Well, I made up my appointment from yesterday today.
However this involves hours of sitting still, and my
laptop was not an old computer, and the situation was
such that I didn't have a free arm to write with.

On the other hand I had a good conversation about
climate. This person pointed out that her parents had
in effect been environmentalists in that plastic bags
had not come into widespread use, neither cars really.

They had the same sentiment that environmentalism is
doomed because other countries will put middle class
short term greed over anything else. So while it is a
nice thought to preserve the environment, the world is
doomed anyway. Another person I spoke to had had a 
similar view.

While wiling away the hours penless, I had the thought
to put my bunzy hash-tables together with my prover
planner plans.

A recap: I drafted a package, bunzy hop (binry-hop,
bunny-hop) in which the :test was a hopfield net
lookup of the key (against other keys). This data
structure really is nuts. And by nuts I mean deep
and non-linear.

Yesterday I was pondering - given how robot labor is
almost the same as human labor anyway, except less
dynamic and faster, but not by a lot, what would a
robot planner want to be like, left to their own
devices? Planning problems are the same complexity
for a robot as for a human.

Since I am using an automatic prover as a planner
/ model, this case is relevant: Since the advent of
automatic provers, first order proofs started
emerging- things that are mathematically true, but
that are so large they are not hand-verifiable by
humans. Such as the four-color-theorem. (Later work
produced a smaller four-color-theorem proof, which
would in principle only take two solid months of
expert human effort to manually verify, assuming no
mistakes, though nobody has done it).

So how can I get my theorem planner to produce plans
that are deliberately out of human scope but within
comfortable robot scope?

Classical hopfield nets have cryptographic properties.
Basically impossible-for-human plans can be created by
using Hopfield nets if only the robot ever knows about
the Hopfield net per se. Kind of cheating, but the
technology is having proofs that are verifiable, but
not ~ decryptable by humans.

My simple case is like this:

(1) *classical* binry-hop hash. The bit-arrays are 
the keys, :test is Hopfield net convergence 
(then 'equal).
The values of the hash-table are lambdas; an action
to make in the context of the key being sensed.

(2) The robot takes a picture of its surroundings.
Could be a feature vector, could be a binarized
image or a sound, it doesn't matter.
After converging, the original picture is added
(in the classical hopfield net sense) to the matrix
calculation, and the image is discarded (= lost).

The image is still inside the hopfield net, and in
principle could be matched, and affects its
behaviour, and, I think, can be used in theorems
involving lookups from this hash, but the system
is also cryptographically opaque.

Aside: My sketch of how to encrypt a pictrogram
(or anything) into a classical hopfield net:
Have the pictogram. As long as it fits in the
net, it's fine.
Have a bernoulli sequence the same size.
logand the pictogram and bernoulli sequences.
Add this destination to the classical hopfield
net.

Now a naieve but accurate sketch of the pictogram
will be a 50% (ish) match with that pictogram,
presumably closer to it than anything else, hence
it should converge there.

The pictogram could be attached to a secret
message.

The hopfield net is then stuffed to capacity with
random garbage, or other pictogram secret messages.

Anyway, simply using this cryptographic classical
hopfield net approach in the development of a plan
(or other theorem) is a shortcut to being robots-
only.

In order to not be reliant on the cryptography to
gain opacity, the snapshot-taking could also be
something unattainable by humans (pretty much
anything with a high data density but low enough
for robot handling).

(Olde robot handling).

If we have our bunny robot advance forward in time,
making and executing plans that modify itself, in
a manner opaque to humans (by data) and other
robots (the cryptographic complexity beats both
humans and robots).

We could also do the same sans cryptography with a
super-high capacity dense hopfield net (which do
not have the cryptographic property, at least in
simple update formulation).

My intuition is that this idea is crazy beyond
normal crazy. I do owe 40-50 years of literature
review on surrounding topics though... Which I
guess I could be doing on printed paper.