Lightnews — Scholar-powered news

Katia Schwerzmann @katschwerzmann.bsky.social · Aug 29

Love your music rec!

Reposted by Katia Schwerzmann

Kulturwissenschaftliches Institut Essen (KWI) @kwi-essen.bsky.social · May 26

#KWIBlog: In today's blog text, former Thyssen Fellow @katschwerzmann.bsky.social investigates the role of reading in the context of new developments in AI, stressing the need for ongoing investment in close and critical reading that considers AI practices and limitations.
🔎 tinyurl.com/25upvwyj

3 9

Katia Schwerzmann @katschwerzmann.bsky.social · May 22

Ich hätte Interesse. Ist das Ticket noch zu haben?

1

Reposted by Katia Schwerzmann

Ben Williamson @benpatrickwill.bsky.social · Feb 9

A social sciences and humanities reading list on AI in education 🧵

53 170 470

Katia Schwerzmann @katschwerzmann.bsky.social · May 19

Thank you for building this reading list on AI and eduction!

1 3

Reposted by Katia Schwerzmann

Roland Meyer @bildoperationen.bsky.social · May 19

Very timely and necessary critical intervention by @katschwerzmann.bsky.social ‬– highly recommended: «By embracing LLMs before developing any critical framework for their use in pedagogical and research contexts, the University allows itself to be governed by the contingent, ever-evolving ...
1/

Ruled by the Representation Space: On the University’s Embrace of Large Language Models

arxiv.org

2 20 53

Katia Schwerzmann @katschwerzmann.bsky.social · May 19

Thank you for sharing your reading @bildoperationen.bsky.social. I am glad it resonated. Working at a critique of generative AI in the context of research and education is currently a somewhat lonely endeavor. I hope more researchers will join. We need to tackle this from a plurality of approaches.

1 1 7

Reposted by Katia Schwerzmann

Alex Campolo @alexcampolo.bsky.social · Apr 30

Totally vague PR blog post as we now expect from OpenAI - remember when they used to publish research? But I am nonetheless interested in the conceptualization of ML "pathologies" - will have to think further about "sycophancy" as political problem... openai.com/index/sycoph...

Sycophancy in GPT-4o: What happened and what we’re doing about it

We have rolled back last week’s GPT‑4o update in ChatGPT so people are now using an earlier version with more balanced behavior. The update we removed was overly flattering or agreeable—often describe...

openai.com

1 1 5

Katia Schwerzmann @katschwerzmann.bsky.social · Apr 13

Genius: www.paloaltoonline.com/news/2025/04...

Silicon Valley crosswalk buttons apparently hacked to imitate Musk, Zuckerberg voices

Crosswalk buttons along the mid-Peninsula appear to have been hacked, so that when pressed, voices professing to be Mark Zuckerberg or Elon Musk begin speaking.

www.paloaltoonline.com

2

Reposted by Katia Schwerzmann

Louise Amoore @amoorelouise.bsky.social · Mar 21

Fabulous paper from @alexcampolo.bsky.social and @katschwerzmann.bsky.social mapping an anatomy of the concept of #alignment in #machinelearning #algorithmicsocieties link.springer.com/article/10.1...

“Desired behaviors”: alignment and the emergence of a machine learning ethics - AI & SOCIETY

The concept of alignment has undergone a remarkable rise in recent years to take center stage in the ethics of artificial intelligence. There are now numerous philosophical studies of the values that ...

link.springer.com

6 15

Katia Schwerzmann @katschwerzmann.bsky.social · Mar 21

So happy to see this paper cowritten with @alexcampolo.bsky.social finally published!!

Alex Campolo @alexcampolo.bsky.social · Mar 21

New paper with @katschwerzmann.bsky.social in AI & Society. Our basic question is: "Where has the ethical imperative of alignment come from? how, exactly, does it propose to regulate machines’ and peoples’ conduct?" It's quite critical - interested to hear thoughts link.springer.com/article/10.1...

“Desired behaviors”: alignment and the emergence of a machine learning ethics - AI & SOCIETY

The concept of alignment has undergone a remarkable rise in recent years to take center stage in the ethics of artificial intelligence. There are now numerous philosophical studies of the values that ...

link.springer.com

2 9

Reposted by Katia Schwerzmann

Ars Technica @arstechnica.com · Feb 24

Big Tech data center buildouts have led to $5.4 billion in public health costs

Cancers and asthma among illnesses linked to air pollution from powering data centers.

arstechnica.com

3 23 72

Reposted by Katia Schwerzmann

Alex Campolo @alexcampolo.bsky.social · Feb 24

Really looking forward to the "New Reading Scenes" workshop later this week, organized by @katschwerzmann.bsky.social at @kwi-essen.bsky.social & @sfb1567.bsky.social. Remote registration is available: docs.google.com/forms/d/e/1F...

