Analysing goals and potential outcomes
Once projects are started, they can take on a life of their own.
It’s similar to the course taken by the monster created by Dr Frankenstein in Mary Shelley’s ground-breaking novel. A project – especially one with high prestige – can acquire an intrinsic momentum that will carry it forward regardless of obstacles encountered along the way. The project proceeds because people involved in the project:
- Tell themselves that there’s already a commitment to complete the project
- View themselves as being in a winner-takes-all race with competitors
- Feel constrained by a sense of loyalty to the project
- Perceive an obligation to fellow team members, or to bosses, or to others who are assumed to be waiting for the product they are developing
- Fear that their pay will be reduced, and their careers will stall, unless the project is completed
- Desire to ship their product to the world, to show their capabilities.
But the result of this inertia could be outcomes that are, later, bitterly regretted:
- The project produces results significantly different to those initially envisioned
- The project has woeful unexpected side-effects
- Even if it is successful, the project may consume huge amounts of resources that would have been better deployed on other activities.
Accordingly, there’s an imperative to look before you leap – to analyse ahead of time the goals and potential outcomes we can expect from any particular project. And once such a project is underway, that analysis needs to be repeated on a regular basis, taking into account any new findings that have arisen in the meantime. That’s as opposed to applying more funding and other resources regardless.
The bigger the potential leap, the greater the need to look carefully, beforehand, at where the leap might land.
The Singularity Principles address projects that seek to develop or deploy new technology that might, metaphorically, leap over vast chasms. The first six of these principles act together to improve our “look ahead” capability:
- Question desirability
- Clarify externalities
- Require peer reviews
- Involve multiple perspectives
- Analyse the whole system
- Anticipate fat tails
Read on for the details.
The principle of “Question desirability” starts with the recognition that, just because we believe we could develop some technology, and even if we feel some desires to develop that technology, that’s not a sufficient reason for us actually to go ahead and develop it and deploy it.
Therefore, the principle urges that we take the time, at the start of the project, to write down what we assume are the good outcomes we will obtain from the technology to be developed. Once these assumptions have been written down, it allows for a more thoughtful and considered review.
The principle also urges that we consider more than one method for achieving these intended outcomes. We should avoid narrowing our choice, too quickly, to a particular technology that has somehow caught our fancy.
This separation in analysis of desired outcomes, sometimes known as “requirements”, from possible solutions, is a vital step to avoiding unintended consequences of technologies:
- Requirements: The outcomes we desire to obtain, as a result of this project, or possibly from other, different, projects
- Solutions: Potential methods of meeting our requirements – though, if we’re not careful, we can become preoccupied with achieving a particular solution, and lose sight of key aspects of the underlying requirements.
For example, a requirement could be “reduce the likelihood of extreme weather events”. One possible solution is “accelerate the removal of greenhouse gases that have built up in the atmosphere”. But a preoccupation with that solution might lead to experimentation with risky geo-engineering projects, and to a failure to investigate other methods to avoid extreme weather events.
Again, a requirement could be “reduce the threats posed by the spread of weapons of mass destruction”. One possible solution is “accelerate the introduction of global surveillance systems”. But a preoccupation with that solution can have its own drawbacks too.
Once we have documented our requirements, it can make it easier to find better, safer, more reliable ways of achieving the outcomes that we have in mind.
The principle of “Question desirability” also recommends that we should in any case challenge assumptions about which outcomes are desirable, and we should be ready to update these assumptions in the light of improved understanding.
Indeed, we should avoid taking for granted that agreement exists on what will count as a good outcome.
That takes us to the next principle, “Clarify externalities”.
Recall that an externality is an effect of a project, or the effect of an economic transaction, that is wider than the people directly involved.
Examples of negative externalities include noises, smells, pollution, resource depletion, cultural chaos, and a general loss of resilience. Examples of positive externalities include:
- People learning skills as a result of interacting with each other
- A reduction in the likelihood of non-vaccinated people catching an infection (because the prevalence of the infection in the population is reduced by the people who are vaccinated)
- The free distribution of second-hand books and magazines.
The principle of “Clarify externalities” draws attention to possible wider impacts (both positive and negative) from the use of products and methods, beyond those initially deemed central to their operation. The principle seeks to ensure that these externalities are included in cost/benefit calculations.
Therefore we should not just consider metrics such as profit margin, efficiency, time-to-market, and disabling competitors. We need to consider broader measures of human flourishing.
What makes this analysis possible is the effort taken, in line with the “Question desirability” principle, to write down the intended outcomes of the technology to be developed. What makes this analysis more valuable are the principles of “Require peer reviews” and “Involve multiple perspectives” to which we turn next.
Require peer reviews
The alternative to requiring peer reviews is that we trust the people who are behind a particular project. We may feel they have a good track record in creating technologies and products. Or that they have outstanding talent. In that case, we might feel a peer review would be a waste of time.
That may be acceptable for projects that are sufficiently similar to those undertaken in the past. However, new technologies have a habit of bringing surprises, especially when used in novel combinations.
That’s why independent peer reviews should be required, involving external analysts who are not connected with the initial project team. These analysts should ask hard questions about the assumptions made by the project team.
