On Mon, Mar 23, 2026, 1:54 PM James Bowery <[email protected]> wrote:

>
>
> On Sat, Mar 21, 2026 at 12:28 PM Matt Mahoney <[email protected]>
> wrote:
>
>> ...
>>
>> There's something I'm not getting. Why does the brain need 10^15
>> synapses to store 10^9 bits? Maybe it's a speed optimization, like how
>> a server farm has a million copies of Linux, or your body has 10^13
>> copies of your DNA. Or is it something else? Is it the reason we
>> didn't solve AI in 2000?
>>
>
> I suspect it has to do with Bekenstein bound placing data points in such a
> high dimensional space that they are all on a surface where they can be
> treated as orthogonal.
>

How is that so? I realize that random bit vectors all have an average
Hamming distance of n/2, which puts them all on a hypersphere surface
surrounding any one of them. But word vectors are not like that. Some are
more correlated than others. They would have to be, because otherwise text
would not be predictable and we wouldn't have AI.

I realize that other parts of the brain are highly repetitive, like
thousands of copies of line and edge detectors in the visual cortex, or
thousands of motor neurons controlling the same muscle. Language evolved
relatively recently and there is not a lot of evolutionary pressure to
optimize it. It uses maybe 10% of our brain, or 2% of resting metabolism.

The Bekenstein bound is different. Space only has 3 dimensions. It might
explain the size of the proton* but I don't see how it explains language.
We already have LLMs that are not far off the 0.3 bits per parameter stored
in a Hopfield net.

* The Bekenstein bound of the Hubble radius is A/ln(16) of its surface area
in Planck units, or 2.95 x 10^122 bits. This is about the number of protons
or neutrons that would fit inside, which is a strange coincidence given
that the number depends only on h, G, c, and the age of the universe.


------------------------------------------
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/Tc9fe35df94409188-Mad595c3e3a8538380e77b94f
Delivery options: https://agi.topicbox.com/groups/agi/subscription

Reply via email to