One of the core principles of Wicker is that its design must conform to individual ideas. Its atom of content must be the proposition. However, if propositions are to be presented for discussion on their own pages, the number of possible ideas that could be expressed presents a combinatorial problem. Consider the following proposition.
Proposition
Universal basic income is unlikely to be implemented in the United States before 2028.
It looks alright — fairly clear and easy to understand — but consider its permutations. You could substitute:
- Other likelihoods (e.g., “…likely…”, “…very unlikely…“)
- Other times of varying specificities (e.g., “…before 2029”, “…after March 6”)
- Other locations of varying geographical scale (e.g., “…Canada…”, “…California…”, “…Denver…“) Even with, say, 5 discrete likelihoods to choose from, dates no more granular than months and confined to the next 10 years, and locations no more granular than states and confined to the US and Canada, number of possible valid propositions is already .
With the additional permutations implied by ambiguity in the definitions of “universal basic income” and the word “implemented”, we can see that accepting every semantically unique proposition as its own page with its own user agreement rating, list of sources, etc., would be difficult. In general, this difficulty probably arises as propositions become more specific; the more variables or context qualifiers you add, such as date and location, the more permutations you imply.
The shape of the problem
This problem arises from the interplay between continuous variables and discrete claims. Each claim is like a large, chunky pixel that samples from a smooth underlying landscape of possible claims. The smoothness of that landscape means that an unlimited number of nearly identical yet technically different claims could be sampled over even a small area.
Why is the landscape smooth?
Let us refer to the dimensions of a claim that are “smooth” as variables. Time is a variable since, in the real world, time is a smooth continuum. Though we generally discretize time into seconds, minutes, and hours, there is, in principle, nothing stopping any claim that refers to a time from being duplicated with the new copy having its time adjusted by some arbitrary amount such as a microsecond.
What other variables are there?
All other units of measure such as distances, weights, volumes, etc., would constitute variables if used in a claim. Geographic locations are variable. Ethnic and cultural designations are variable. Moral sentiments, prices, probabilities, levels of concern, interest, and urgency — even verbs and nouns can be variable; did she toss it or did she throw it? The world is a pretty “continuous” place — a fact which, when combined with the infinite degrees of variation between our subjective experiences of it, can make communication in a discrete medium pretty complicated. A functioning collaborative reasoning system would essentially allow us to “comb” the numerous, crisscrossing strands of our collective thought into greater alignment, even if just enough so that it flows in a more or less coherent pattern.
Solvable in principle
The vast majority of the claim space doesn’t need to be sampled. It is, in principle, possible to select only the most meaningful samples on the claim space. The questions are:
a) how to select the most meaningful samples in a way that suits most people and that accounts for the rephrasings that will probably be required in the process, and
b) whether, once everyone has reached an acceptable threshold of satisfaction with the claims, there will still be too many to manage.
Opposite of graph splitting
This is the problem of managing potentially large numbers of similar claims; it’s trying to reduce the burden of many claims. This is kind of the inverse of the problem of graph splitting which is concerned with how to bifurcate the graph into new, localized layers.
Some level of duplication will be required
// Before we get into solving the problem, it should be acknowledged that some level of duplication will be required. You’ll probably want to debate AI progress by a nice round date such as, say, 2030. But there may also be those are particularly interested in debating AI progress by December 16, 2029 when their shareholder meeting is. These are overlapping but slightly differing debates. There will always be cases where different groups of people care about slightly differing reference points on the same set of variables.
Re-evaluation of the shape of the problem
- Differences in context/variables (combinatorially explosive), esp. under high canonicality. (Context and variables are actually kind of the same. I think the only difference might be whether they’re embedded in the claim text or represented with some kind of metadata.)
- Differences in phrasing, i.e., having to do with how a problem or claim is framed conceptually including issues with claims being misleading, undecidable, or unproductive to debate. A subset of these can probably be caught at submission time.
- Differences in sentiment or degree of assertion, e.g., “significantly” vs “extremely”, “bad” vs “terrible”
- Differences in specificity. May also be kind of the same as context.
Approaches to solving the problem
The question is how to determine and focus on only the claim variants that make the most sense to debate about. This is really hard, not least because you’ll probably always leave some people unsatisfied about the way that claims are phrased which violates an important UX requirement.
1 / User Likert-style scores where possible
It may be possible to reduce the number of claims by compressing sentiments such as agreement into 5-point, Likert-style scores.
2 / Embed spectra into claims
This might make for its own type of claim. Or, I guess, it would sort of constitute a whole bunch of implicit claims. In a debate about when the earliest humans existed, you could have each piece of evidence contribute a date range that they suggest and then overlay all the date ranges, sort of when2meet style in a way that shows you where they overlap and are most highly concentrated. This is probably a really good idea.
3 / Choose variables according to what is present in source material
For variables like dates, you might be able select claims according to the specific variable values that are listed in its source material. However, as soon as you have more than one source for a claim you lose your mechanism for choosing. You could have claims associated with each source, which you may want to do anyway, but I don’t think that really helps. This topic needs some examples to continue understanding it though.
