Mathematician Checks the Numbers Behind OpenAI's "Breakthrough" Fourier Algorithm

OpenAI, the company behind ChatGPT, published an improvement to the Fourier transform. This is a mathematical method for extracting frequencies from signals, which is why it is used in audio and video processing. The current benchmark algorithm for computing the Fourier transform is the Cooley-Tukey algorithm, also known as the Fast Fourier Transform (FFT), published in 1965. That is the algorithm OpenAI's new one gets compared against. Professor Ben Recht of Berkeley went through the details of the paper. He found that for the new algorithm to actually beat the 1965 version, a hidden coefficient in the formula would need to grow to 10^100. And even after that growth, the 1965 algorithm would still be 10^99 times faster than the new one. So he called it a "malicious exploit."

We need to check every number before we believe it. This is a good example of why. A factor of 10^100 is larger than the total number of atoms in the observable universe. That is how a "breakthrough" turns into something completely useless. OpenAI recently published a whole batch of math claims like this one. So always check what is actually hiding behind the headlines.