The Aaronson Oracle

An interactive program that predicts whether you will press one of two keys builds on a seventy-year line of lab machines that outguessed human choice. Early work by David Hagelbarger produced SEER, which tracked three yes/no facts about the previous two rounds and used eight small counters that purposely “forgot” ancient history; Claude Shannon simplified the idea and added a hardware source of randomness, analyzing that his machine could be beaten 3:1 in principle yet still beat real people repeatedly. Reimplementations and museum records show these simple pattern-matchers beat human opponents much more often than chance (Shannon’s exhibit tally ~59% machine wins; a 2020 software replay between the two designs averaged 55.8% vs 44.2%). Scott Aaronson’s classroom program similarly reached 70-80% prediction rates, though one student achieved 50% by producing truly random choices.

The prediction method is straightforward and concrete: every keypress is recorded as six entries corresponding to context runs of length 0-5 (so a full session encodes 63 possible contexts). To guess the next press the program looks for the longest recent run it has seen at least twice and takes the majority-next move; if that run is unseen, seen once, or tied it shortens the context progressively down to no context and finally flips a coin. That “shorten-when-unsupported” rule prevents overfitting to single coincidences and is the key practical trick distinguishing this implementation from the earlier relay machines.