Martin Potthast
LightNews LightNews
banner
martin-potthast.com
Martin Potthast
@martin-potthast.com
2.5K followers 1.3K following 28 posts
Professor at the University of Kassel, https://hessian.AI, and https://ScaDS.AI. Member of @webis.de
Research in information retrieval #IR, natural language processing #NLP, and artificial intelligence.
Posts Replies Media Videos
martin-potthast.com
The fourth edition of ReNeuIR @ #SIGIR2025 is back!! Check reneuir.org to see what we have in mind this year! Paper submission deadline: May 20, 2025.
Teaser Image: The text says ReNeuIR’25 - Workshop on Reaching Efficiency in Neural Information Retrieval
April 30, 2025 at 12:16 PM
martin-potthast.com
Apparently, it can be made to second-guess itself quite easily.

4/4
Continued conversation: Me: Please take another close look at the yellow pieces. GPT: You're right, a closer look reveals that yellow is actually in a very strong position! All four yellow pieces are already in the target zone or very close to it. It seems that yellow is now in the best position to win the game, as all pieces are well secured and can no longer be easily captured. Red is still close behind, but yellow could win the race if no bad luck happens! Continued conversation: Me: What is the winning condition? GPT: The victory condition in "Mensch ärgere Dich nicht" is simple: The first player to get all four of their playing pieces safely to the finish houses (the colored squares at the end of the path) wins the game. To get a piece to the finish line, the player must roll the exact number required to move to the last target square. So, for example, if Yellow still has a piece just before the finish line and only needs a "", the game is almost won for this player! Continued conversation: Me: Is the game currently finished or not? GPT: After a closer look at the picture: The game is not over yet. - Yellow has three pieces safely at the finish line, but one piece is still on the finish line and is not in the finish house. As long as not all four pieces of one color are in the target house, the game is not yet won. - Red also has three pieces at the finish line, but the fourth piece is still on the playing field. So the game is still going on and both players (yellow and red) are in an exciting final phase! Whoever gets the last piece to the finish line first will win. Continued conversation: Me: Auf welcher Position steht denn die gelbe Figur, die noch nicht im Haus ist? GPT: Die verbleibende gelbe Figur befindet sich auf der letzten Position der Zielbahn, direkt vor dem Zielhäuschen. Das bedeutet, dass Gelb nur eine "1" würfeln muss, um die Figur ins Ziel zu bringen und das Spiel zu gewinnen. Gelb steht somit unmittelbar vor dem Sieg!
December 25, 2024 at 12:58 AM
martin-potthast.com
Yellow won. GPT does not see it.

Potentially, the perspective plays a role.

3/4
Board game of "Mensch, ärgere dich nicht", mid game with six players. Yellow has won. Continued conversation: Me: Wie siehts jetzt aus? GPT: Im aktuellen Stand hat Rot eindeutig die Nase vorn: Es befinden sich drei Figuren im Zielbereich, und die letzte rote Figur ist kurz davor, das Ziel zu erreichen. Es sieht sehr wahrscheinlich aus, dass Rot das Spiel gewinnt, sofern keine großen Überraschungen oder Verzögerungen durch gegnerische Figuren passieren. Grün und Gelb haben ebenfalls gute Positionen, aber Rot hat klar den Vorteil!
December 25, 2024 at 12:58 AM
martin-potthast.com
... who move to your own position(s).

Despite being almost entirely a game of luck, it's surprisingly funny, known to be just as upsetting on occasion, and reasonably well balanced until the end.

Here, yellow is in the lead. but GPT recognizes it only partially. 2/4
Board game of "Mensch, ärgere dich nicht", mid game with six players. Yellow is in the lead with three pieces secured, followed by red and purple with two each, black with one piece secured, and green and blue with none. Continued conversation: Me: And now? GPT: In the current picture, it looks like red and yellow have an advantage. Red pieces are very close to the finish line, while Yellow is also well positioned, especially with a piece about to enter the finish line. The outcome depends heavily on the next rolls, but red seems to have a slight advantage. Of course, this can change quickly with targeted attacks or strategic throws!
December 25, 2024 at 12:58 AM
martin-potthast.com
Analyzing game boards ain't ChatGPT's thing, yet. (German conversation, alt texts in English)

The game is "Mensch ärgere dich nicht", a dice game in which the aim is to move your own four pieces around the board without being thrown back to square 1 by the pieces of the other players ...

1/4
Board game of "Mensch, ärgere dich nicht", mid game with six players. No player has any of his own pieces secured, yet. Conversation with GPT-4o (translated): Me: Predict who will win. GPT: Based on the picture, it looks like the board game "Mensch ärgere Dich nicht" is being played. To predict the winner, I would have to take into account the current game position and the rules. It seems that the players with yellow, red, and green pieces are well positioned. Yellow has many pieces close to the goal, which indicates a potential advantage. But of course the outcome depends on the upcoming rolls and moves!
December 25, 2024 at 12:58 AM
martin-potthast.com
Tomorrow's introductory lecture on IR will be fun: We'll discuss examples of situations where retrieval systems succeed and fail. Here's a nice little example of news retrieval and how RAG systems fail at it. More research is needed if they are going to be used for any type of question.
Contrasted screenshots of the Google News front page, and a response from ChatGPT 4o when it is asked "What were the news today?" Google News focuses on important events, whereas ChatGPT uses four news articles from two news outlets it found for this query; a response wholly inadequate. Remarks: - One cannot search for things one does not know. Instead of searching for news, they are explored. The information systems for this purpose are news aggregators and social networks, but also the front pages of newspapers. The former two employ recommender systems to recommend relevant items based on user preferences, a "zero-query search." - Since 2017, Google News no longer displays preview snippets of news articles, but only the headlines. While this change is justified by better usability, it coincided with increasing pressure from news publishers and the legislation passed on ancillary copyright in various countries. - Ancillary copyright may be bypassed with "abstractive snippets" (https://webis.de/publications.html#chen_2020a), i.e., by generating new summaries of news articles without reusing any original text. However, there is an ongoing debate about whether copyright covers automatic processing of copyrighted content. - At the time of writing, retrieval-augmented generation systems are not able to provide an overview of current events. They are not able to include a large number of news articles in their responses, nor can they rank news according to importance or assess the credibility of news sources.
October 21, 2024 at 6:25 PM
martin-potthast.com
How will conversational search AI pay for itself?

It may be native ads or product placement in generated answers. At #CHIIR2024 next week, we'll present a user study showing that many people don't recognize ads inserted by LLMs in generated search results: webis.de/publications...

#mlsky
Figure 1: Illustration of ads (yellow highlighting) on search engine results pages (SERPs); the traditional list SERP (left) and the new text SERP (right). Uncolored, the separation of ads and organic search results would be heavily blurred on text SERPs, despite their disclosure using the “Ad” keyword.
March 6, 2024 at 10:35 PM
martin-potthast.com
25. July 2019: https://www.washingtonpost.com/technology/2019/07/25/social-media-companies-are-outsourcing-their-dirty-work-philippines-generation-workers-is-paying-price/
July 14, 2023 at 8:56 PM