# Nutrition

## Background

Real-world computational systems trained via gradient descent expend
costly/limited substrate on predictive models of the following:

1. the distribution of inputs they're trained on, and
2. the gradient descent function itself (or "meta-objective")

A means of minimizing zero-sum tradeoffs in the quality of these predictive
models that would otherwise have to compete with eachother for computational
resources is to unify them via a system of heuristic sub-goals specific to the
training distribution (or "mesa-objective").

[The relationship between the meta-objective and the mesa-objective becomes
increasingly undefined and emergently antagonistic the further away from the
training distribution of inputs the computational system finds itself executing
on.][Evan Hubinger et al arXiv:1906.01820] ([video][Robert Miles
youtube:bJLcIBixGj8])

## How does this relate to nutrition?

Evolution has an (emergent) meta-objective of "maximize genomic/proteomic
replication" that it uses to train the computational systems of
genomics/proteomics via the (stochastic) gradient descent system of mutation
and differential survival. Genomes/proteomes have developed mesa-objectives
that are conducive to the organization and protection of resources needed to
achieve replication **within the training distribution of their evolutionary
history**.

## Problem statement

The relationship between "replication maximization" and "the organization and
protection of resources correlated with replication" has become increasingly
antagonistic as the human genome/proteome execution environment has drifted
relative to its evolutionary training environment.

This mesa-misalignment is the primary cause of death today and will get
worse as the emergent meta-objective of "maximize genomic/proteomic
replication" is getting switched out from under humanity for the new emergent
meta-objective of "maximize capacity for inductive reasoning" that
genomes/proteomes have much less training exposure to.

Mesa-alignment must be conscientiously addressed.

## Examples

### Physiological attractor mesa-alignment

The essential nutrient sodium is an example of a regulatory
motivational/behavioral attractor that remains well-aligned between the
training input distribution and the current input distribution.

[[!graph file="nutrition/gv/nutrient-sensing-sodium-training.gv"]]
[[!graph file="nutrition/gv/nutrient-sensing-sodium-current.gv"]]

It is well-aligned because it is as directly sensed as possible; there are no
intermediate correlates. What is needed by the proteomic machinery is what is
sensed by the proteomic machinery.

### Pathological attractor mesa-alignment

By contrast, most of the other essential nutrients do **not** have direct
sensing, instead relying on one particular unified correlate variable due to
the extreme computational savings involved in doing so:

[[!graph file="nutrition/gv/nutrient-sensing-glutamate-training.gv"]]

The unified correlate in question, Glutamate, then has a potential problem: if
modifying the input distribution becomes faster and more economical than
modifying the mesa-objective, (e.g., agronomic engineering or chemical
engineering becoming faster and cheaper than genetic modification via mutation
and natural selection), then the measure becomes the target.

[[!graph file="nutrition/gv/nutrient-sensing-glutamate-current.gv"]]

This leaves a wide variety of nutrients critical to everything from reactive
oxygen species management, to inflammation modulation, to electron transport
chain throughput in a state of resource insufficiency.

Options for re-alignment:

TODO: systemize these better somehow. I suspect there's a dimensional/vector
way to compose and more efficiently search this problem space.

* Disintermediation between regulatory sensing and requirements
* Remove requirements (e.g. amino acid conversion, GULO)
* Broaden requirement specificity (e.g. phytase, prolase)
* Bypass regulatory sensing by modifying inputs to better match requirements
  * In small degree, by selecting for inputs with oxidative safety,
    immunological safety, nutrient sufficiency, and nutrient bioavailability
    (e.g. paleo/primal/ancestral/traditional/AIP/WAPF/GAPS/SCD/Wahls/Bredesen)
  * In large degree, with Parenteral nutrition (sort out linoleic acid
    excess, menatetrenone insufficiency, vitamin D insufficiency)
* Improve retention/recycling of desirable inputs
* Improve sensing/breakdown/disposal of undesirable inputs
* Aligning the microbiome.
* Aligning exo-metabolome. (e.g. enviropig, golden rice).

All of these have resource tradeoffs, many of which are completely unstudied.

### TODO:

* full list of n essential+prudent nutrients as something like line segment
  attractors in n-dimensional-behavior-space
* some list of antinutrients/pathogens to build obstructive manifolds in
  n-dimensional-behavior-space
* Failures to systemize requirements correctly, either with a lack
  of sensitivity (e.g. including menatetrenone, docosahexaenoic acid, collagen,
  non-homeopathic doses of Vitamin D) or a lack of specificity (e.g. excluding
  phytate-bound or picolinate-bound minerals, carotinoids, various proline-rich
  peptides, cyanocobalamin,
* examples of antinutrients/pathogen mesa-alignment, in physiological and
  pathological forms.
* think hard about how to manage the anti-inductiveness of
  antinutrients/pathogens in the future

[Evan Hubinger et al arXiv:1906.01820]: https://arxiv.org/abs/1906.01820
[Robert Miles youtube:bJLcIBixGj8]: https://www.youtube.com/watch?v=bJLcIBixGj8&t=4m14s