Frank de Groot

In the early noughties, I showed that Bayesian pattern learning works surprisingly well to predict the played moves in 42% of positions of previously untrained-for professional Go games. I published my algorithm in great detail, as well as all essential info on my blog on its technical details and speedup techniques (Mersenne Twister random, Zobrist hashing, Hamming distance, MMX assembly speedups etc.). Microsoft Research Cambridge used that information to replicate my improved Bayesian learning algorithm and after they credited me in their published paper, Google contacted me, offering a job interview.

My Machine Learning software for the game of Go (2005)
Download it for free (317 MB) at http://additionalintelligence.com/MoyoGoStudio.zip

Tore Graepel, the team lead for Microsoft’s team that replicated my work, became the CTO for DeepMind, which integrated my work into their plausible-move generator for AlphaGo.

I now work on General AI. This is my R&D machine, standing in an airconditioned storage room into which I remote with AnyDesk (due to the noise and cold):

128 EPYC cores, 2 TB RAM, 400 TB in 3 x RAID6, dual UPS, 16-tape library. Airconditioned.