Description vs simulated prediction

By Rick Korzekwa, 22 April 2020

Throughout our investigation into discontinuous progress, we had some discussions about what precisely it’s that we’re attempting to do after we analyze discontinuities in technological progress. There was some disagreement and (at the very least on my half) some confusion about what we have been attempting to study from all of this. I feel this comes from two separate, extra basic questions that may come up when forecasting future progress. These questions are intently associated and each may be answered partly by analyzing discontinuities. First I’ll describe these questions generically, then I’ll clarify how they will grow to be confused within the context of analyzing discontinuous progress.

Query 1: How did tech progress occur up to now?

Realizing one thing about how historic progress occurred is essential to creating good forecasts now, each from the standpoint of understanding the underlying mechanisms and establishing base charges. Merely having the ability to describe what occurred and why might help us make sense of what’s occurring now and may allow us to make higher predictions. Such descriptions fluctuate in scope and element. For instance:

• The variety of transistors per chip doubled roughly each two years from 1970 to 2020
• Wartime funding throughout WWII led to the fast improvement and scaling of penicillin manufacturing, which contributed to a 90% lower in US syphilis mortality from 1945 to 1967
• Massive jumps in metrics for technological progress weren’t all the time accompanied by basic scientific breakthroughs

This type of evaluation could also be used for different work, along with forecasting charges of progress. Within the context of our work on discontinuities, answering this query largely consists of describing quantitatively how metrics for progress evolve over time.

Query 2: How would now we have fared making predictions up to now?

That is truly a household of questions aimed toward creating and calibrating prediction strategies based mostly on historic information. These are questions like:

• If, up to now, we’d predicted that the present development would maintain, how usually and by how a lot would now we have been improper?
• Are there domains through which we’d have fared higher than others?
• Are there heuristics we will use to make higher predictions?
• Which strategies for characterizing traits in progress would have carried out one of the best?
• How usually would now we have seen hints {that a} discontinuity or change in fee of progress was about to occur?

These questions usually, however not all the time, require the identical method

Within the context of our work on discontinuous progress, these questions converge on the identical strategies more often than not. For a lot of of our metrics, there was a transparent development main as much as a discontinuity, and describing what occurred is basically the identical as making an attempt to (naively) predict what would have occurred if the discontinuity had not occurred. However there are occasions after they differ. Specifically, this will occur when now we have totally different data now than we might have had on the time the discontinuity occurred, or when the naive method is clearly lacking one thing necessary. Three instances of this that come to thoughts are:

The development main as much as the discontinuity was ambiguous, however later information made it much less ambiguous. For instance, advances in steamships improved instances for crossing the Atlantic, nevertheless it was not clear whether or not this progress was exponential or linear on the time that flight or telecommunications have been invented. But when we have a look at progress that occurred for transatlantic ship voyages after flight, we will see that the general development was linear. If we wish to reply the query “What occurred?”, we’d say that progress in steamships was linear, in order that it will have taken 500 years on the fee of development for steamships to carry crossing time all the way down to that of the primary transatlantic flight. If we wish to reply the query “How a lot would this discontinuity have affected our forecasts on the time?”, we’d say that it seemed exponential, in order that our forecasts would have been improper by a considerably shorter period of time.

We now have entry to data from earlier than the discontinuity that no one (or nobody particular person) had entry to on the time. Prior to now, the world was a lot much less related, and it isn’t clear who knew about what on the time. For instance, constructing heights, altitude data, bridge spans, and navy capabilities all confirmed progress throughout totally different elements of the world, and it appears possible that no one had entry to all the data that now we have now, in order that forecasting might have been a lot tougher or yielded totally different outcomes. Info that’s actively stored secret might have made this downside worse. It appears believable that no one knew the state of each the Manhattan Mission and the German nuclear weapons program on the time that the primary nuclear weapon was examined in 1945.

The within view overwhelms the surface view. For instance, the second transatlantic telegraph cable was a lot, significantly better than the primary. Utilizing our methodology, it was practically as massive an development over the primary cable as the primary cable was over mailing by ship. However we lose quite a bit by viewing these advances solely when it comes to their deviation from the earlier development. The primary cable had extraordinarily poor efficiency, whereas the second carried out about in addition to a typical excessive efficiency telegraph cable did on the time. If we have been attempting to foretell future progress on the time, we’d give attention to questions like “How lengthy do we predict it can take to get a standard cable working?” or “How quickly will or not it’s till somebody is keen to fund the subsequent cable laying expedition?”, not “If we draw a line by way of the factors on this graph, the place does that take us?” (Although that exterior view should be price consideration.) Nonetheless, if we’re simply attempting to describe how the metric developed over time, then the right factor to do is to only draw a line by way of the factors as greatest we will and calculate how far off-trend the brand new development is.

Causes for focusing extra on the descriptive method for now

Each of those questions are necessary, and we will’t actually reply one whereas completely ignoring the opposite. However for now, now we have centered extra on describing what occurred (that’s, answering query 1).

There are a number of causes for this, however first I’ll describe some benefits to specializing in simulating historic predictions:

1. It mitigates what could also be some deceptive outcomes from the descriptive method. See, for instance, the outline of the transatlantic telegraph above.
2. We’re attempting to do forecasting (or allow others to do it), and actually good solutions to those questions could be extra beneficial.

However, for now, I feel the benefits to specializing in description are larger:

1. The outcomes are extra readily reusable for different initiatives, both by us or by others. For instance, answering a query like “How a lot of an enchancment is a typical main development over earlier expertise?”
2. It doesn’t require us to mannequin a hypothetical forecaster. It’s laborious to foretell what we (or another person) would have predicted if requested about future progress in weapons expertise simply earlier than the invention of nuclear weapons. To me, this course of feels prefer it has quite a lot of transferring elements or at the very least quite a lot of subjectivity, which leaves room for error, and makes it tougher for different folks to guage our strategies.
3. It’s simpler to construct from query 1 to query 2 than the opposite manner round. An outline of what occurred is a fairly cheap place to begin for determining which forecasting strategies would have labored.
4. It’s simpler to check throughout applied sciences utilizing query 1. Query 2 requires taking a extra inside view, which makes comparisons tougher.

Notes