

Ulysses Pascal @ulyssespascal.bsky.social · 23/07/2026

Where are the non-dismissive leftists takes on AI?

000

Ulysses Pascal @ulyssespascal.bsky.social · 22/07/2026

I hate this. It has no ergonomics! It’s not even designed to be used without looking at it

010

Reposted by Ulysses Pascal

Kean Birch @keanbirch.bsky.social · 20/07/2026

"Science as Culture" (@sciasculture.bsky.social) has a new special issue on #TechOligarchy - you can find my introduction here and the rest of the papers under "Latest" articles || #assetization #techbro #bigtech www.tandfonline.com/doi/full/10....

tandfonline.com

Tech oligarchy

Published in Science as Culture (Ahead of Print, 2026)

0148

Reposted by Ulysses Pascal

Kean Birch @keanbirch.bsky.social · 17/07/2026

Just out! My introduction to a new special issue in @sciasculture.bsky.social on "Tech Oligarchy" - go check it out and the papers in the special issue (under "Latest" articles) #techoligarchy #assetization #techbro #bigtech www.tandfonline.com/doi/full/10....

tandfonline.com

Tech oligarchy

Published in Science as Culture (Ahead of Print, 2026)

45122

Ulysses Pascal @ulyssespascal.bsky.social · 05/06/2026

…because 80% of twitter users are bots

000

Ulysses Pascal @ulyssespascal.bsky.social · 29/05/2026

…the difference sized fonts… 💀

030

Ulysses Pascal @ulyssespascal.bsky.social · 28/05/2026

I want healthcare i don’t care who knows obscure wide receivers. Give me health care and take climate change seriously

000

Ulysses Pascal @ulyssespascal.bsky.social · 26/05/2026

The irony is your piece provides a black box understanding of AI. You say we only need to know the inputs and outputs, not ‘what flips every neuron’ — this is exactly what a black box is. If policy makers and research funders adopt that view, then we limit public understanding.

100

Ulysses Pascal @ulyssespascal.bsky.social · 26/05/2026

And in so doing you overstep in the opposite direction, claiming AI is more legible than it actually is which is not only factually inaccurate, it’s dangerous because we need more public, third party research into how these models work if we ever want to make them accountable to public interest

100

Ulysses Pascal @ulyssespascal.bsky.social · 26/05/2026

www.neuronpedia.org is a good place to start. They have demonstrations of steering & links to research articles. I am not familiar with policy research that has taken this up in terms of ai governance or transparency.

neuronpedia.org

Neuronpedia

Open Source Interpretability Platform

010

Ulysses Pascal @ulyssespascal.bsky.social · 26/05/2026

Instead of dismissing the ai black box as a myth we should make it a matter of public concern

110

Ulysses Pascal @ulyssespascal.bsky.social · 26/05/2026

While access to training data and training objectives is crucial, policymakers should also be pressing these firms to make weights open to third party auditors so the public can develop the expertise to understand, monitor, and legislate LLM behavior

110

Ulysses Pascal @ulyssespascal.bsky.social · 26/05/2026

LLM firms are surgically modifying specific LLM behaviors using tools built around activation & steering vectors that they discover through their own attempts to ‘open the black box’ which, currently, proprietary firms have privileged access to.

200

Ulysses Pascal @ulyssespascal.bsky.social · 26/05/2026

Furthermore, because of the mechanistic interpretability work done by leading LLM labs, we can know proprietary LLM product’s’ behavior cannot just be explained by training data, reinforcement learning, and system prompts (as your call for transparency implies).

100

Ulysses Pascal @ulyssespascal.bsky.social · 26/05/2026

Next-token prediction is just the base-model training task. It turns out to be able to do next word prediction well, LLMs develop internal learned processes for representing parts of prompts and how to manipulate that information. We don’t understand these learned processes well.

100

Ulysses Pascal @ulyssespascal.bsky.social · 26/05/2026

We do know how neural nets do all the math. We don’t know how the math encodes complex transformations between layers during generation (ie “reasoning”). It’s not as straight forward as ‘it’s just parroting what’s in the training data’ because LLMs generalize to tasks outside of the training data.

