Some Lessons from the OpenAI FrontierMath Debacle

They really could have done all this in the open (test database aside) and would have had same outcome.

They gained literally nothing and took another knock on rep. It’s not like it was a mystery that they’re working on reasoning models.

The mathematicians are training AI for a pittance and are not told about it.

It is their fault though to participate in writing a test set in the first place that will obviously be used in a non-scientific manner. "Trust-me-bro" science on other people's closed servers.

But AI ethics is a dangerous field that can get you suicided.

> Also, a number of the mathematicians who worked on FrontierMath would possibly have not contributed to this if they knew about the funding and exclusive access. It feels concerning that OpenAI could have inadvertently paid people to contribute to capabilities when they might not have wanted to.

This is odd. The issue isn't o3's "capabilities" or AINotKillEveryoneism, it's the spreading corrosion of OpenAI's dishonest marketing. Presumably those contributors thought they were making a good benchmark. Instead they got misled into making an infomercial.

This specifically hurts Terence Tao, because it raises the question about whether or not he knew that OpenAI had privileged access. Epoch and OpenAI tarnished his reputation in order to improve o3's reputation. Truly despicable.

We now know the exact amount of solutions OpenAI got beforehand: 23-25%

Because 2% they could have solved without cheating