Richard Hollerith

When I learned about goal system zero back in 1992 I saw it (among other things) as a way to get clarity on the moral environment, e.g., on the many moral claims that bombard me. I told myself that goal system zero does not have to be completely correct for it to help me to evaluate moral claims and to organize my actions. I explicitly compared it to Newtonian dynamics — another system that has helped people organize their actions in spite of not being completely correct.

Since I started posting on Overcoming Bias in Oct 2007 I have defined my purpose as advocating goal system zero to the singularitarians. Well, I am no longer going to advocate goal system zero — at least not to the singularitarians — for the following reasons:

• advocacy of goal system zero scares singularitarians, and a scared singularitarian does not think as effectively as a not-scared singularitarian
• advocacy of goal system zero scares singularitarians, including singularitarians I would like to have as friends
• it will be good for my psychological health for me to act more from my natural human feelings and less from an intellectual ethical structure.

Although I will no longer advocate goal system zero, I will continue to share what I know about goal system zero. I have answered questions about goal system zero from over a dozen people. I welcome more discussions of that sort. I would be particularly inclined to be helpful toward anyone trying for a mathematical definition of goal system zero.

Whereas my old role in the rationalist/singularitarian community was an advocate, my new role is probably best described by the concept or metaphor of a churchgoer. In other words, I will participate on Less Wrong and attend meetups for essentially the same reasons that many attend church. For more about this metaphor of the churchgoer, see the page that says what the blog is about, which I just rewrote.

(more…)

This is a comment on Eliezer’s series (from December?) on personhood, sentience and consciousness and their relation to morality. My comment relies on a knowledge of the cognitive skill sometimes known as mind reading or empathy. Mirror neurons are part of the implementation of that skill. By waiting till today to comment, I get to piggyback on Eliezer’s post Sympathetic Minds in which those things are explained.

In the series on personhood, Eliezer says that it is important when designing a superintelligence not to create a person as a side effect. For example, if the superintelligence has to predict the behavior of a person, Eliezer says that it is important that in computing the prediction, the AI does not simulate the person with such accuracy that the simulation itself constitutes a person.

My position is that the designer of a superintelligence does not have to worry about creating a new person. Contrary to what most people think, the whole question of whether an intelligent agent or a part of an intelligent agent is a person or whether it is conscious is irrelevant in most situations in which one is considering the far future or considering the use of extremely potent technological means like a superintelligent AI.

Any important decision is impinged on by thousands of factors, but turns on much fewer. If we build a superintelligence, we have to decide what goal to give it (including what ethical standards are always observed in the pursuit of the goal). The probability that the concept of personhood or consciousness is useful or important in making that decision is IMHO so low as to be not worth thinking about.

(more…)

This is the next step of a dialog that starts here and continues with Virge’s asking me,

Richard, do you assign zero value to your autonomy? Do you also assign zero value to your personal enjoyment of the process of achieving your goals?

(more…)

Given a few simple conditions we will enumerate in a minute, what an agent would do seems to be independent of whether the agent’s goal is to maximize, e.g., the number of moments of happiness experienced by humans so far or, e.g., the number or gold atoms “at the current time”. (That last needs to be made more precise because of course there is no such thing as the current time in relativity theory. Nevertheless, I will continue to use the phrase “at the current time” because additional precision is not necessary to get the basic idea across.) In other words, there seems to be a wide range of goals that has the property that pursuing a goal in the wide range entails doing the same things initially as pursuing any other goal in the wide range. And this initial period during which what is done is invariant might easily last billions of years.

The goal system that has my tentative loyalty, which I call goal system zero (GSZ), does whatever things an agent loyal to a goal in this wide range of goals would do, but it does them as an end in themselves, not as a means to some other end.

Actually we have to add a few conditions to that last statement. One condition is that the goal in the wide range of goals must be time-indifferent: if for example the goal is to maximize the number of gold atoms, there may not be a time limit on when the maximization happens or even a future discount rate applied to the number of gold atoms. Actually, the precise definition of GSZ refers to effects rather than instances in time. I prefer sometimes to use words like “time” and “before” because they are easier for the reader to understand, but for deep understanding, those words should be translated to words like “cause” and “effect”. So, a more precise definition is as follows: a goal system is time-indifferent iff the chain of effects leading to the state in which utility is at a maximum can be as long as you like.

