In brief
In one flash-anzan performance, 30 three-digit numbers were added in 3.29 seconds. Years of specialised practice make this possible. But a speed limit claimed for all human thought has to account for the fastest human performances too.
Picture the task
Imagine these 30 three-digit numbers appearing one after another. A top flash-anzan competitor added 30 such numbers in 3.29 seconds:
384 + 719 + 256 + 903 + 147 + 628 + 571 + 842 + 395 + 164 + 736 + 219 + 487 + 950 + 302 + 675 + 813 + 426 + 591 + 748 + 263 + 809 + 135 + 674 + 927 + 318 + 452 + 786 + 541 + 690 = ?
Illustrative numbers, not those from the reported performance.
What do bits per second mean here?
One bit represents a choice between two equally likely possibilities; eight bits make a byte. A Wi-Fi connection running at 100 megabits per second transfers 100 million bits each second, and a 3 MB phone photo contains about 24 million bits. In mental calculation, the same unit describes uncertainty in the numbers a person sees and in the intermediate results they work through. This makes information rates comparable as quantities, while showing how differently the information is represented: digital traffic carries encoded data; a trained calculator can update a compact mental abacus as each number appears.
What was studied?
The paper analyzes high-level mental calculation using results from international competitions and record attempts. For a concrete example, the Japan Flash Anzan Association reports a 2026 performance by Yuichiro Takakura: 30 three-digit numbers added in 3.29 seconds. The association describes this as a Guinness application record, not yet a confirmed Guinness record. The study relates the information needed to perceive numbers and perform calculations to elapsed time. Its reported rates are model-based estimates, not recordings of individual neurons or a direct measure of all thought.
What did the analysis find?
The fastest short calculations imply rates far above some earlier estimates for high-level cognition. The model reaches more than 215 bits per second: roughly 125 for perceiving the numbers and 90 for the intermediate operations of the algorithm. High rates also persist in some longer performances. The competitors are not merely reading an answer from memory; they are continually updating a learned mental representation. The 3.29-second performance occurred after the paper's 2026 revision and illustrates the task rather than contributing to its calculations.
What does it not show?
The measured tasks do not imply that everyone thinks at 215 bits per second, or that anyone can reach this rate without extensive training. A proposed limit on all human cognition must nevertheless account for the highest observed human performances. The numerical estimates depend on how information and calculation steps are modelled.
Why bring this to a talk?
We all make rapid decisions, for example while driving on a motorway. Perception, prediction, habit and movement are intertwined there, making a meaningful bit rate difficult to calculate. Timed arithmetic provides a task in which the input and the steps are much easier to count. It gives a measurable glimpse of how quickly a trained human mind can operate on an internal model. Computers and language models process vastly more digital data, but the bit rate of their hardware alone does not describe how efficiently they—or we—turn representations into useful decisions. This question links the research to my keynotes on learning, attention and information processing.