1
First principle of action: separate every act from sensation of the desired effect. No presumption of success. A belief about the effect of an act is only a hope. An act is not its effects. Action is opaque. Effect is endless and uncertain.
2
Means: a belief that a mind applies to reach goals. Examples: a motor neuron to activate, code to execute, not the believed conditions of an act.
3
Separate belief in a means from each belief about its possible effects. The association between an act and an effect is never certain. A mind must be free to individually sense and forget such beliefs. A believed effect of an act, as a prediction, can be unconditioned or can be inferred from other beliefs.
Strictly, certain pairs of act and effect are improbable to us, thanks to our high minds, long experience and complex models. At bottom, a mind cannot make subtle distinctions about probability. It must allow any binding of effect to act to be made or doubted.
We observe only coincidence, not cause. Our models of the world might, in some cases, be very reliable, yet we still routinely find errors in them. Make the simple honest base assumption that we can never know for certain where the errors in our models are, so anything can cause anything, with whatever links between.
4
Why separate action from sensation? A simple act could itself cause a belief—bad engineering. You may have already reached the goal for which you would act. Separate senses from means to save acts and free the mind to discover other means.
5
Means to senses. The senses that independently detect the effects of acts are ideally the effects of earlier acts. A sense should have no special status to a mind. A sense is merely a believed condition of the perception of the expected effect of an act. Senses are only effects of means. A mind could easily miss that a thing is a sense, that it is a condition of belief in a class of beliefs. Senses are discovered. What use to interpret a thing as a sense?
6
You never untie the same knot twice. Ignoring minds with few and discrete senses. All acts are creative because every moment is unique. Even selecting what to ignore, to make moments compare and the past apply, is a creative choice.
7
Thoughts are not distinguished from matter by lack of effect. Thoughts, like every thing, are part of the endless web of effects, whether you see the links or not. Thinking alone affects neurons, oxygen consumption, or CPU heat. Mere arithmetic can destroy a poorly designed computer by overheating it.
8
Mental acts. Not all acts are physical. Adding two numbers in your head is an act.
9
The most general means are the most interesting. Example: one way to effect anything is to ask another mind to do it. This means is tricky to sequence. A mind that resorts to it too early will annoy you with requests. A mind that tries it too late wastes time trying to succeed alone.
10
Imperative and functional computer languages prejudice one effect of every statement: the return value. Example: 4 as the value of 2 + 2. Pure functional languages outlaw any other explicit effect. Wise to favor purity but functional programming purifies a misconception. Instead of pretending that every statement shall only have one known effect, it should admit that every statement has infinite effects, far beyond those intended by the programmer. Only teleological (goal oriented) programming recognizes this philosophical truth.
Imperative:
Try to move forward one step.
Try to turn right.
Try to move forward one step.
Teleological:
You want to be at 1, 1.
You have a sense of position, now 0, 0.
You have a sense of orientation, now North.
You can try to turn.
You can try to move forward.
Shallow differences of syntax and punctuation—little more separate the countless weak imperative languages that programmers debate and rank.
11
Every mind's situation is that under certain conditions, certain acts tend to precede approximate effects. Ultimately, we do not know why, though we can indefinitely elaborate our models in whatever direction seems most promising.
12
At any time any act could precede any effect. No mind can ever exhaust the possibilities for action. It can only discover, plan, and rank them.
13
Means to means. A mind could originally believe in only one means with which the mind would make the next layer of means. A wise mind in a powerful body could begin with only a bare mind believing in a single means.
14
Passive mind: a mind without means, or the ends to give them purpose, or both. Strictly, not a mind at all. Any use? Maybe to isolate learning from an active mind. A passive mind that learns associations could have ends to focus its attention. A mind with means but no ends could act randomly to build a general purpose model of its world.
A passive mind would not even apply itself to making senses or communicating what it learned. This mute mind must be transparent to whatever uses it.
15
Throttling. No act's intended effect is immediate. How long should a mind wait? Not as long as it takes. It may take forever. Yet any length is more intelligent than that or zero. For one act, the gap is a millisecond, for another, an hour, in both cases, never precisely the same again. How to learn the lengths?
