iiinova — where your digital self keeps evolving

Dominik Peters @dominik-peters.de · 07/02/2026

Not that new, but an excellent thesis!

Dominik Peters @dominik-peters.de · 12/06/2025

My first PhD student just graduated with an amazing dissertation. It was a great pleasure working together and to watch Théo pick up all the skills to be a great researcher. Excited for the future!

031

Dominik Peters @dominik-peters.de · 23/03/2025

Résultats de la votation de la Ville de Paris ("Pour ou contre végétaliser et rendre piétonnes 500 nouvelles rues dans Paris ?") par arrondissement. www.datawrapper.de/_/2HKVf/

000

Reposted by Dominik Peters

arxiv cs.GT @arxiv-cs-gt.bsky.social · 21/03/2025

Christian Kroer, Dominik Peters Computing Lindahl Equilibrium for Public Goods with and without Funding Caps arxiv.org/abs/2503.16414

021

Dominik Peters @dominik-peters.de · 17/03/2025

@dividedargument.bsky.social @williambaude.bsky.social @epps.bsky.social Was just relistening to an old episode and it turns out to include a timely side discussion about auto pen.

000

Dominik Peters @dominik-peters.de · 06/03/2025

Yeah, it would be nicer if Russia stopped its aggression and Westeners would be interested in going there again.

Dominik Peters @dominik-peters.de · 26/02/2025

Similar for some commuter trains around Paris. > The vehicle's width is 3.06 metres (10 ft 0 in), which is wider than previous trains [which were are 2.82 metres], achieved by having a relatively short car length of 13.24 metres (43 ft 5 in). en.wikipedia.org/wiki/SNCF_Cl...

en.wikipedia.org

SNCF Class Z 50000 - Wikipedia

000

Dominik Peters @dominik-peters.de · 25/02/2025

“HB 1297 passed 75 to 16. A similar bill passed in the 2023 session – but was vetoed by then-Governor Doug Burgum – and the Legislature failed to override that veto.” Sounds like they now have enough votes to override a veto 🥲

Dieses Ergebnis war für mich überraschend -- ich hatte nicht mit so viel "Kollateralschaden" außerhalb Bayerns gerechnet. Wahrscheinlich sollte das Wahlrecht noch weiter angepasst werden, um das Prinzip der personalisierten Verhältniswahl besser zu bewahren.

000

Die anderen 20 Wahlkreise haben ihr Direktmandat nur verloren, weil die Zweitstimmen erst nach Land aufgeteilt werden, was die Zweitstimmendeckung sehr unflexibel macht. Wenn man auf Bundesebene zählen würde, wären tatsächlich nur die 3 bayrischen Wahlkreise verloren gegangen.

Der Grund, warum der Bundestag in letzter Zeit angewachsen ist, ist der fallende Zweitstimmenanteil der CSU in Bayern (wobei die CSU bei den Erststimmen weitenteils stärkste Kraft geblieben ist). Aber nur 3 der 23 Wahlkreise ohne Direktmandat befinden sich in Bayern!

100

Die Wahlkreise ohne Direktmandat sind genau diejenigen, die am meisten umkämpft sind, häufig mit einem Kopf-an-Kopf-Rennen, wie hier in Frankfurt, wo der CDU-Kandidat knapp gewonnen hat, aber dann doch kein Direktmandat zugeteilt bekommen hat. Das wirkt problematisch.

100

Um Frankfurt, Darmstadt, Heidelberg, und Mannheim gibt es eine Ansammlung von Wahlkreisen ohne Direktmandat -- in diesen Städten haben Bürger also keine Bundestagsmitglieder, die speziell für sie zuständig sind. Link zur Karte: www.datawrapper.de/_/bsCzu/

100

Der Grund dafür ist das neue Konzept der "Zweitstimmendeckung", nachdem Direktkandidaten mit den meisten Stimmen in ihrem Wahlkreis nur dann in den Bundestag einziehen, wenn ihre Partei genügend Sitze auf Grundlage ihres Zweitstimmenanteils erhalten hat.

100

Das deutsche Wahlrecht gibt jedem Wähler zwei Stimmen: für einen Direktkandidaten und eine Liste. Aber das Wahlrecht wurde für dieses Jahr geändert, um den Anwuchs der Sitze im Bundestag zu stoppen. Das führt aber dazu, dass 23 Wahlkreise diesmal kein Direktmandat erhalten. 🧵

100

I think this result will be surprising to people. I hadn't expected their to be so much "collateral damage" outside Bavaria. Maybe there is potential for further tweaking of the German election system.

000

The reason why the other 20 districts lost their representative is because the votes for a party are divided between individual states, leading to inflexible quotas. But one could easily operate the process nationwide, in which case only the 3 Bavarian districts would have lost their representative.

100

The underlying reason why the number of Bundestag members was rising in recent years was the falling popularity of the CSU party in Bavaria (in terms of % but not in terms of districts won). But only 3 of the 23 districts without a representative are in Bavaria!

100

The unrepresented districts are those where the local representative got the lowest % of votes, which tend to be the most competitive districts, such as this one in Frankfurt where the CDU candidate narrowly won -- but then actually lost and got no seat. This seems problematic.

