This is a working research archive for an essay on my personal website, on what machine-chess means to human players.
The other day I was discussing AI and how it may change how we do science in the next few years with my friend Becca. She said something along the lines of: 'what's important in the future few years is understanding what the value of human + AI can be, not what the value of AI can be'. I think chess is a good place to start understanding this since chess players have already spent decades adapting to engines and using them to improve.
This reminds me of Garry Kasparov's 2010 account of the 2005 PAL/CSS Freestyle tournament (standard chess with engine assistance), won by two amateurs working with three computers: 'Weak human + machine + better process was superior to a strong computer alone'. What did the humans add, and how has that changed as the engines improved?
Gaessler and Piezunka's Training with AI (2023) found that access to chess computers improved players' tournament performance, whilst reducing their ability to exploit human mistakes.
The outline pairs chess and workplace evidence; the claims ledger gives source pages and limits for each claim.
The earlier plot showed the yearly median first position absent from earlier recorded games, using Lumbra's archive. Most of the long-run rise disappears when the reference is held at 1,000 games; the files are retained in the local archive.
The 2005 winners described how they chose between engine suggestions. The clock reconstruction checks their account against recorded move times and modern searches.
The first run covered eight games from the four matches that I know well (1984-85, 2000, 2008, and 2023). The later release comparison uses 500 independently selected positions; methods gives the sampling and reproduction commands.
Check whether the variety of named openings and the variety of positions actually reached at fixed plies have moved together across the archive, or diverged. My guess is that the vocabulary has expanded whilst the sentences have shortened; but the divergence would need to survive rarefaction and plausible ply choices before I trust it.