In addition to being time-indifferent, GSZ is “agent-indifferent”: it does not matter to GSZ which agent causes reality to end up in a state considered desirable by GSZ. Note that most human goals are not agent-indifferent. For example, when a scientist tries to discover a new law of nature, he probably prefers to make the discovery himself (so that he gets credit for it). He probably prefers that outcome to an outcome in which he helped another scientist to make the discovery even if the latter outcome is easier for the scientist to achieve.

To summarize, what has utility under GSZ is the capacity to achieve some wide range of goals, and this wide range includes goals such as maximizing the number or gold atoms.

Parenthetically, my opinion as to the best way to increase this capacity to achieve some wide range of goals mirrors Eliezer’s opinion on how to build a superintelligence on all points I am aware of: self-improving AI, decision theory, causal models a la Pearl, Solomonoff induction, Kolmogorov complexity, reflectivity. Of course Eliezer’s opinion is better grounded than my opinion because learning that stuff has been Eliezer’s day job for about ten years.

Define creativity as the ability to achieve a goal in the wide range of goals we have been speaking of. I wonder whether the goal of maximizing creativity itself is a member of this wide range of goals.

If it is a member, is that by the definition of creativity or does it rely on the nature of the reality in which we find ourselves? In other words, if we find ourself in a different reality in which different laws, might that change the answer?

And if it is a member, then is it not the case that it does not make sense to speak of making a choice that reduces creativity in the short term in exchange for increasing creativity in the long term?

Certainly it is possible to trade off a short term reduction in some other expected measure of utility such as moments of happiness or atoms of gold for a long term increase in that expected measure. But that does not mean the same is true of creativity.

Certainly it is possible for an agent loyal to GSZ whose model of reality is incomplete (like the model of any agent in our reality must be unless we are missing some very vital fact about our reality) to mistakenly choose an action that increases creativity less than an alternative action would have. Here is a simple example. You are the emperor of ancient Rome and you try to increase creativity by spending public monies to provide running water to the common citizen. But you use lead pipes to provide the water, which damages the brain of the common citizen. (The ancient Romans did not know lead was a poison.) So the agent ends up actually decreasing creativity. But we can fix up our question by replacing the word “creativity” with “expected creativity”. The adjective “expected” here is meant to indicate that the thing is to be evaluated relative to the agent’s current model of reality rather than relative to reality itself. So, now our question reads,

And if it is a member, then is it not the case that it does not make sense to speak of making a choice that reduces expected creativity in the short term in exchange for increasing expected creativity in the long term?

I’ve said this before, but it bears repeating. The reason I am in this public conversation is that I have a perspective — a system of values — that I believe deserves to be represented or advocated, and I see no one else doing so. Unfortunately I am not a very good advocate because my life is a mess, with the result that I cannot afford to give the role of advocate the time it deserves. This summer and fall is a particularly messy time in my life.

I have indicated in the past (e.g., here) that indefinitely long causal chains are central to the system of values I advocate. I understand now that although those words (”indefinitely long causal chain”) are sufficient to remind me of certain concepts and arguments, they are woefully inadequate to the task of informing or persuading someone who is not already informed and persuaded.

When I speak of a causal chain, I have in mind a directed path through a directed acyclic graph where each node of the graph represents an event and where there is an edge from node A to node B iff event A is a cause of event B. Graphs like that are known as Bayesian networks, and to a first-order approximation, a Bayesian network is the kind of thing an AGI or an SI will use for its model of reality. Probably the best author on causal models is Judea Pearl.

I wish I had time to present at this point an example from the beginning of one of Pearl’s books of a very simple Bayesian network. But I do not have time.

The purpose of a system of values or a model of reality is to help the agent make choices; if the agent never makes a choice there is no need for the agent to have either of these things.

Consider a statement I made in the past, namely, that causal chains that do not persist indefinitely are unimportant. Let us try to unpack that statement a little.

It is not really causal chains, but rather choices between alternative actions that are either important or unimportant. Each alternative action will initiate a causal chain. In other words, the action will cause an event, which will in turn cause another event, and so on.

