In Intelligence Explosion analysis draft: introduction, Luke Muehlhauser and Anna Salamon wrote
Norman Rasmussen's (1975) analysis of the safety of nuclear power plants, written before any nuclear accidents had occurred, correctly predicted several details of the Three Mile Island incident that previous experts had not (McGrayne 2011, 180).
I investigated this further a part of the project "Can we know what to do about AI?". The analysis that Muehlhauser and Salamon reference is WASH-1400, 'The Reactor Safety Study', which was produced for the Nuclear Regulatory Commission.
Upon investigating, I formed the impression that this case study little relevance to the question of whether we can know what to do about AI:
- The issue was one that a lot of people were already concerned about.
- The issue was highly domain specific.
- Rather than there being a few salient predictions, there were a huge number of small predictions.
One way in which the situation is relevant is that risk of nuclear power plant accidents was not adequately addressed, but that's only one of many examples people not adequately addressing risks.
A few details below.
- The report predicted one core damage accident per 20,000 years of nuclear power plant operation.
As a point of comparison, at the time of the Three Mile Island incident, only 500 years of nuclear power plant operation had occurred. This could be a fluke. The Chernobyl accident occurred only 6 years later. If the cause was the same, that strongly suggests that the Three Mile Island incident was not a fluke.
However, the report was based on reactor designs that didn't include the Three Mile Island type.
- The report did discuss tidal waves as a potential cause for nuclear disaster, anticipating the recent disaster in Japan. But the report is 21 volumes long, and so this (weak) prediction could have been cherry picked retrospectively.
- The report is considered to be obsolete, and its main lasting value seems to have been pioneering use of probabilistic risk assessment.