Titus von der Malsburg 📖👀💭
@tmalsburg.scholar.social.ap.brid.gy
69 followers 2 following 140 posts
Linguist, cognitive scientist, tt prof at University of Stuttgart. I study language and how we understand it one word at a time. #eyetracking, large-scale […] 🌉 bridged from ⁂ https://scholar.social/@tmalsburg, follow @ap.brid.gy to interact
Posts Media Videos Starter Packs
tmalsburg.scholar.social.ap.brid.gy
More observations: Finding bugs introduced by LLM is frustrating because it have no mental model of the bugs that LLM typically introduce. They can be unlike human bugs or at least unlike my own human bugs.

When making revisions, the LLM also likes to leave dead code or code that is executed […]
Original post on scholar.social
scholar.social
tmalsburg.scholar.social.ap.brid.gy
One big problem is that the LLM doesn't see the result and therefore can't critique the solution. So I had to explain the shortcomings over and over again, which didn't feel like a good use of my cognitive resources.

In sum: Was vibecoding faster in the end than coding from scratch? Perhaps. Is […]
Original post on scholar.social
scholar.social
tmalsburg.scholar.social.ap.brid.gy
I "vibecoded" a little thing for #emacs (KWIC index based on helm). The LLM was able to come up with a solution that works (impressive given the niche domain), but the solution was ugly and slow. I could get the LLM to fix some of these issues but it was getting increasingly tedious as the code […]
Original post on scholar.social
scholar.social
tmalsburg.scholar.social.ap.brid.gy
Zur Verlängerung meiner Mensakarte muss ich ein Formular auf Papier ausfüllen es dann per Post oder Fax an das Dezernat Personal schicken. Dort gibt es wahrscheinlich eine Person, die nichts anderes macht, als diese Formulare zu bearbeiten. Das Formular wird mir dann per Post zurückgeschickt und […]
Original post on scholar.social
scholar.social
tmalsburg.scholar.social.ap.brid.gy
#ediff, #emacs tool for comparing and merging versions of a text file, is pretty neat. Worth actually reading the manual and learning its features. Can be a huge time-saver.
tmalsburg.scholar.social.ap.brid.gy
There is a second person in my department who is using #emacs with #vim keybindings (evil mode). What are the odds?!
tmalsburg.scholar.social.ap.brid.gy
So Microsoft unilaterally decided to install some bigger update on our lab workstations, which left them unusable. To be able to use these computers, we were forced to create Microsoft accounts and provide personal details such as name and date of birth, which makes no sense at all on a shared […]
Original post on scholar.social
scholar.social
tmalsburg.scholar.social.ap.brid.gy
@felwert True, git is certainly not ideal for non-technical users but in some settings it’s simply a given or there’s no good alternative.
tmalsburg.scholar.social.ap.brid.gy
What’s the best software for non-technical users to get started with #git? Use case is simple project repositories for experimental studies: stimuli in .csv, experiment implementationl, analysis in R, manuscripts in #markdown.
tmalsburg.scholar.social.ap.brid.gy
Die unzuverlässige Bahn ist ein Geschenk für das Hotelgewerbe. Geschäftsreisende müssen jetzt bei wichtigen Terminen oft schon einen Tag früher anreisen. Das schlägt sich auch immer deutlicher im Reisekostenbudget meiner Arbeitsgruppe nieder.
tmalsburg.scholar.social.ap.brid.gy
@rikefranke.bsky.social "NATO is restrained by the fact that there is no certainty about the reaction and position of the U.S. under the Trump administration." So, with a different president in the White House, NATO would be unrestrained? Scary stuff.
tmalsburg.scholar.social.ap.brid.gy
Q about LLM tokenizers: If a string can be tokenized in multiple ways, which tokenization will be preferred? I guess I assumed that the algorithm is greedy in the sense that it chooses the longest possible token at each step, but that doesn't necessarily produce a globally optimal tokenization […]
Original post on scholar.social
scholar.social
tmalsburg.scholar.social.ap.brid.gy
I fully blame Silicon Valley for ruining the internet. It could have continued to be fun and useful, but no, the top priority was to monetize it, ads had to be shown, as many as possible. Google is the worst offender IMO closely followed by Meta. Shame on these companies.
tmalsburg.scholar.social.ap.brid.gy
Can't we just shut down the internet and collectively pretend that it never happened?
tmalsburg.scholar.social.ap.brid.gy
Hab gerade die zweite Nosferatu-Spinne in der Wohnung gefunden – in meiner Hose, als ich sie anziehen wollte. Die Nachbarin sagt, es gäbe jetzt viele in der Nachbarschaft. Leute, schüttelt eure Kleider vorm anziehen aus ☝️
tmalsburg.scholar.social.ap.brid.gy
I'd be much more bullish on AI if it was better at suggesting acronyms for research projects.
tmalsburg.scholar.social.ap.brid.gy
Habe gerade eine Nosferatu-Spinne in meiner Wohnung gefangen und in den Garten gebracht. Verdammt schnell die Biester, und groß. 😳
tmalsburg.scholar.social.ap.brid.gy
I don't make that kind of money, but hey, at least people think I do!
tmalsburg.scholar.social.ap.brid.gy
A local Porsche dealership contacted me with an offer (unsolicited). They're seriously overestimating how much professors earn 😂
tmalsburg.scholar.social.ap.brid.gy
Three elderly people next to me on the train to #amlap2025 are having a lively conversation in a wild mix of high German, Swiss German, Swabian, Spanish, and Italian. They switch between up to three languages in each conversational turn.
tmalsburg.scholar.social.ap.brid.gy
Inbox zero OMG!!!! Please don't send me e-mails for the next 24h 🙏🙏🙏
tmalsburg.scholar.social.ap.brid.gy
I made a poster for #amlap2025 (academic conference) using #latex #beamer. I chose Beamer because I wanted the design part to be as painless as possible and focus on content. I think the result looks pretty decent. What do you think? And what's your preferred way to make posters?
Screenshot of a poster for an academic conference. Title is "Transformers Fail to Predict Consistent Effects for Agreement Attraction Configurations". Poster talks about the research question, study design, and results. Pretty basic look, probably too text-heavy for this type of poster.