Actually, the action will cause zero or more events, and each of those will cause zero or more additional events. The fiction that each events has exactly one effect is a fiction I have been using to save words.

I will sometimes refer to such the chain of effects of an action as a “possible future”.

An example is in order. Consider Conway’s Game of Life and consider an agent that has some ability to affect the evolution of the game. Perhaps the agent can choose the starting configuration. Perhaps the agent can choose the starting state of one of the squares.

This is an example of what we will call a predicament. The predicament currently under consideration consists of

• A board or grid of squares. Every square is “mutable,” which means it can change over time.
• A mapping that takes a grid state into another, successor grid state. This mapping can be viewed as the “rules” of Conway’s game.
• An agent able to make a choice that affects the outcome of the game (by for example affecting the state of one of the squares of the grid).

The biggest difference between the predicament currently under consideration and the actual predicament in which we find ourselves is that we (or rather human civilization in general) must discover the laws of our reality whereas we will postulate that the agent just introduced already knows the “laws” or rules of Conway’s game. There is in other words no problem of induction in the predicament under consideration.

Another big difference is that the means (e.g., our brains, our computers and things like pencil, paper, desks and friends) by which we learn about and keep track of our reality are an integral part of our reality whereas we will postulate that the agent under consideration is external to the board or grid of Conway’s game. In other words, the same laws of physics that apply to everything else in our reality apply to our selves and our minds. The agent under consideration must be able to make a choice, but let us suppose that the agent has no knowledge — no causal model — of how it is able to choose — or indeed how it is able to know anything (which is not too far from the predicament in which the ancient Greeks found themselves).

Now I want the reader to assume that the predicament under consideration, which again consists of an agent, a grid and a mapping, is the whole of reality. It is a good exercise for us to imagine finding ourselves in a reality much different from the reality we actually find ourselves in and to try to get a feel for the “moral dynamics” (for lack of a better term) of that alien reality.

Suppose the agent is able to make a binary choice — is able, that is, to choose between action A1 and action A2.

Suppose further that A1 causes the grid eventually to reach state S1 and that A2, too, causes the grid eventually to reach state S1 although the sequence of grid states between A1 and S1 might differ from that between A2 and S1.

Since Conway’s game has the property that the state of the grid at any point completely determines the evolution of the game forever after, the evolution of the game after S1 is the same regardless of how S1 was reached.

Now I will state a normative belief of mine about the simple predicament consisting of an agent, a grid and a mapping. Provided that the predicament constitutes the entirety of reality, the choice between A1 and A2 is trivial or unimportant. Why? Because the predicament (which we are assuming is reality) will end up the same way regardless of which alternative is chosen.

The above is a restatement with a little more detail of my assertion that only causal chains that persist indefinitely are important.

The original assertion does not make a whole lot of sense when taken literally because regardless of the starting configuration, every run of Conway’s game persists indefinitely: the game can go on for as many moves as one cares to make. The game might be very boring (e.g., might consist forever of only blank squares) but even a very boring game is a causal chain of indefinite length.

So, taken literally, every causal chain is important in this predicament — this alien reality — we are considering. Or so it seems to me right now. But not every choice is important.

In summary, a choice between two actions, both of which cause reality to end up “the same way”, is IMHO an unimportant choice.

One principle or intuition that leads me to believe that is that the proper way to determine the desirability or the rightness of an event is to examine the effects of the event. If we are applying the principle or intuition uniformly, we are naturally led to an examination of the effects of the effects. And so on indefinitely.

There are other arguments for the conclusion, some of which I mention in the post of a few days ago titled Goal System Zero.

Supermassive black holes are the kind of thing that persist a very long time. (The more massive the hole, the slower it loses mass due to Hawking radiation, so supermassive holes have a very, very long life according to current physical models.)

Is a universe consisting two supermassive black holes orbiting around each other better than a universe consisting of ten smaller pairs of black holes of the same total mass? The principle I am explaining today does not say. Yes, the smaller black holes will tend to evaporate sooner, but even if they evaporate, the universe goes on (and the black holes and the Hawking radiation caused by the black holes continue to have effects that might persist forever).

The principle I am explaining today does says that black holes are the kind of thing that might be important in deciding whether possible future A is better or worse than possible future B.

In contrast, moments of happiness are not the kind of thing that might be important in deciding whether possible future A is better or worse than possible future B.

Consider a reality much like our own except that the reality contains an agent that happens to know that supermassive black holes are approaching Earth from all directions. When the holes reach Earth, they will collide to form an even more massive hole, and the solar system will be swallowed up by the resulting hole, and there is nothing the agent can do to prevent that from happening, and there is no time to evacuate the humans or save any product of the humans.

Note that most physicists with an opinion believe that when something is swallowed by a black hole, any information in that thing (e.g., what kind of particles it is composed of, how those particles are arranged) is destroyed except that the mass of the thing (and perhaps one or two other properties like charge and angular momentum) is added to the mass of the black hole.

Suppose further that the agent is faced with a choice between two actions H and S. Action H (”H” for “Happy”) leads to a flourishing of human civilization in which death, disability and suffering are eradicated and the humans come to enjoy exquisite control and choice over their lives, their selves and their experiences. Then a thousand years later, the black holes arrive and everything is swallowed up.

Action S (”S” for “Sad” or “Sadistic”) leads to lots of degradation, death and torture. Then the black holes arrive and everything is swallowed up.

The same reasoning we used in the predicament of the Conway game applies to this predicament: regardless of which action is chosen, after a thousand years or so, reality is “the same way”. Consequently, the choice between H and S is unimportant! In other words, it is not the kind of thing to which an agent might need to pay attention to when making a choice about something important. And it will never be necessary to consider the choice when deciding between goodness and badness or rightness and wrongness.

Now you can see why I feel the need to participate in this public conversation. The vast majority of participants in the conversation do not agree with the conclusion just made. My viewpoint is in a small minority. Yet I am almost sure I am right, which is why I think it is important for me to participate in the public conversation about seed AI, “Friendly” AI, and so on.

If there is even a tiny probability that a product of human civilization (a colony of humans, say, or a robotic spacecraft or even a message) can escape the collision, then the conclusion that the choice between H and S is unimportant no longer holds because that effect of human civilization which escapes the collision can exert a decisive effect on the evolution of reality (especially if humans are the only intelligent agents in reality.

Let us return to the question, Is a universe consisting two supermassive black holes orbiting around each other better than a universe consisting of ten smaller pairs of black holes of the same total mass?

There is a way to begin to crack that question that I have not mentioned yet because I did not want to distract from my main point. But now that I have made my main point, it probably will not hurt anything to mention it.

The maximally entropic state of a space-time continuum like ours is (according to Roger Penrose) a universe consisting of a single black hole containing the vast majority of the mass of the universe and a very thin gas of baryons and photons.

Like I indicated in the post titled Goal System Zero, one principle a person can properly use to decide between two actions is to choose the action that (roughly speaking) leaves the agents affected by the action with the most options. In the Goal System Zero post, I imply that all other things being equal, that action is the one that maximizes the “creativity” (intelligence and knowledge, roughly) of the agents.

Well, it seems to me that increasing the entropy of reality necessarily closes off options. That is the main reason that the most evil action I can imagine a person doing is to launch the seed of a superintelligence whose goal is to pile as much matter as possible into black holes as massive as possible. In contrast, if the future light cone is filled with paperclips, there is at least some possibility that that vast region of paperclips will evolve somehow into something interesting.

So, there is a rationale for preferring the action that results in the ten pairs of black holes over the action that results in a single pair of black holes. The latter has more entropy and consequently offers any agents in that possible future fewer options.

The agents have more opportunities, for example, to create free energy by colliding two of the holes together.

But that is almost a distraction from the main point, which is that the principle that causal chains that do not persist are unimportant does not by itself help us decide whether one indefinitely-long causal chain is better than another.

Very parenthetically, there is a significant probability (according to professional cosmologists) that it will be impossible for black holes or intelligent agents or even quarks to persist indefinitely in our space-time continuum because dark energy or whatever you call it will become so dense that it will tear all these things apart. (That possible future often goes under the name of the Big Rip.) If that is the case, then the only hope for you or I to initiate an indefinitely-long causal chain entails the discovery of a second “compartment” of reality with which agents in this, first compartment can communicate. I hate to even bring this up because it (unjustifiably) will cause a lot of readers to conclude I am a crank, but it might help one of my more thoughtful readers by reminding him that the models of reality possessed by the current generation of physicists and cosmologists might leave out some very important things. In particular, if we have a more accurate or more complete model of reality, the “moral landscape” painted by this blog entry might not look so sterile now.

