onsdag 15 augusti 2018

Singularities

Just today, I came across the 2017 paper Singularities and Cognitive Computing. It deals with AI futurology, a topic I am very much interested in. Author of the paper is Devdatt Dubhashi. Here are four things that struck me, from a mostly rather personal perspective, about the paper:
    (1) The name Häggström appears four times in the short paper, and in all four cases it is me that the name refers to. I am flattered by being considered worthy of such attention.
So far so good, but my feelings about the remaining points (2)-(4) are not quite as unambiguously positive. I'll refrain from passing moral judgement on them - better to let them speak for themselves and let the reader be the judge.
    (2) The paper was published in the summer of 2017, a large fraction of it is devoted to countering arguments by me, and the author is a Chalmers University of Technology colleague of mine with whom I've previously had fruitful collaborations (resulting in several joint papers). These observations in combination make it slightly noteworthy that the paper comes to my attention only now (and mostly by accident), a full year after publication.

    (3) The reference list contains 10 items, but strikingly omits the one text that almost the entire Section 2 of the paper attempts to engage with, namely my February 2017 blog post Vulgopopperianism. That was probably not by mistake, because at the first point in Section 2 in which it is mentioned, its URL address is provided. So why the omission? I cannot think of a reason other than that, perhaps due to some grudge against me, the author wishes to avoid giving me the bibliometric credit that mentioning it in the reference list would yield. (But then why mention my book Here Be Dragons in the reference list? Puzzling.)

    (4) In my Vulgopopperianism blog post I discuss two complementary hypotheses (H1) and (H2) regarding whether superintelligence is achievable by the year 2100. Early in Section 2 of his paper, Dubhashi quotes me correctly as saying in my blog post that "it is not a priori obvious which of hypotheses (H1) and (H2) is more plausible than the other, and as far as burden of proof is concerned, I think the reasonable thing is to treat them symmetrically", but in the very next sentence he goes overboard by claiming that "Häggström suggests [...] that one can assign a prior belief of 50% to both [(H1) and (H2)]". I suggest no such thing in my blog post, and certainly do not advocate such a position (unless one reads the word "can" in Dubhashi's claim absurdly literally, meaning "it is possible for a Bayesian to set up a model in which each of the hypotheses has probability 50%"). If the sentence that he quoted from my blog post had contained the passage "as far as a priori probabilities are concerned" rather then "as far as burden of proof is concerned", then his claim would have been warranted. But the fact is that I talked about "burden of proof", not "a priori probabilities", and it is clear from this and from the surrounding context that what I was discussing was Popperian theory of science rather than Bayesianism.1 It is still possible that the mistake was done in good faith. Perhaps, despite being a highly qualified university professor, Dubhashi does not understand the distinction (and tension) between Popperian and Bayesian theory of science.2

Footnotes

1) It is very much possible to treat two or more hypotheses symmetrically without attaching them the same prior probability (or any probability at all). As a standard example, consider a frequentist statistician faced with a sample from a Gaussian distribution with unknown mean μ and unknown variance σ2, making a 95% symmetric confidence interval for μ. Her procedure treats the hypotheses μ<0 and μ>0 symmetrically, while not assigning them any prior probabilities at all.

2) If this last speculation is correct, then one can make a case that I am partly to blame. In Chapter 6 of Here Be Dragons - which Dubhashi had read and liked - I treated Popperianism vs Bayesianism at some length, but perhaps I didn't explain things sufficiently clearly.

Inga kommentarer:

Skicka en kommentar