100

There is a sort of cluster where this happened around Frankfurt, Darmstadt, Mannheim, and Heidelberg -- major cities where people will not have a local representative in the Bundestag that they can call. Link to map: www.datawrapper.de/_/H6dRv/

100

Dominik Peters @dominik-peters.de · 23/02/2025

The German election system lets people vote for both a local representative and a party. But it changed this year: to stop the number of Bundestag members from fluctuating, some candidates that would've become local representatives lost if their party got too few votes. This happened in 23 districts

100

Feels so good to be able to vent to someone wise and well-read and who tends to agree with you. I have been procrastinating returning this page proof for months now because I get pissed off everytime I open the PDF.

000

Dominik Peters @dominik-peters.de · 19/02/2025

I can imagine that escalators that aren’t being used (and thus don’t move except during tests) don’t need much maintenance. (I feel like I’ve seen emergency shaft escalators a few times, but maybe I just saw the Munich plans on multiple occasions.)

They put palm trees all over Place de la Concorde — Paris Plage in February!

Some caveats to the data: additional conversations have happened outside these websites (e.g. on Poe just after GPT-4 was released and with GitHub Copilot because it is free for eduction). Conversations not growing doesn't mean that number of tokens isn't growing.

Made by exporting my data from ChatGPT and Claude, then analyzing the conversations.json in the resulting zip files using a python script (claude.site/artifacts/55..., I renamed the jsons to openai.json and claude.json) and having Claude make a bar chart with the resulting CSV file.

Dominik Peters @dominik-peters.de · 12/02/2025

How many conversations I have with different LLMs, per month. Peaks in months where I've paid for the plus versions (e.g. 2024-04 for Claude 3 Opus, 2024-09 for o1-preview, 2025-01 for o1). Surprising that there is no clear growth since Summer 2023.

120

Probably a chart showing a distribution for each trade across months/days would be more informative than a potentially discontinuous “most common” metric

000

Dominik Peters @dominik-peters.de · 11/02/2025

Unfortunately mathematicians typically 1-index things. Say you have a sequence x_1, x_2, … then defining natural numbers as excluding 0 lets you say “for all i in N, x_i is non negative” etc

000

Dominik Peters @dominik-peters.de · 04/02/2025

That vegan meal sounds pretty good! Better than the previous veggie burger (pictured) for sure

030

Dominik Peters @dominik-peters.de · 03/02/2025

It will definitely be possible from January 2026: the tickets for UK visitors are already on sale. Quite a bit pricier than what it cost in prior years without Epic Universe! ukstore.universalorlando.com/park-tickets

Dominik Peters @dominik-peters.de · 01/02/2025

These are two differences which can be analyzed separately: separation of powers between executive and legislature, and proportional representation. A priori the former adds a way to avoid abuses of power not available in parliamentary systems: the executive can choose not to enforce unjust laws.

120

Dominik Peters @dominik-peters.de · 25/01/2025

Not a one-off either, last week I posted an issue at 23:41 and they marked it as fixed at 1:55.

000

Dominik Peters @dominik-peters.de · 25/01/2025

16:31: I notice a pedestrian signal facing the wrong way, open an issue in the Paris DansMaRue app. 18:36: The city emails me a photo of having fixed it. @paris.fr 👏🔥

100

Dominik Peters @dominik-peters.de · 23/01/2025

New application for LLMs: if you generate a list of counterexamples, you can ask it which one is the nicest looking.

Dominik Peters @dominik-peters.de · 21/01/2025

There are too many computer scientists who don't know essential optimization concepts like Farkas lemma and KKT conditions. These seem to not be part of the standard CS undergrad curriculum, but should be.

Dominik Peters @dominik-peters.de · 17/01/2025

Why riskier? Possible tree damage during transport and planting?

Dominik Peters @dominik-peters.de · 10/01/2025

Academics are slowly starting to get back to work, judging from how often they check the documentation of TikZ (a figure drawing package for LaTeX)

Chart showing how often people click on links to tikz.dev from Google Searches, showing a long big drop over christmas and new years

011

Dominik Peters @dominik-peters.de · 09/01/2025

Congrats, great news! To be fair, ILPs are solved by rather advanced algorithms.

Dominik Peters @dominik-peters.de · 06/01/2025

If you want your ideas to become more influential, it seems worthwhile to post translations of articles and papers in many languages. Then they end up in the training corpus of LLMs multiple times, and the copies are expensive to filter out. Plus you get broader accessibility to humans.

120

Dominik Peters @dominik-peters.de · 22/12/2024

I just published my own digitized version of the FSI French Basic course from the 1960s and 70s. Super boring but quite effective, and now much easier to use thanks to audio files synced to the textbook. dominik-peters.de/fsi/

dominik-peters.de

FSI French Basic Course

Enhanced digital version of FSI French Basic course, with audio, text, and exercises. Available free on the web or as a podcast.

020

Dominik Peters @dominik-peters.de · 17/12/2024

Glad you're back! It was indeed your longest ever hiatus:

Bar chart showing the gap in days between successive episodes. The latest episode had a gap of 82 days. The next highest gap was 60 days, attained in late 2021.