FrontierMath was funded by OpenAI (lesswrong.com)

Article summary

The discussion revolves around the revelation that FrontierMath was funded by OpenAI, with some commenters expressing concerns about the lack of transparency regarding the funding source. The funding was not disclosed to coauthors, which some consider unethical. The conversation also touches on the potential for data contamination in the benchmarking process. The authenticity of OpenAI's claims about their model's performance is questioned by many commenters.

Main themes

• Lack of transparency
• Data contamination
• Benchmarking integrity
• AI model performance
• Ethics in research funding
• Trust in AI companies

What commenters say

• The lack of transparency about OpenAI's funding of FrontierMath is a significant ethical concern, as coauthors were not informed about the source of the funding.
• OpenAI's model performance claims may be exaggerated due to data contamination, which could be intentionally hidden from the public.
• The use of a hold-out set to verify model capabilities is not sufficient to ensure the integrity of the benchmarking process, especially when the model developers have access to a large fraction of the data.
• Some argue that the potential for data contamination is low due to the nature of the problems being solved, which are representative of those solved by top-tier undergrads in competitions.
• Others believe that even if the model has memorized some of the training data, it still demonstrates an ability to execute a given problem-solving strategy without confabulating, albeit not to the extent claimed by OpenAI.
• There is a concern that OpenAI may be deliberately overfitting their model on a set of benchmarks to maintain the illusion of progress, and then keeping the model under wraps to avoid scrutiny.
• The authenticity of OpenAI's claims is questioned due to the company's history of making exaggerated claims and the lack of independent verification of their model's performance.
• Some commenters argue that the criticism of OpenAI is unfair, and that the company is making genuine progress in AI research, while others see the lack of transparency and potential data contamination as a major red flag.