Show HN: How random are you?
Humans are surprisingly bad at generating truly random sequences, a fact highlighted by a web game that tries to predict numbers people speak or type and compares them to statistical baselines. Commenters explore why people over-alternate, how external sources like π or coin flips can improve unpredictability, and how feedback from the system itself lets players adversarially game the predictor. The thread also touches on implementation details such as speech-recognition models and relates the concept to classic probability anecdotes, mentalism tricks, and earlier “mind-reading” experiments like the Aaronson oracle.
Overall Reactions and Concerns
- Many found the game clever and entertaining, especially the “house vs human” framing and dual typed/voice modes.
- A subset was uneasy about providing voice samples, raising concerns about potential model training or social engineering, though others noted there are already abundant voice datasets.
Human Randomness Biases
- Multiple comments referenced prior experiments showing humans are poor random generators, especially over-alternating choices and avoiding long runs.
- The classic coin-flip demonstration (real vs invented sequences) was discussed as evidence that humans underproduce streaks.
- Some argued that while all sequences are equally likely for a fair RNG, human-generated sequences differ statistically and can be detected post hoc.
Player Strategies and Outcomes
- Many tried to “beat” the predictor by:
- Using external randomness (digits of π or Fibonacci, math constants, clocks, word lengths, PRNG automation).
- Playing adversarially by reacting to the site’s guesses, or intentionally lagging/mirroring them.
- Repeating a single number many times, or alternating from ends of the number range.
- Several reported very low prediction rates (near or below the blind 11% baseline) and were told they were among the “most random” relative to a referenced psychology study.
- Others found the predictor quickly locked onto certain patterns (e.g., repeated digits, lag-following), though a few said it sometimes “stuck” on one guess, which the author attributed to a heuristic being overweighted.
Randomness vs Unpredictability
- Some stressed the difference between true randomness and mere unpredictability to a particular algorithm.
- Proposed schemes (e.g., combining two internally chosen digits with modular arithmetic) were debated; critics noted they are deterministic and only help insofar as they fool the specific pattern detectors.
- Classic debiasing ideas like von Neumann’s coin trick were raised to illustrate extracting unbiased bits.
Implementation, Comparisons, and Improvements
- The site reportedly uses on-the-fly n‑gram Markov models plus heuristic predictors; it was inspired by another well-known “oracle” that predicts keypresses.
- Speech recognition is implemented with a lightweight model chosen for browser compatibility; some criticized its accuracy and suggested alternatives, noting model size vs performance trade-offs.
- Several pointed to similar or related games (number-uniqueness contests, prior “mind reader” or RPS competitions) and suggested adding a stricter “hard mode” without visible feedback to more closely match original lab conditions.