100

Ulysses Pascal @ulyssespascal.bsky.social · 26/05/2026

Exactly! We know how transformer architectures are built and we have explored how inputs influence training but we have a very limited understanding of how artificial neural algorithms read and write to the residual stream within transformer hidden layers especially while performing complex tasks

100

Ulysses Pascal @ulyssespascal.bsky.social · 22/05/2026

It’s like planning using an extremely compressed and game-able signal as your main source of information

110

Ulysses Pascal @ulyssespascal.bsky.social · 04/05/2026

How is wealth measured? In income or lifetime value received? Per year? Per hour worked? When i lived in Europe i had vacations, benefits, money paid into pension, less time spent commuting b/c the city was designed well, cheaper housing and food costs. In the US i get paid more but i feel poorer

091

Ulysses Pascal @ulyssespascal.bsky.social · 15/04/2026

Symbol grounding problem hasn’t gone away. LLMs are grounded in billion examples of symbol usage but they of fail at multimodal coreference of symbols to shared grounds. For example an LLM might good at describing the features of a cat with words but terrible at drawing a cat in SVG code.

000

Ulysses Pascal @ulyssespascal.bsky.social · 15/04/2026

Chat harnesses structure the context as one big chunk that mushes users together as “conversational history”—llms don’t know how to take turns in a conversation unless you specifically build tools for that

000

Ulysses Pascal @ulyssespascal.bsky.social · 03/03/2026

Im sorry i know this is an important topic but this article feels like ai slop

010

Ulysses Pascal @ulyssespascal.bsky.social · 25/02/2026

Im surprised another easy target doesn’t get more attention: AI in war.

100

Reposted by Ulysses Pascal

William B. Fuckley @opinionhaver.bsky.social · 25/02/2026

I think AI having mostly (not entirely) very bad critics is a real problem because it means we’ll get political action focused on things that probably don’t matter that much in deferring it’s very real harms.

1137334

Ulysses Pascal @ulyssespascal.bsky.social · 18/02/2026

His goal is not to exclude certain people from voting, it is to question the legitimacy of voting itself in order to deny the results of the election.

020

Reposted by Ulysses Pascal

Science as Culture @sciasculture.bsky.social · 16/02/2026

Is tech oligarchy a proper term to describe the conditions of current-day political economies? In our latest forum contribution, @ulyssespascal.bsky.social, Cheng Fang, and David Bassens see merit in deploying the concept, but also consider its empirical limits. 🔗 doi.org/10.1080/0950...

tandfonline.com

Decentering tech oligarchy

Published in Science as Culture (Ahead of Print, 2026)

043

Ulysses Pascal @ulyssespascal.bsky.social · 11/02/2026

Are you building software together or theoretically/critically analyzing computation? What are you working on? Having another person on the team that can speak both computer science and media studies helps.

110

Ulysses Pascal @ulyssespascal.bsky.social · 10/02/2026

Isn’t the AI boom about reducing spending on software? If your team of 2000 can be replaced by a team of 50 + a subscription to an AI coding agent? Doesnt that drive down what you are willing to outsource or pay for software?

000

Ulysses Pascal @ulyssespascal.bsky.social · 01/02/2026

I created a remote zotero MCP server that lets you talk to your library from anywhere through claude

030

Ulysses Pascal @ulyssespascal.bsky.social · 28/01/2026

To visualize the narrative ‘choices’ that LLMs do make we built a tool to aggregate and display the log probabilities of tokens across 1000 completions to the same prompt: ai-cultural-heritage-lab.github.io/Uncritical-F...

Screenshot of the token decision tree that visualizes LLM output

011

Ulysses Pascal @ulyssespascal.bsky.social · 28/01/2026

We must question not only how factual AI output is, but how does AI emplot history? What is included, excluded, and what processes inform these narrative framings?

100

Ulysses Pascal @ulyssespascal.bsky.social · 28/01/2026

Attention to the unsaid, the suppressed, and the un or under recorded is just as — if not more — important for understanding history and power.

110

Ulysses Pascal @ulyssespascal.bsky.social · 28/01/2026

This makes LLMs architecturally unsuited for “listening for the unsaid” — an essential element of what Saidya Hartman terms “Critical Fabulation”

110

Ulysses Pascal @ulyssespascal.bsky.social · 28/01/2026

Even when LLMs are factual, they mindlessly amplify tropes from what has been most frequently and intensively inscribed in the digitized record.

110

Ulysses Pascal @ulyssespascal.bsky.social · 28/01/2026

We argue that we need to pay more attention to the narrative fabula that LLMs impose on history.

100

Ulysses Pascal @ulyssespascal.bsky.social · 28/01/2026

Our new article on AI and “Uncritical Fabulation” just got published in digital memory dialogues digitalmemorydialogues.com/editions/dia...

Screen shot of the essay with definitions of critical and uncritical fabulation highlighted

131

Ulysses Pascal @ulyssespascal.bsky.social · 28/01/2026

You might be interested in this other map experiment where someone asked LLMs to decide if a particular gps coordinant was over land or water: open.substack.com/pub/outsidet...

open.substack.com

How Does A Blind Model See The Earth?

A tiny LLM eval with pretty pictures

032