AI girls create a system to predict horse races and win a bet with odds of 2275.3 to 1
When I exhibited a system where an AI girl cheerfully predicts horse races, I won a bet with a 2000 to 1 odds (1) Design Concept #LLM - Qiita
https://qiita.com/oktamajun/items/34c2287a9f4d9ac9c298
GALLOPIA, created by Tamaki, is a group chat where eight AI girls live, and when you specify a race and send a request for a prediction, they will discuss and suggest recommended betting tickets. The purpose of this system is to 'enjoy the AIs talking to each other with some dubious knowledge and claims like, 'This horse will win!' 'No, this one will win!'' and not to create a serious horse racing prediction AI. Therefore, Tamaki wrote, 'Probably a serious prediction AI would use DNN (deep neural network) instead of language-based LLM, and that would be more correct. Probably. (I don't know the details).'
In GALLOPIA, there are eight AIs, each in charge of a different field. Each girl's role is as follows:
Each girl presents her own opinion on the analytical field she is in charge of, Shirabe, who is in charge of data, keeps an eye on their progress, Fumino, who is in charge of writing on the board and as secretary, organizes the information, and finally, the main character, Michiru, presents her predictions for the race.
Below is a video of GALLOPIA, created by Tamaki, in action. Several girls express their opinions, with Michiru in charge of prediction judgment, Chiaki in charge of pedigree analysis, Yukari in charge of jockey and related party analysis, and Hayate in charge of lap and training analysis. In the end, Michiru summarizes the opinions of the girls and presents five betting patterns, and it can be confirmed that the betting ticket with odds of 2275.3 that Tamaki actually won is included in these predictions.
GALLOPIA in action (24/11/16 Tokyo 12R predictions reproduced) - YouTube
The models that Tamaki used to realize GALLOPIA and their uses are as follows. Tamaki wrote, 'There are many use cases where it is better to combine multiple LLMs rather than a single LLM.'
gemini-1.5-pro: Extracting advantages and disadvantages from large amounts of data
GPT-4o: deriving advantages and disadvantages from small amounts of data, Michiru's decision-making
GPT-4o-mini: Writing summary of discussions on a whiteboard
claude-3-5-sonnet-20241022: Creating the outline of each person's argument
claude-3-haiku-20240307: Character dialogue creation, individual summary
In a field with a lot of thought-provoking data, such as horse racing predictions, it is a bad idea to feed all the information to a large-scale language model (LLM) and try to get an answer in one go, Tamaki explains. 'If the prompt exceeds a certain length, the performance on the task tends to deteriorate.' For this reason, GALLOPIA improves the quality of thinking by calling the LLM multiple times with short prompts.
In addition, because horse racing predictions are based on various specialized knowledge such as pedigree, jockeys, and training, the method of using multiple LLMs seems to have been particularly effective. In the case of GALLOPIA, it was thought that by analyzing each domain knowledge in a separate LLM and then merging them later into another LLM, it would be possible to output more reliable predictions.
Tamaki calls this method of using multiple LLMs to output answers in an interactive format 'multi-agent LLM orchestration' and has explained its advantages and disadvantages on his blog, so if you're interested, be sure to check it out.
Related Posts:
