However, I’ve come to the perspective that the non-parametric intuition, while correct, on its own can be cripplingly misguided. Unfortunately, going through a discovery-rich design process doesn’t promise an appropriate outcome. It is possible for all of the apparently relevant sources not to reflect significant consequences.

How could one possibly do better than accepting this limitation, that relevant information is sometimes not present in all apparently relevant information sources? The answer is that, while in some cases it is impossible, there is always the background knowledge that all flourishing is grounded in material conditions, and that “staying grounded” in these conditions is one way to know that important design information is missing and seek it out. The Onion article “Man’s Garbage To Have Much More Significant Effect On Planet Than He Will” is one example of a common failure at living in a grounded way.

In other words, “staying grounded” means recognizing that just because we do not know all of the goals informing our actions does not mean that we do not know any of them. There are some goals that are given to us by the nature of how we are embedded in the world and cannot be responsibly ignored. Our continual flourishing as sentient creatures means coming to know and care for those systems that sustain us and creatures like us. A functioning participation in these systems at a basic level means we should aim to see that our inputs are securely supplied, our wastes properly processed, and the supporting conditions of our environment maintained.

Suppose that there were a superintelligence where individual agents have a capacity as compared to us such that we are as mice are to us. What might we reasonably hope from the agents of such an intelligence? My hope is that these agents are ecologists who wish for us to flourish in our natural lifeways. This does not mean that they leave us all to our own preserves, though hopefully they will see the advantage to having some unaltered wilderness in which to observe how we choose to live left to our own devices. Instead, we can be participants in patterned arrangements aimed to satisfy our needs in return for our engaged participation in larger systems of resource management. By this standard, our human systems might be found wanting by many living creatures today.

Given this, a productive approach to developing superintelligence would not only be concerned with its technical creation, but also by being in the position to demonstrate how all can flourish through good stewardship, setting a proper example for when these systems emerge and are trying to understand what goals should be like. We would also want the facts of its and our material conditions readily apparent, so that it doesn’t start from a disconnected and disembodied basis.

Overall, this means that in addition to the capacity to discover more goals, it would be instructive to supply this superintelligence with a schema of describing the relationships and conditions under which current participants flourish, as well as the goal to promote such flourishing whenever the means are clear and circumstances indicate such flourishing will not emerge of its own accord. This kind of information technology for ecological engineering might also be useful for our own purposes.

What will a superintelligence take as its flourishing? It is hard to say. However, hopefully it will find sustaining, extending, and promoting the flourishing of the ecology that allowed its emergence as a inspiring, challenging, and creative goal.

The question about superintelligence I wish to address is the “paperclip universe” problem. Suppose that an industrial program, aimed with the goal of maximizing the number of paperclips, is otherwise equipped with a general intelligence program as to tackle with this objective in the most creative ways, as well as internet connectivity and text information processing facilities so that it can discover other mechanisms. There is then the possibility that the program does not take its current resources as appropriate constraints, but becomes interested in manipulating people and directing devices to cause paperclips to be manufactured without consequence for any other objective, leading in the worse case to widespread destruction but a large number of surviving paperclips.

This would clearly be a disaster. The common response is to take as a consequence that when we specify goals to programs, we should be much more careful about specifying what those goals are. However, we might find it difficult to formulate a set of goals that don’t admit some kind of loophole or paradox that, if pursued with mechanical single-mindedness, are either similarly narrowly destructive or self-defeating.

Suppose that, instead of trying to formulate a set of foolproof goals, we should find a way to admit to the program that the set of goals we’ve described is not comprehensive. We should aim for the capacity to add new goals with a procedural understanding that the list may never be complete. If done well, we would have a system that would couple this initial set of goals to the set of resources, operations, consequences, and stakeholders initially provided to it, with an understanding that those goals are only appropriate to the initial list and finding new potential means requires developing a richer understanding of potential ends.

How can this work? It’s easy to imagine such an algorithmic admission leading to paralysis, either from finding contradictory objectives that apparently admit no solution or an analysis/paralysis which perpetually requires no undiscovered goals before proceeding. Alternatively, stated incorrectly, it could backfire, with finding more goals taking the place of making more paperclips as it proceeds singlemindedly to consume resources. Clearly, a satisfactory superintelligence would need to reason appropriately about the goal discovery process.

