MIRI aims to do research now that increases humanity’s odds of successfully managing important AI-related events that are at least a few decades away. Thus, we’d like to know: To what degree can we take actions now that will predictably have positive effects on AI-related events decades from now? And, which factors predict success and failure in planning for decades-distant events that share important features with future AI events?

Or, more generally: How effectively can humans plan for future decades? Which factors predict success and failure in planning for future decades?

To investigate these questions, we asked Jonah Sinick to examine historical attempts to plan for future decades and summarize his findings. We pre-committed to publishing our entire email exchange on the topic (with minor editing), just as Jonah had done previously with GiveWell on the subject of insecticide-treated nets. The post below is a summary of findings from our full email exchange (.pdf) so far.

We decided to publish our initial findings after investigating only a few historical cases. This allows us to gain feedback on the value of the project, as well as suggestions for improvement, before continuing. It also means that we aren’t yet able to draw any confident conclusions about our core questions.

The most significant results from this project so far are:

1. Jonah’s initial impressions about The Limits to Growth (1972), a famous forecasting study on population and resource depletion, were that its long-term predictions were mostly wrong, and also that its authors (at the time of writing it) didn’t have credentials that would predict forecasting success. Upon reading the book, its critics, and its defenders, Jonah concluded that many critics and defenders had  seriously misrepresented the book, and that the book itself exhibits high epistemic standards and does not make significant predictions that turned out to be wrong.
2. Svante Arrhenius (1859-1927) did a surprisingly good job of climate modeling given the limited information available to him, but he was nevertheless wrong about two important policy-relevant factors. First, he failed to predict how quickly carbon emissions would increase. Second, he predicted that global warming would have positive rather than negative humanitarian impacts. If more people had taken Arrhenius’ predictions seriously and burned fossil fuels faster for humanitarian reasons, then today’s scientific consensus on the effects of climate change suggests that the humanitarian effects would have been negative.
3. In retrospect, Norbert Weiner’s concerns about the medium-term dangers of increased automation appear naive, and it seems likely that even at the time, better epistemic practices would have yielded substantially better predictions.
4. Upon initial investigation, several historical cases seemed unlikely to shed substantial light on our  core questions: Norman Rasmussen’s analysis of the safety of nuclear power plants, Leo Szilard’s choice to keep secret a patent related to nuclear chain reactions, Cold War planning efforts to win decades later, and several cases of “ethically concerned scientists.”
5. Upon initial investigation, two historical cases seemed like they might shed light on our  core questions, but only after many hours of additional research on each of them: China’s one-child policy, and the Ford Foundation’s impact on India’s 1991 financial crisis.
6. We listed many other historical cases that may be worth investigating.

The project has also produced a chapter-by-chapter list of some key lessons from Nate Silver’s The Signal and the Noise, available here.

Further details are given below. For sources and more, please see our full email exchange (.pdf).

Read more »

In this post, I aim to summarize one common view on AI transparency and AI reliability. It’s difficult to identify the field’s “consensus” on AI transparency and reliability, so instead I will present a common view so that I can use it to introduce a number of complications and open questions that (I think) warrant further investigation.

Here’s a short version of the common view I summarize below:

Black box testing can provide some confidence that a system will behave as intended, but if a system is built such that it is transparent to human inspection, then additional methods of reliability verification are available. Unfortunately, many of AI’s most useful methods are among its least transparent. Logic-based systems are typically more transparent than statistical methods, but statistical methods are more widely used. There are exceptions to this general rule, and some people are working to make statistical methods more transparent.

The value of transparency in system design

Nusser (2009) writes:

…in the field of safety-related applications it is essential to provide transparent solutions that can be validated by domain experts. “Black box” approaches, like artificial neural networks, are regarded with suspicion – even if they show a very high accuracy on the available data – because it is not feasible to prove that they will show a good performance on all possible input combinations.

Unfortunately, there is often a tension between AI capability and AI transparency. Many of AI’s most powerful methods are also among its least transparent:

Methods that are known to achieve a high predictive performance — e.g. support vector machines (SVMs) or artificial neural networks (ANNs) — are usually hard to interpret. On the other hand, methods that are known to be well-interpretable — for example (fuzzy) rule systems, decision trees, or linear models — are usually limited with respect to their predictive performance.¹

But for safety-critical systems — and especially for AGI — it is important to prioritize system reliability over capability. Again, here is Nusser (2009):

strict requirements [for system transparency] are necessary because a safety-related system is a system whose malfunction or failure can lead to serious consequences — for example environmental harm, loss or severe damage of equipment, harm or serious injury of people, or even death. Often, it is impossible to rectify a wrong decision within this domain.

Read more »

1. Quote from Nusser (2009). Emphasis added. The original text contains many citations which have been removed in this post for readability. Also see Schultz & Cronin (2003), which makes this point by graphing four AI methods along two axes: robustness and transparency. Their graph is available here. In their terminology, a method is “robust” to the degree that it is flexible and useful on a wide variety of problems and data sets. On the graph, GA means “genetic algorithms,” NN means “neural networks,” PCA means “principal components analysis,” PLS means “partial least squares,” and MLR means “multiple linear regression.” In this sample of AI methods, the trend is clear: the most robust methods tend to be the least transparent. Schultz & Cronin graphed only a tiny sample of AI methods, but the trend holds more broadly. ↩