Thanks to @[email protected] for the links!
Here’s a link to Caltech’s press release: https://www.caltech.edu/about/news/thinking-slowly-the-paradoxical-slowness-of-human-behavior
Here’s a link to the actual paper (paywall): https://www.cell.com/neuron/abstract/S0896-6273(24)00808-0
Here’s a link to a preprint: https://arxiv.org/abs/2408.10234
There was a study in 2019 that analyzed 17 different spoken languages to analyze how languages with lower complexity rate (bits of information per syllable) tend to be spoken faster in a way that information rate is roughly the same across spoken languages, at roughly 39 bits per second.
Of course, it could be that the actual ideas and information in that speech is inefficiently encoded so that the actual bits of entropy are being communicated slower than 39 per second. I’m curious to know what the underlying Caltech paper linked says about language processing, since the press release describes deriving the 10 bits from studies analyzing how people read and write (as well as studies of people playing video games or solving Rubik’s cubes). Are they including the additional overhead of processing that information into new knowledge or insights? Are they defining the entropy of human language with a higher implied compression ratio?
EDIT: I read the preprint, available here. It purports to measure externally measurable output of human behavior. That’s an important limitation in that it’s not trying to measure internal richness in unobserved thought.
So it analyzes people performing external tasks, including typing and speech with an assumed entropy of about 5 bits per English word. A 120 wpm typing speed therefore translates to 600 bits per minute, or 10 bits per second. A 160 wpm speaking speed translates to 13 bits/s.
The calculated bits of information are especially interesting for the other tasks (blindfolded Rubik’s cube solving, memory contests).
It also explicitly cited the 39 bits/s study that I linked as being within the general range, because the actual meat of the paper is analyzing how the human brain brings 10^9 bits of sensory perception down 9 orders of magnitude. If it turns out to be 8.5 orders of magnitude, that doesn’t really change the result.
There’s also a whole section addressing criticisms of the 10 bit/s number. It argues that claims of photographic memory tend to actually break down into longer periods of study (e.g., 45 minute flyover of Rome to recognize and recreate 1000 buildings of 1000 architectural styles translates into 4 bits/s of memorization). And it argues that the human brain tends to trick itself into perceiving a much higher complexity that it is actually processing (known as “subjective inflation”), implicitly arguing that a lot of that is actually lossy compression that fills in fake details from what it assumes is consistent with the portions actually perceived, and that the observed bitrate from other experiments might not properly categorize the bits of entropy involved in less accurate shortcuts taken by the brain.
I still think visual processing seems to be faster than 10, but I’m now persuaded that it’s within an order of magnitude.
Thanks for the link and breakdown.
It sounds like a better description of the estimated thinking speed would be 5-50 bits per second. And when summarizing capacity/capability, one generally uses a number near the top end. It makes far more sense to say we are capable of 50 bps but often use less, than to say we are only capable of 10 but sometimes do more than we are capable of doing. And the paper leans hard into 10 bps being a internally imposed limit rather than conditional, going as far as saying a neural-computer interface would be limited to this rate.
“Thinking speed” is also a poor description for input/output measurement, akin to calling a monitor’s bitrate the computer’s FLOPS.
Visual processing is multi-faceted. I definitely don’t think all of vision can be reduced to 50bps, but maybe the serial part after the parallel bits have done stuff like detecting lines, arcs, textures, areas of contrast, etc.
The problem here is that the bits of information needs to be clearly defined, otherwise we are not talking about actually quantifiable information. Normally a bit can only have 2 values, here they are talking about very different types of bits, which AFAIK is not a specific quantity.
This is of course a thing.
I think everyone agrees on the definition of a bit (a binary two-value variable), but the active area of debate is which pieces of information actually matter. If information can be losslessly compressed into smaller representations of that same information, then the smaller compressed size represents the informational complexity in bits.
The paper itself describes the information that can be recorded but ultimately discarded as not relevant: for typing, the forcefulness of each key press or duration of each key press don’t matter (but that exact same data might matter for analyzing someone playing the piano). So in terms of complexity theory, they’ve settled on 5 bits per English word and just refer to other prior papers that have attempted to quantify the information complexity of English.