4 / Use AI to heavily prune propositions
Use AI to prune down to the most useful few. This is kind of what we do in real life. We aren’t immobilized by infinity of questions we could ask or things we could consider. We prune. But though it’s easy to prune in our own heads, to establish a set of pruning heuristics that everyone can agree on sounds difficult. Indeed, it is partly the fact that idea-pruning schemes differ so widely between individuals that makes collaboration so powerful. In the case of the UBI example, though, it’s not too hard to imagine. “Before 2028” is a fairly reasonable timeline to discuss because of the election cycle.
5 / Consolidate claim variants into “topics” and use scoring to choose specifics to highlight
Instead of saying “X is unlikely to Y” and listing a truth score of Z based only on sources that support it, you could have some kind of collection object like “Likelihood of X to Y” where you can view a range of likelihood proposals along with the relevant sources that support and detract from them. Some kind of topic object like this could also allow the most popular phrasings to bubble to the top via votes. The idea is that you’d be consolidating little nodules in the graph that are related to similar things and then using these as access points to show the top-rated samples of the claim space. If you wanted to use some kind of Similarity Mesh to manage variables such as phrasings and specificity, then this would be just sampling from local regions of that mesh.
6 / Use AI + heuristics to prevent poor phrasings in the first place
This is probably one of the first things you’d want to try — keeping bad phrasings out of the graph in the first place. One approach would be to write up a set of heuristics for good claims and let an LLM evaluate new claims on each. You could imagine gamifying this interface slightly with something akin to a password strength indicator. You want to get your claim phrasings up into the green. Or maybe there are check boxes for each heuristic so you can see what you’re aiming for. Perhaps some heuristics are optional and others are mandatory. Some heuristics may not apply in all situations so the submitter should be able to select “doesn’t apply to this claim” next to the heuristics on which they’re scoring poorly. This might trigger a developer review of the claim and heuristic to re-evaluate it. If you submit claims that are in violation of certain heuristics, perhaps the claims will be demoted in the visibility rankings. You would be made aware of this demotion as the claim submitter and could appeal to have the design modified if you think your claim should not be demoted.
7 / Other
- You could allow all variations and try to model some kind of high dimensional space where you’re interpolating between claims and things like that (think: DOE process).
- Start from sources and map the entire space. The axes of time and location create a kind of space which source information illuminates like a flashlight. You could start with your sources instead of your propositions; creating props from info that’s available and determining the specific permutations you implement based on that. After all, there’s not much point creating a prop for something if there’s no source. Still, a given source may provide insight over a certain time period, but each day within that period could constitute its own prop.
- Just let people create whatever props they want. Don’t enforce any pruning scheme and just hope that people don’t create useless props.
- Start from questions such as “when are we likely to get UBI in the United States?” Then maybe you could identify specific time points on the horizon to make other propositions about or maybe you could fill in a continuous probability graph (e.g., an increasing trend line over the years). The probability graph would likely imply way more certainty than is actually possible and therefore be misleading.
Other thoughts
It’s not just about context
Though the likelihood component contributes relatively little to the combinatorial problem (only a factor of 5 rather than of 120 for the months) this is only because probability, in this context, can be sufficiently quantized into 5 discrete options. In other words, it’s probably not a general property of non-contextual variables to contribute little to the combinatorial problem. If the context were about error rates in self-driving vehicles, for example, then probabilities would be better expressed in terms of percentages which, even assuming a reasonable binning scheme (e.g., 0.6%–1.0%, 1.1%–1.5%, etc.), would lead to far more permutations. This means that the combinatorial problem cannot be resolved by handling context alone in a clever way.
Continuity between internal and external modelling domains
There is, in principle, nothing stopping the subject of a Wicker model from extending into the realm of what unfolds inside someone’s (or a group of people’s) internal, emotional, or mental spaces. Indeed, making sense of the world often involves attempting to model the internal experiences of other people, whether they’re people you might have sympathy for, e.g., refugees, or people you might disagree with.
Time vs location
I suspect that time and location can actually be treated as the same kind of variable with respect to the combinatorial problem. I had thought that time would have specific landmarks which AI could determine but that location need be specified by any user who wants to make a proposition. But actually, location is just as variable as time. And furthermore, location admits of just the same kind of reasonable landmarks that AI can know about. For example, there is the country level, there is the state level and maybe the city level. If there are other geographical delineations which are reasonable, you could argue for them or the AI may already know them.
On context in general
It really seems like location should have a special place on proposition pages. Also, the AI should be very aware of the connections between context of, say, city and state level. Policy discussions here are always, I imagine, both somewhat distinct and somewhat related. AI can probably help in organizing propositions in a way that makes sense here.
Other other
- Different embeddings of the prop graph require different levels of specificity. For example, it’s sufficient — and in fact, preferable — in many cases, to say something like “we’re not likely to get UBI anytime soon”. “Anytime soon” is a pretty vague time horizon but if it suffices for the purposes of the NL sentence in which its embedded, then it’s better than including too much detail.