I have a comment on today’s post by Eliezer over at Overcoming Bias:

(more…)

Now let us turn to Marcello’s objection that an agent with Roko’s goal system will try to destroy rival agents. My answer is that it seems to me that most agents will try to destroy certain classes of rival agents. There are reasons to think that Goal System Zero (GSZ) is more cooperative than most goal systems are.

If a person wanted to, he could probably specify a goal system that will try to help all agents it encounters, resolving all conflicts between the helped agents as peacefully and as fairly as possible. Such a goal system would be more cooperative than GSZ, but although it is a little reassuring to me that GSZ is cooperative, neither universal helpfulness nor fairness nor conflict minimization are terminal values of mine.

Let us examine what happens when an agent loyal to GSZ meets up with an agent loyal to the goal system of maximizing human happiness.

First, note that most people alive today who advocate the terminal value of human happiness have a strong preference for moments of human happiness that will occur in the near future. If you explain to them that what maximizes human happiness is for the humans alive now to spend all their time and energy designing and building superintelligences (SIs) to expand through the universe to turn the resources of the universe into more SIs so that billions of years from now all that engineered intelligence can be applied to creating humans and keeping them happy, well, they do not want to hear that: they want humans alive now to be happy; they do not want human alive now to devote themselves to a project that take billions of years to bear fruit. In other words, most people with goal systems that emphasize human happiness have goal systems that assign greater utility or moral value to moments in time close to here and now. Let us call such a goal system “non-time-indifferent”.

Let us suppose a GSZ agent meets up with an agent with the time-indifferent goal of maximizing human happiness. To the first order, the GSZ agent is perfectly happy to see the happiness maximizer dominate the resources of the part of the future shared by both agents — at least if the happiness maximizer is agent-indifferent as well as time-indifferent. Again, GSZ has no preference as to what eventual end the creativity of the universe is applied to: human happiness is just as good as anything else as far as GSZ is concerned. It would however disapprove of the premature pursuit of happiness.

I included the disclaimer “to the first order” because the GSZ agent has to consider what happens if the other agent meets a third agent intent on maximizing, e.g., gold atoms. The second and third agents might fight, and of course a fight has the potential to expend resources that could have gone into maximizing creativity. It is in the interest of GSZ to try to prevent the fight. Yes, one way to do that is for the GSZ agent to destroy the second agent — or to deny it resources.

But life as an SI loyal to an agent-indifferent goal system differs significantly from life as a human. For one thing, any SI Eliezer or I would endorse would know how to create SIs with any goal system the SI cares to specify. (This is a side effect of its knowing how to improve itself). This gives two SIs a powerful method to reach a compromise that is unavailable to humans: in this method, the SIs agree to destroy themselves while simultaneously creating a new SI whose goal system is a composite of the goal systems of the “parent” SIs. (E.g., the utility function of the child SI is the sum of the utility functions of the parent SIs. The child SI can be used to verify the destruction of the parent SIs.)

The availability of this high-reliability method for compromise suggest that if two SIs with conflicting agent-indifferent goal systems meet, they will expend very little of their resources fighting. So, GSZ’s motivation to destroy or hobble another agent-indifferent goal system to prevent it from fighting a third agent-indifferent goal system is likely to be weak. It seems to me that GSZ is more likely to provide positive assistance to an agent-indifferent time-indifferent goal system than to hobble or destroy it.

Note that when the goal systems under discussion are agent-indifferent, it seems harmless to blur the distinction between a goal system and an agent loyal to the goal system, which is what we have done in the previous couple of sentences.

The helpfulness that GSZ is expected to extend to an agent loyal to an agent-indifferent time-indifferent goal system does not extend to an agent with an goal system that lacks one or both of those two properties. GSZ is likely to consider these other agents to be rivals, but as far as I can tell, GSZ is not different than most goal systems in that regard — the exception being goal systems deliberately designed for cooperativeness or helpfulness.

This blog entry and the following one are replies to Eliezer’s Invisible Frameworks. I will continue to describe the system of values that currently has my loyalty, a system I call goal system zero. My initial description of goal system zero is here.

