From the video, I think they used mathematical optimization. Multiobjective vectorial optimization if I had to guess. The big breakthrough here is that instead of OCR'ing the image they tried to rerun the captcha construction algorithm controlling the random choices the algorithm makes. Each choice is a variable here. Them you implement a function that measures how close this variables get to the CAPTCHA image. Now you use optimization to get to the global minimum of this function.
At least that is how I would have done it.