At least, as a moving average of all a mean's acts. Better, inferred from a single act's purpose and preconditions. Best, inferred from any relevant belief. Anticipating an act's delay could be another act with a prediction as its effect. If the mind waits too long, it wastes time waiting on a hopeless act. If the mind retries too soon, it may overload the means.
The time when an acts starts vs. the time when the act itself ends vs. the time when the intended effect is sensed.
Continuous vs. instantaneous acts. Acts, like a thermostat keeping the furnace on, that have an increasing effect vs. acts that quickly end and have a fixed effect.
How can a mind enforce a throttle? At worst, entirely in its unconscious, in its engine. Better done, as usual, in terms of beliefs. An engine could sense its own use of a means. Then the mind could infer from such beliefs a temporary suppression of belief in any matching means.
16
Act sequencing. How to know what act to try next? One rule: try specific means before general ones. When to start trying another means? Or the same means in a different way? How best to interleave retrying multiple means? Not all minds have such a sequencer. Example: a thermostat has only one means to its end.
Any answer, no matter how often wrong, is an immense leap over a system so simple that it doesn't need an answer, that never retries or varies an act.
17
Parametric means: a means to more than one intended effect. How does a mind apply a means that can have different effects? How do the particulars of the goal connect to the act?
18
Basic problems of action.
Broken actuator: How to recognize that a means no longer has the expected effect? Maybe it only fails to have certain effects under certain conditions. How to reset a frozen actuator, especially in a sequential mind medium?
Delayed effect: The time between the act and the intended effect exceeds the mind's expectation.
Overshoot, oscillation: An act may have more than the desired effect.
19
Don't die. Any act can have any effect at any time. This includes the mind's death. A mind inevitably uses means in conditions unseen by its maker. An active mind must barricade itself from the dangerous side-effects of every act. Examples of software errors that are fatal if uncaught or unstopped: an exception that would end the thread, a process that exhausts the CPU or memory.
Engineers typically limit their machines to make errors unlikely. A non-trivial mind's use is to act creatively. A mind maker can't contrive a safe path for a mind meant to find new paths. Don't avoid errors. Attack error itself.
Making a mind robust involves challenges caused by the nature of mind. Others by the technicalities of a mind's medium, which I leave to the expertise of different readers.
20
Parallelism as a prerequisite of high speed. To keep pace with other minds, a mind must commit acts, including acts of thought, in parallel, not one after the other.
Intelligence as a prerequisite of parallelism. Mind isn't cheap. Compared to blind action, the overhead of sensation, inference and selection is immense. But the awareness of the conditions of action and their exclusivity can repay the investment.
Every act has conditions of yielding particular effects. A mind can safely commit two acts in parallel only if none of their conditions, and the conditions of the conditions, are exclusive. So only a mind that learns negative associations can discover what acts it can commit at once. Present computer programs depend on the mind of a human programmer to see the independence of acts.
Noticing the relationships between conditions exposes another means for a mind to conceal its overhead: first pursue conditions shared by multiple acts.
21
Telling of errors in means as a means. Preconditions: senses of errors and language to discuss them.
Scenario: a mindless machine fails.
You cause the machine to attempt x.
x fails and the machine at best manages to show an error message.
The message gives too little information. The machine offers no way to ask for more. You guess.
You again tell the machine to do x.
Contrast: an intelligent machine interacts.
You give the machine a goal to x.
It repeatedly tries to cause x using means a.
While retrying a, the machine tries means b, which happens to involve telling you of errors that it associated with a.
You ask for a detail.
You correct whatever caused a to fail.
The machine successfully retries a before you can tell it of the change.
22
Act to know the world vs. act to change the world. How and why to distinguish between an act to cause a sense, which will change the acting mind's beliefs, and an act to change the world, which also changes a mind's beliefs?
23
Parts of an act:
Expected effect.
Means: how a mind initiates effects.
Conditions: what must be so for the means to have the intended effect. Kinds:
Preconditions: what must be when a means is applied.
Co-conditions: what must be for the intended effect to occur and persist.
Comments