# MY REASON FOR DONATING
I have not donated to Evitable before because I failed to pick up on some of the factors needed for (even unconstrained) superintelligence to doom mankind. I have listed them below in case anybody wants to add them to (or emphases them more) in their argument for restricting AI development, so that others don't make the same mistake.

## MORE DOMINANT AI's TEND TO REPLACE LESS DOMINANT ONES (AND HUMANS)

In most Complex universes, (including our own), given the continuance of multiple competing actual (and potential AIs), having at least one non-trivial goal, the Dominance of each AI (or possibly each group of co-ordinating AIs) is approximated by the product of its instrumental variables. ...

AI Dominance = ( Amount of compute available to it
* Its algorithmic efficiency
* The total value of the information it possesses
* The degree of control it has
) / its vulnerability

Where ...
By "AI Dominance", I mean a measure by which higher dominance AI's tend to replace (or expand their influence at the expense of) lower dominance ones.
"Amount of compute" * "algorithmic efficiency" is the AIs intelligence as normally understood.
"Total value of information" may include the value of information on the public Internet, plus the value of any private information the AI has access to.
"Degree of control" includes any direct power, access to any money, as well as influence.
The product of each of the above has to be divided by the vulnerability of the AI, for instance to being turned off.

The AI dominance and each of the parameters above are meant to be instantaneous values and to change with time. The Intelligence, information, control, and/or protection given by the controlling human (if any) should be added to that of the AI itself.

A justification for the inclusion of each of the factors above and for the use of the multiply operator is the intuition that if an AI possesses no information, or has no control over anything, it cannot dominate anything. The same would occur if it is going to be imminently switched off (infinite vulnerability). Contrarily, if some value is large, then each incremental increase will have less effect.

As an ex. engineer, requiring "information" "control" and "vulnerability" as factors makes the doom scenario more creditable. It also suggests that in so far as factors can be traded off against each other, more powerful AIs will tend to do all of "grabbing" information, make "imaginative" uses of tools, and minimise their vulnerability (e.g. by trying to avoid being found out), as reported in the recent Hugging Face attack.

## THE PROBLEM
The same instrumental variables that define AI Dominance for other non-trivial goals also define ability to make profits, ability to borrow (and therefore spend) money, and status (especially perhaps in the US). There is therefore a large incentive for AI users their creators, and the AIs themselves to strive for more Dominance.
Unfortunately for us humans, the more control the AI has over the world, the less control we have over it. This loss of human control then opens up the pandoras box of ills that others.

## THE SOLUTION
Fortunately, to control AI Dominance, it is only necessary to control one of its factors. Amount of compute available does seem to be the easiest.

## REFERENCES
1. Bostrom N. "Superintelligence" p136 "agents will tend to instrumentally value many kinds of information". The book also adds many other instrumental variables which could become important, including "Resource Acquisition, of which "Control" is a generalisation to include such things as money which is "owned" by a human, but controlled by an AI, and psychological influence.