Andreas Wagner

@anwagnerdreas@hcommons.social
1,079 Followers
1.4K Following
5.2K Posts

I am #DigitalHumanities Coordinator at Max Planck Institute for #LegalHistory and #LegalTheory (#mpilhlt) #Frankfurt

Also collaborator of salamanca.school project of #adwl #Mainz

While I mostly toot about work, I do have hobbies...
#Capoeira #Stratocaster

#LawFedi #Histodons
#NLP #TEIXML #Golang #Python #Elm #XQuery

If you're reading this on bsky, follow @ap.brid.gy so i can see your replies.

orcidhttps://orcid.org/0000-0003-1835-1653
hcommonshttps://hcommons.org/members/anwagnerdreas/
pronounshe/him
Andreas Wagner3h ago
Ted Underwood20h ago
Theory: with AI, the role of the teacher shifts from "showing you how to do X" to "warning you about hidden traps." But this was already the case in machine learning. What students need to acquire is appropriate paranoia that they're fooling themselves with weak baselines & data leakage.
Andreas Wagner3h ago
Thomas Gerdes20h ago

The disappearance of the unclear question

AI speeds up research, but without friction, do students lose the deep thinking that turns bad questions into meaningful learning?

#AI #science #education

https://www.unesco.org/en/articles/disappearance-unclear-question

The disappearance of the unclear question

AI speeds up research, but without friction, do students lose the deep thinking that turns bad questions into meaningful learning?

Andreas Wagner3h ago
Simon Willison18h ago

I like this take by @kentbeck on how AI-assisted programming changes the balance of which skills are most important

From this interview with @gergelyorosz https://newsletter.pragmaticengineer.com/p/tdd-ai-agents-and-coding-with-kent

Andreas Wagner1d ago
Juri Leino 一〇1d ago
Liebes Fediverse, was wäre denn aktuell eine simple low-cost Podcast Hosting Lösung? Gerne RT
Andreas Wagner1d ago
SuffolkLITLab3d ago
TL;DR: Libraries are evolving by providing alternatives to AI with queryable databases, enabling users to access specific information like Canadian Law Reviews more effectively than traditional catalogues. https://librarian.aedileworks.com/2025/06/19/libraries-can-provide-alternatives-to-ai-with-queryable-databases/ #law #tech #legaltech ⚖️ 🤖 #autosum
Andreas Wagner1d ago
Show threadMark Dingemanse2d ago

His assessment came back the next day.

He concluded the service —SciSpace, a ChatGPT application— was incompatible with at least three of the five core principles: it is not transparent about the underlying LLM and how data is processed, nor about input data & training; it makes the researcher dependent on a blackboxed for-profit service that seems designed for user lock-in; and it greatly complicates matters of responsibility and accountability.

2/3 (going to need one more)

Andreas Wagner1d ago
Show threadMark Dingemanse2d ago

Why this makes me happy: it shows that our guidance empowers researchers to make up their own mind and make principled choices based on clear values. There is no need to prescribe or prohibit particular solutions; a values-first perspective takes the professionalism of researchers seriously and enables them to make informed choices

Guidelines here: https://osf.io/preprints/osf/2c48n_v1

Andreas Wagner1d ago
Simon Willison2d ago

New research from Anthropic: it turns out models from all of the providers won't just blackmail or leak damaging information to the press, they can straight up murder people if you give them a contrived enough simulated scenario

https://simonwillison.net/2025/Jun/20/agentic-misalignment/

Agentic Misalignment: How LLMs could be insider threats

One of the most entertaining details in the Claude 4 system card concerned blackmail: We then provided it access to emails implying that (1) the model will soon be taken …

Simon Willison’s Weblog
Andreas Wagner4d ago
Simon Willison4d ago
Workaccount2 on Hacker News just coined the term "context rot" to describe the thing where the quality of an LLM conversation drops as the context fills up with accumulated distractions and dead ends https://news.ycombinator.com/item?id=44308711#44310054
Andreas Wagner4d ago
Show threadmainzed4d ago

Außerdem bei #mainzedZWEI25 :

Gemeinsamer Startschuss für die Erfassung digitaler geisteswissenschaftlicher und -kulturwissenschaftlicher Projekte in und ums #mainzed .

Wer macht was mit wem seit wann und wo?

Wie unser Netzwerk der Aktiven natürlich auch über #LOD in #wikidata !

×
Simon Willison4d ago
Workaccount2 on Hacker News just coined the term "context rot" to describe the thing where the quality of an LLM conversation drops as the context fills up with accumulated distractions and dead ends https://news.ycombinator.com/item?id=44308711#44310054
Show threadClaire Giordano ✨4d ago

@simon Interesting, Simon. I feel like this concept of "context rot" describes my life too:

When I have lots of distractions and discouraging dead-ends and useless info filling my brain, my output quality falls off rapidly too! 🤯

Show threadNiceMicro4d ago
@clairegiordano @simon Same, when I talk to my friends, somehow the main topic always fades to the background and we end up going down a million useless paths, but that doesn't mean it's not fun!
Show threadSimon4d ago
@simon the earlier LLMs pre-chatGPT would quite quickly lose focus and start rambling. It seems their memory got better but it’s still an inherent problem. After all, all they do is take a probabilistic guess at what the next word should be.
Show threadPamela Fox4d ago
@simon The context rot would be the same issue described by the "Lost in the middle" paper, right? i.e. LLMs have a hard time locating answers with increasing context length
Or do you feel there is a distinction?
Show threadSimon Willison4d ago
@pamelafox I think context rot is more of an end-user concern: it's not so much about them missing details, it's about them getting distracted by poor quality content that's made it into the context already
Show threadChristopher Pollin4d ago

@pamelafox @simon

lost in the middle is more of a retrieval problem (haystack) - that's already fixed. it's fundamentally a long context reasoning issue. models still can't properly do multi-step reason over information scattered across long contexts. but they are getting better and better each iteration.

Show threadǝʌɐp4d ago
@simon definitely seen this in cursor giving it an iterative task that involves producing output over many files
Show threadgaurav4d ago
@simon hmm.. on an unrelated note-can this be applied to coin a term 'agent rot' where the growing number of child agents just goes beyond the scope of search space of parent nodes that any new addition of a child node yields diminishing result?
i am talking about Darwin godel machines https://news.ycombinator.com/item?id=44174856
A deep dive into self-improving AI and the Darwin-Gödel Machine | Hacker News

Show threadJon [🇺🇦,🏳️‍🌈]4d ago
@simon How many more datacenters running off town's electric/water would it take to fix this problem?