The value of these peer reviews depends on:
- The extent to which reviews are indeed independent, rather than being part of some cosy network of “I’ll scratch your back – give your project a favourable review – if you scratch mine”
- The extent to which reviewers have up-to-date relevant understanding of the kinds of things that could go wrong with particular projects.
In turn, this depends upon society as a whole placing sufficient priority on supporting high quality peer reviews.
Involve multiple perspectives
The peer review phase, into the proposed goals and likely outcomes of a project, should involve people with multiple different skill sets and backgrounds (ethnicities, life histories, etc).
These reviewers should include not just designers, scientists, and engineers, but also people with expertise in law, economics, and human factors.
A preoccupation with a single discipline or a single perspective could result in the project review overlooking important risks or opportunities.
To be clear, these independent analysts won’t necessarily have a veto over decisions taken by the project team. However, what is required is that the project team, along with their sponsors, take proper account of questions and concerns raised by independent analysts.
That proper account should observe two further principles: “Analyse the whole system” and “Anticipate fat tails”.
Analyse the whole system
What’s meant by the “whole system” is the full set of things that are connected to the technology that could be developed and deployed – upstream influences, downstream magnifiers, and processes that run in parallel. It also includes human expectations, human beliefs, and human institutions. It includes aspects of the natural environment that might interact with the technology. And, critically, it includes other technological innovations.
When analysing the potential upsides and downsides of using the new technology that we have in our mind, we need to consider possible parallel changes in that wider “whole system”.
- Rather than just forecasting that a new intervention in a biological ecosystem might reduce the presence of some predator species with unpleasant characteristics, we need to consider whether a reduction of that population would trigger a sudden rise in the population of another species, preyed on by the first, with knock-on consequences for the flora consumed by the second species, and so on
- Rather than extrapolating the level of public interest in a forthcoming new technology from what appears to be only a modest interest at the present time, we should consider the ways in which public interest might significantly change – potentially even causing a panic or stampede – once there are visible examples of the technology changing people’s lives
- Rather than simply analysing how a piece of new artificial intelligence might behave in the environment as it exists today, we should consider possible complications if other pieces of new artificial intelligence, including adversarial technology, or novel forms of hacking, are introduced into the environment as well.
This kind of analysis might lead to the conclusion that a piece of new technology would, after all, be more dangerous to deploy than was first imagined. Or it could lead to us changing aspects of the design of the new technology, so that it would remain beneficial even if these other alterations in the environment took place.
Anticipate fat tails
The principle of “Anticipate fat tails” urges us to remember that not every statistical distribution follows that of the famous Normal curve, also known as the Gaussian bell curve.
For Normal distributions, once we observe the mean of a set of observations, often denoted by the Greek letter mu (μ), and also the standard deviation of these observations, known as sigma (σ), we can be confident that new measurements more than three standard deviations away from the mean will be unlikely. They’ll be seen only around three times in a thousand. And for a new measurement that is more than six standard deviations away from the mean, you would have to wait on average more than one million years, if a new measurement was made every single day.
However, our initial observations of the data might lead us astray. The preconditions for the distribution of results being Normal might not apply. These preconditions require that the outcomes are formed from a large number of individual influences which are independent of each other. When, instead, there are connections between these individual influences, the distribution can change to have what are known as “fat tails”. In such cases, outcomes can arise more often that are at least six sigma away from the previously observed mean – or even twenty sigma away from it – taking everyone by a horrible surprise.
That possibility would change the analysis from “how might we cope with significant harm”, such as a result three sigma away from the mean, to “could we cope with total ruin”, such as a result that is, say, twenty sigma distant.
In practical terms, this means our plans for the future should beware the creation of monocultures that lack sufficient diversity – cultures in which all the variations can move in the same direction at once.
We should also beware the influence of hidden connections, such as the shadow links between multiple different financial institutions that precipitated the shock financial collapse in 2008.
For example, consider a cry of exasperation in August 2007 from David Viniar, the Chief Financial Officer for Goldman Sachs. Viniar was offering his explanation for a dismal reversal of fortune in the bank’s Global Alpha investment fund. This was no ordinary fund: it used what was described as “sophisticated computer models” to identify very small differences in market prices, and to buy or sell securities as a result. The fund had stellar financial results for a number of years, before experiencing a major setback as the global financial crash gathered pace. Viniar’s shocked comment: “We were seeing things that were 25-standard deviation moves, several days in a row”.
Viniar was by no means alone, as a banking executive, at being caught out by the scale of deviations which occurred in the prices of key financial instruments in 2007. John Taylor of Stanford and John Williams of the Federal Reserve Bank of San Francisco have calculated some stunning “before and after” statistics for the so-called “spread” between the overnight interbank lending rate and the London interbank offer rates (Libor). The baseline statistics covered the period from December 2001 to July 2007, that is, the period before the financial crisis. However, the spread on 9th August 2007 exceeded the previous mean by seven standard deviations of the baseline statistics. By 20th March 2008, the spread exceeded the previous mean by sixteen standard deviations.
The takeaway: the mere fact that performance trends seem to be well behaved for a number of years provides no guarantee against sharp ruinous turns of fortune.
Indeed, whenever there are reasons to foresee fat tail outcomes, it means we need to rethink our plans for the new technology. Otherwise, the world might experience a shock outcome from which there is no prospect of any recovery – perhaps for generations, perhaps indefinitely.