Sep 15, 2012
Due in part to Eliezer's writing style (e.g. not many citations), and in part to Eliezer's scholarship preferences (e.g. his preference to figure out much of philosophy on his own), Eliezer's Sequences don't accurately reflect the close agreement between the content of The Sequences and work previously done in mainstream academia.
I predict several effects from this:
I'd like to counteract these effects by connecting the Sequences to the professional literature. (Note: I sort of doubt it would have been a good idea for Eliezer to spend his time tracking down more references and so on, but I realized a few weeks ago that it wouldn't take me much effort to list some of those references.)
I don't mean to minimize the awesomeness of the Sequences. There is much original content in them (edit: probably most of their content is original), they are engagingly written, and they often have a more transformative effect on readers than the corresponding academic literature.
I'll break my list of references into sections based on how likely I think it is that a reader will have missed the agreement between Eliezer's articles and mainstream academic work.
(This is only a preliminary list of connections.)
An Intuitive Explanation of Bayes' Theorem, How Much Evidence Does it Take, Probability is in the Mind, Absence of Evidence Is Evidence of Absence, Conservation of Expected Evidence, Trust in Bayes: see any textbook on Bayesian probability theory, e.g. Jaynes (2003) or Friedman & Koller (2009).
What's a Bias, again?, Hindsight Bias, Correspondence Bias; Positive Bias: Look into the Dark, Doublethink: Choosing to be Biased, Rationalization, Motivated Stopping and Motivated Continuation, We Change Our Minds Less Often Than We Think, Knowing About Biases Can Hurt People, Asch's Conformity Experiment, The Affect Heuristic, The Halo Effect, Anchoring and Adjustment, Priming and Contamination, Do We Believe Everything We're Told, Scope Insensitivity: see standard works in the heuristics & biases tradition, e.g. Kahneman et al. (1982), Gilovich et al. 2002, Kahneman 2011.
Eliezer's metaethics sequences includes dozens of lemmas previously discussed by philosophers (see Miller 2003 for an overview), and the resulting metaethical theory shares much in common with the metaethical theories of Jackson (1998) and Railton (2003), and must face some of the same critiques as those theories do (e.g. Sobel 1994).
Eliezer's free will mini-sequence includes coverage of topics not usually mentioned when philosophers discuss free will (e.g. Judea Pearl's work on causality), but the conclusion is standard compatibilism.
How an Algorithm Feels From Inside and Dissolving the Question suggest that many philosophical problems can be dissolved into inquiries into the cognitive mechanisms that produce them, as also discussed in, for example, Shafir (1998) and Talbot (2009).
Thou Art Godshatter, Not for the Sake of Happiness Alone, and Fake Utility Functions make the point that value is complex, a topic explored in more detail in affective neuroscience (Kringelbach & Berridge 2009), neuroeconomics (Glimcher 2010; Dolan & Sharot 2011), and other fields.
Newcomb's Problem and the Regret of Rationality repeats a common debate among philosophers. Thinking that CDT must be right even though it "loses" to EDT on Newcomb's Problem, one group says "What can we do, if irrationality is rewarded?" The other group says "If you're so smart, why aren't you rich? What kind of rationality complains about the reward for irrationality?" For example, see Lewis (1981).
Eliezer's TDT decision algorithm (2009, 2010) had been previously discovered as a variant of CDT by Wolfgang Spohn (2003, 2005, 2012). Both TDT and Spohn-CDT (a) use Pearl's causal graphs to describe Newcomblike problems, then add nodes to those graphs to represent the deterministic decision process the agent goes through (Spohn calls them "intention nodes," Yudkowsky calls them "logical nodes"), (b) represent interventions at these nodes by severing (edit: or screening off) the causal connections upstream, and (c) propose to maximize expected utility by summing over possible values of the decision node (or "intention node" / "logical node"). (Beyond this, of course, there are major differences in the motivations behind and further development of Spohn-CDT and TDT.)
Many of Eliezer's points about intelligence explosion and machine ethics had been made in earlier writings Eliezer did cite, e.g. Williamson (1947), Good (1965), and Vinge (1993). Others of Eliezer's points appear in earlier writings he did not cite but probably had read: e.g. Minsky (1984), Schmidhuber (1987), Bostrom (1997), Moravec (1999). Others of Eliezer's points appear in earlier writings he probably hadn't read: e.g. Cade (1966), Good (1970), Versenyi (1974), Lukasiewicz (1974), Lampson (1979), Clarke (1993, 1994), Sobel (1999), Allen et al. (2000). (For a brief history of these ideas, see here and here.)