Obviously an agent that tries to maximize the number of moments of happiness in the world is a rival of an agent that tries to maximize the number of, e.g., gold atoms.

Consider an agent that tries to maximize the universe’s ability to get things done, but that has no preference as to what eventual end to which that ability will be put. My reason for prefering that agent to, e.g., one that maximizes moments of human happiness is that I perceive no valid reason to prefer happiness moments and gold atoms. (Yes, I concede that I have an austere and dry way of looking at the world.) Consequently, the superintelligence I would set into existence if I had to power to set one into existence would have no preference in that regard.
Although it pleases me that that agent would be more cooperative than a agent set on maximizing the number of gold atoms, that is not my main reason for prefering it.

Note that I say, “maximize the universe’s ability,” not “maximize the agent’s own ability.” That is because the agent I am defining differs from most humans in that its goal system is “agent-indifferent”. An “agent-indifferent” goal system does not refer to itself or indeed to any agent. It is perfectly happy to choose an action that causes itself to cease to exist as long as the same action causes the creation of another agent that will prove at least as instrumental to its goal system.

Teilhard de Chardin is probably the first author to describe in detail an agent-indifferent goal similar to the ones under consideration here. He spoke of the universe becoming aware of itself and of its own history. He held that out as a new motivating principle for our human civilization. I have not read much Teilhard, so I do not know if Teilhard also held out a broader motivating principle, namely, the ability of the universe to transform itself and steer its own future, but John David Garcia has, and John describes himself as an intellectual heir to Teilhard. Obviously, becoming aware of oneself and one’s history is necessary to gain control over oneself and one’s future, so the one goal is a subgoal of the other.

To avoid repeating the phrase “ability to get things done,” I will follow Garcia and say, “creativity.”

It is possible to implement a superintelligence (SI) that seeks to maximize the creativity of the parts of reality under its control, but that has no preference as to what purpose that creativity is eventually put. At first that seems self-contradictory, but it is not, as I try to explain here. The trick is that maximizing creativity is an unambiguous guide to action only if the reality in which the agent finds itself is “big” (which I define in the linked document). The goal system I want the superintelligence (SI) to have has no preference among futures if the SI finds itself in a “small” reality: in that situation, it just lets any other agents in the vicinity determine the evolution of that reality.

This is not a lost purpose if the agent or the creator of the agent never had any purpose beyond maximizing creativity to begin with. There is nothing impossible or self-contradictory about the goal system, which I call goal system zero (GSZ).

GSZ is not going to motivate any agent to drive around in a car with no destination because that would waste energy that can be used to try to increase the creativity of the universe (or the parts of the universe over which the agent has influence).

I have not questioned Roko enough to know whether he endorses GSZ, but his values seem close, and his argument for often instrumental values is a strong argument for GSZ. Here is a quote from Roko’s first full explanation of his argument:

Any agent who acts in the world to achieve certain goals has to contend with two fundamental facts about the nature of the interaction of an agent with the real world. The first fact is that my desire to achieve some goal (”I want to be in Los Angeles”) does not make that desired state happen. In order to impose our goals on the world, we have to manipulate the world, and those manipulations follow a set of rules, called the laws of physics.

This seems trivial, but as far as finding an objective system of ethics is concerned it is very important, and in fact it is a good thing. If it were the case that as soon as I desired some state, the world instantly transformed itself into that state with no side-effects, then there would be no mathematical structure to the set of goal states that an agent could have. In the case of a set of possible goal states with no mathematical structure, i.e. such that there are no objective relations between those goals, there is clearly no objectively best goal. Like elements of an abstract set, goals without relations between them cannot be superior to one another.

But our world is not like this! Goals do have relations between them. Steve Omohundro wrote two papers about the relations between various goals that an agent can have.

The most important relation that goals can have is the following: Goal A is instrumental to Goal B. That is to say, if we first achieve Goal A, then it will be easier to achieve Goal B.

This is not the only consideration that makes me prefer GSZ. I am motivated also by the fact that the goal system is agent-indifferent (does not refer to particular agents or classes of agents), time-indifferent (does not hold some moments of time to be more important or valuable than other moments) and very simple.