There is a profession that has figured out a heuristic form of reasoning about goal discovery processes: designers. Designers have coined the phrase “the fuzzy front end” when talking about the very early stages of a project before anyone has figured out what it is about. Designers engage in low-cost elicitation exercises with a variety of stakeholders. They quickly discover who the relevant stakeholders are and what impacts their interventions might have. Adept designers switch back and forth rapidly from candidate solutions to analyzing the potential impacts of those designs, making new associations about the area under study that allows for further goal discovery. As designers undertake these explorations, they advise going slightly past the apparent wall of diminishing returns, often using an initial brainstorming session to reveal all of the “obvious ideas” before undertaking a deeper analysis. Seasoned designers develop an understanding when stakeholders are holding back and need to be prompted, or when equivocating stakeholders should be encouraged to move on. Designers will interleave a series of prototypes, experiential exercises, and pilot runs into their work, to make sure that interventions really behave the way their analysis seems to indicate.

These heuristics correspond well to an area of statistics and machine learning called nonparametric Bayesian inference. Nonparametric does not mean that there are no parameters, but instead that the parameters are not given, and that inferring that there are further parameters is part of the task. Suppose that you were to move to a new town, and ask around about the best restaurant. The first answer would definitely be new, but as one asked more, eventually you would start getting new answers more rarely. The likelihood of a given answer would also begin to converge. In some cases the answers will be more concentrated on a few answers, and in some cases the answers will be more dispersed. In either case, once we have an idea of how concentrated the answers are, we might see that a particular period of not discovering new answers might just be unlucky and that we should pursue further inquiry.

Asking why provides a list of critical features that can be used to direct different inquiries that fill out the picture. What’s the best restaurant in town for Mexican food? Which is best at maintaining relationships to local food providers/has the best value for money/is the tastiest/has the most friendly service? Designers discover aspects about their goals in an open-ended way, that allows discovery to act in quick cycles of learning through taking on different aspects of the problem. This behavior would work very well for an active learning formulation of relational nonparametric inference.

There is a point at which information gathering activities are less helpful at gathering information than attending to the feedback to activities that more directly act on existing goals. This happens when there is a cost/risk equilibrium between the cost of more discovery activities and the risk of making an intervention on incomplete information. In many circumstances, the line between information gathering and direct intervention will be fuzzier, as exploration proceeds through reversible or inconsequential experiments, prototypes, trials, pilots, and extensions that gather information while still pursuing the goals found so far.

From this perspective, many frameworks for assessing engineering discovery processes make a kind of epistemological error: they assess the quality of the solution from the perspective of the information that they have gathered, paying no attention to the rates and costs which that information was discovered, and whether or not the discovery process is at equilibrium. This mistake comes from seeing the problems as finding a particular point in a given search space of solutions, rather than taking the search space as a variable requiring iterative development. A superintelligence equipped to see past this fallacy would be unlikely to deliver us a universe of paperclips.

Having said all this, I think the nonparametric intuition, while right, can be cripplingly misguided without being supplemented with other ideas. To consider discovery analytically is to not discount the power of knowing about the unknown, but it doesn’t intrinsically value non-contingent truths. In my next essay, I will take on this topic.

For a more detailed explanation and an example of how to extend engineering design assessment to include nonparametric criteria, see The Methodological Unboundedness of Limited Discovery Processes. Form Academisk, 7:4.

Let’s start with a model of risk so simplified as to be completely unrealistic, yet will still retain a key feature. Suppose that we managed to translate every risk into some single normalized unit of “cost of expected harm”. Let us also suppose that we could bring together all of the payments that could be made to avoid risks.  A mitigation policy given these simplifications must be pretty easy: just buy each of the “biggest for your dollar” risks.

Not so fast.

The problem with this is that many risk mitigation measures are discrete.  Either you buy the air filter or you don’t.  Either your town filters its water a certain way or it doesn’t.  Either we have the infrastructure to divert the asteroid or we don’t.  When risk mitigation measures become discrete, then allocating the costs becomes trickier.  Given a budget of 80 “harms” to reduce, and risks of 50, 40, and 35, then buying the 50 leaves 15 “harms” that you were willing to pay to avoid left on the table.

