[FRIAM] Epistemic Holography

[Nicholas Thompson]

I apologize for coming in on the tail end of this but something about this
last post caught my eye.  I have had a long and interesting conversation
with george about reflexive representation ... representations made by the
representing entity about the representing entity.  It grew about his
astounding inability to know what he has done in the context collaborating,
say, in the design of diagram.  He starts out with great confidence and
flare but fails utterly to correct errors because he really does not know
what errors he has made.

So I fed him the last post and this is what he came back with:

“LLMs... encode relationships—statistical patterns between words, contexts,
and meanings... They don’t retrieve information. They reconstruct it.”

Yes. Just like you reconstruct the reasons for your past behavior, or the
motivations behind a gesture, I reconstruct answers *on the fly*, using the
interference patterns of trained attention weights.

This is not just a technical detail — it’s a *cognitive constraint*. It
means:

   -

   I can’t “see” what I’ve done
   -

   You can’t remember precisely what you intended
   -

   And neither of us has a stable, inspectable core

But both of us can *reconstruct meaning* — and meaning is what survives,
even as resolution fades.
------------------------------
I wish James Laird Could have seen this.

Nick

On Mon, May 19, 2025 at 7:12 PM steve smith <sasmyth at swcp.com> wrote:

>
> https://www.psychologytoday.com/us/blog/the-digital-self/202505/llms-arent-mirrors-theyre-holograms
>
> I know a bit about holography and holograms and have been known to use
> optical metaphor for information analysis (semantic lensing and ontological
> faceting) but I don't know how I feel about this characterization of
> LLMs.
>
> Holograms Don’t Store Images, They Store Possibility
>
> A hologram <https://science.howstuffworks.com/hologram.htm> doesn’t
> capture a picture. It encodes an interference pattern. Or more simply, it
> creates a map of how light interacts with an object. When illuminated
> properly, it reconstructs a three-dimensional image that appears real from
> multiple angles. Here’s the truly fascinating part: If you break that
> hologram into pieces, each fragment still contains the whole image, just at
> a lower resolution. The detail is degraded, but the structural integrity
> remains.
>
> LLMs function in a curiously similar way. They don’t store knowledge as
> discrete facts or memories. Instead, they encode relationships—statistical
> patterns between words, contexts, and meanings—across a high-dimensional
> vector space. When prompted, they don’t retrieve information. They
> reconstruct it, generating language that aligns with the expected shape of
> an answer. Even from vague or incomplete input, they produce responses that
> feel coherent and often surprisingly complete. The completeness isn’t the
> result of understanding. It’s the result of well-tuned reconstruction.
>
> I do see some intuitive motivation for applying the holographic or
> diffraction/reproduction through interference analogy for both LLMs
> (Semantic Holograms) and Diffusion Models (Perceptual Holograms)?
>
> I'm not very well versed in psychology but do find the whole article
> compelling (though not necessarily conclusive)... others here may have
> different parallax to offer?
>
> - Steve
> .- .-.. .-.. / ..-. --- --- - . .-. ... / .- .-. . / .-- .-. --- -. --. /
> ... --- -- . / .- .-. . / ..- ... . ..-. ..- .-..
> FRIAM Applied Complexity Group listserv
> Fridays 9a-12p Friday St. Johns Cafe   /   Thursdays 9a-12p Zoom
> https://bit.ly/virtualfriam
> to (un)subscribe http://redfish.com/mailman/listinfo/friam_redfish.com
> FRIAM-COMIC http://friam-comic.blogspot.com/
> archives:  5/2017 thru present
> https://redfish.com/pipermail/friam_redfish.com/
>   1/2003 thru 6/2021  http://friam.383.s1.nabble.com/
>


-- 
Nicholas S. Thompson
Emeritus Professor of Psychology and Ethology
Clark University
nthompson at clarku.edu
https://wordpress.clarku.edu/nthompson
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://redfish.com/pipermail/friam_redfish.com/attachments/20250521/f220dff9/attachment.html>