1 2 10

Katia Schwerzmann @katschwerzmann.bsky.social · Feb 24

Exciting Alex!!!!

1

Reposted by Katia Schwerzmann

Kulturwissenschaftliches Institut Essen (KWI) @kwi-essen.bsky.social · Feb 12

The workshop is organized by our Thyssen@KWI Fellow
@katschwerzmann.bsky.social
with talks by @floriansprenger.bsky.social | @alexcampolo.bsky.social |
@rainermuehlhoff.bsky.social | @moritzhiller.bsky.social
Further Information here: www.kulturwissenschaften.de/veranstaltun...

Kulturwissenschaftliches Institut Essen (KWI)

Das Kulturwissenschaftliche Institut Essen (KWI) ist ein interdisziplinäres Forschungskolleg für Geistes- und Kulturwissenschaften in der Tradition internationaler Institutes for Advanced Study. Als i...

www.kulturwissenschaften.de

2 3

Katia Schwerzmann @katschwerzmann.bsky.social · Feb 11

Wichtige Stellungnahme bezüglich der aktuellen politischen Angriffe gegen die Freiheit von Forschung und Lehre:

www.fg-gender.de/angriffen-au...

Angriffen auf die Freiheit von Forschung und Lehre entschlossen entgegentreten! | Fachgesellschaft Geschlechterstudien // Gender Studies Association

www.fg-gender.de

3 3

Reposted by Katia Schwerzmann

Megan McIntyre @rcmeg.bsky.social · Feb 5

Not to put too fine a point on it, but a week ago, the CSU closed 6 departments (among them, philosophy, women's and gender studies, and economics) and fired 46 tenure-line faculty at Sonoma State. One of those faculty studies AI ethics.

They fired the AI ethics professor. Then, this.

Hypervisible @hypervisible.blacksky.app · Feb 4

“The rollout will cover 23 campuses of the largest public university system in the United States, enabling students to access personalized tutoring and study guides through the chatbot, while the faculty will be able to use it for administrative tasks.”

OpenAI targets higher education in the U.S. with ChatGPT rollout at California State University

Microsoft -backed OpenAI said on Tuesday it will roll out an education-specific version of its chatbot to about 500,000 students and faculty at California State University as it looks to expand its user base in the academic sector and counter competition from rivals like Alphabet .

www.reuters.com

11 120 230

Katia Schwerzmann @katschwerzmann.bsky.social · Feb 5

This language signals ML community's tendency—or rather desire—to make the human factor, in particular researchers' judgement, evaluation, and labor, disappear from view and from model training. The rule-based component of the reward model is interesting, though.

Katia Schwerzmann @katschwerzmann.bsky.social · Feb 5

One also notices once again the type of naturalizing language that @alexcampolo.bsky.social and I critically analyze in our work. For instance: "During training, DeepSeek-R1-Zero **naturally** emerged with numerous powerful and interesting reasoning behaviors" (p. 3).

1

Katia Schwerzmann @katschwerzmann.bsky.social · Feb 5

So what exactly is new? That LLMs can do well in math reasoning and coding without relying on supervised learning but still don't do well enough in natural language tasks to not rely on supervised fine-tuning in the end?

1 1

Katia Schwerzmann @katschwerzmann.bsky.social · Feb 5

The pure RL phase concerns math and coding problems only. The reward model assesses the base model's solution to "deterministic" math problems through "rule-based verification of correctness," while for coding problem "a compiler can be used to generate feedback based on predefined test cases"(p.6).

1

Katia Schwerzmann @katschwerzmann.bsky.social · Feb 5

"We create new SFT data through rejection sampling on the RL checkpoint, combined with supervised data from DeepSeek-V3 in domains such as writing, factual QA, and self-cognition, and then retrain the DeepSeek-V3-Base model" (p.2). This seems to be a quite classical fine-tuning process.

1 1

Katia Schwerzmann @katschwerzmann.bsky.social · Feb 5

To tackle this issue, the "DeepSeek-V3-Base model" is fine-tuned using the kind of data commonly used in supervised fine-tuning (SFT):

1

Katia Schwerzmann @katschwerzmann.bsky.social · Feb 5

"Pure RL" means here RL that doesn't rely on supervised learning with annotated data. But then, one reads on p. 2 that "DeepSeek-R1-Zero encounters challenges such as poor readability, and language mixing," which is indeed quite a problem for a large language model.

1 1

Katia Schwerzmann @katschwerzmann.bsky.social · Feb 5

Reading DeepSeek's paper: "DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning," I find the paper somewhat misleading. DeepSeek's claim is that its model obtains excellent reasoning abilities by applying "pure reinforcement learning" onto the pre-trained model.

1 1