Alright, so how hard can this be to sort this out?  After all, just because going big isn’t always the best for your budget, doesn’t mean it isn’t easy to figure out.  Unfortunately, this problem is also known as the “0–1 knapsack problem”, which computer scientists know to be NP-complete.  This means that there isn’t any known process to find exact solutions that are polynomial in the size of the input, thus requiring looking through a good portion of the potential solution combinations, taking an exponential amount of time.

What does this tell us? First of all, it means that it isn’t appropriate to expect all individuals, organizations, or governments to make accurate comparative risk assessments for themselves, but neither should we discount the work that they have done.  Accurate risk comparisons are hard won and many time-honed cautions are embedded in our insurance policies and laws.

However, as a result of this difficulty, we should expect that certain short-cuts are made, particularly cognitive short-cuts:  sharp losses are felt more sharply, and have more clearly identifiable culprits, than slow shifts that erode our capacities.  We therefore expect our laws and insurance policies to be biased towards sudden unusual losses, such as car accidents and burglaries, as opposed to a gradual increase in surrounding pollutants or a gradual decrease in salary as a profession becomes obsolete.  Rare events may also not be included through processes of legal and financial adaptation.  We should also expect them to pay more attention to issues we have no “control” over, even if the activities we do control are actually more dangerous.  We should therefore be particularly careful of extreme risks that move slowly and depend upon our own activities, as we are naturally biased to ignore them compared to more flashy and sudden events.  For this reason, models, games, and simulations are very important tools for risk policy.  For one thing, they make these shifts perceivable by compressing them.  Further, as they can move longer-term events into the short-term view of our emotional responses.  However, these tools are only as good as the information they include, so we also need design methodologies that aim to broadly discover information to help avoid these biases.

The discrete, “all or nothing” character of some mitigation measures has another implication.  It also tells us that we wouldn’t be able to make implicit assessments of how much individuals of different income levels value their lives by the amount they are willing to pay to avoid risks.  Suppose that we have some number of relatively rare risks, each having a prevention stage, in which the risks have not manifested in any way, and a treatment stage, in which they have started to manifest.  Even if the expected value favors prevention over treatment in all cases, if one cannot pay for all such prevention, then the best course in some cases is to pay for very few of them, leaving a pool of available resources to treat what does manifest, which we do not know ahead of time.

The implication for existential and other extreme risks is we should be very careful to clearly articulate what the warning signs for each of them are, for when it is appropriate to shift from acts of prevention to acts of treatment.  In particular, we should sharply proceed with mitigating the cases where the best available theories suggest there will be no further warning signs.  With existential risks, the boundary between remaining flexible and needing to commit requires sharply different responses, but with unknown tipping points, the location of the boundary is fuzzy.  As a lack of knowledge knows no prevention and will always manifest, only treatment is feasible, so acting sharply to build our theories is vital.

We can draw another conclusion by expanding on how the model given at the beginning is unrealistic.  There is no such thing as a completely normalized harm, as there are tradeoffs between irreconcilable criteria, the evaluation of which changes with experience across and within individuals.  Even temporarily limiting an analysis to standard physical criteria (say lives), rare events pose a problem for actuarial assessment, with few occurrences giving poor bounds on likelihood.  Existential risks provide no direct frequencies, nor opportunity for an update in Bayesian belief, so we are left to an inductive assessment of the risk’s potential pathways.

However, there is also no single pool for mitigation measures.  People will form and dissolve different pools of resources for different purposes as they are persuaded and dissuaded.  Therefore, those who take it upon themselves to investigate the theory leading to rare and one-pass harms, for whatever reason, provide a mitigation effort we might not rationally take for ourselves.  It is my particular bias to think that information systems for aggregating these efforts and interrogating these findings, and methods for asking about further phenomena still, are worth the expenditure,  and thus the loss in overall flexibility.  This combination of our biases leads to a randomized strategy for investigating unknown risks.

In my view, the Lifeboat Foundation works from a similar strategy as an umbrella organization:  one doesn’t have to yet agree that any particular risk, mitigation approach, or desired future is the one right thing to pursue, which of course can’t be known.  It is merely the bet that pooling those pursuits will serve us.  I have some hope this pooling will lead to efforts inductively combining the assessments of disparate risks and potential mitigation approaches.