- New Bostrom paper on superintelligence control: “Hail Mary, Value Porosity, and Utility Diversification.”
- MIRI’s new technical research agenda overview.
- Frank Arntzenius on infinite ethics.
- Aeon: Why has human progress slowed?
- Gary Marcus on the future of AI (EconTalk). Kind of weird that Russ & Gary wonder why there’d be a risk from general AI, but then don’t mention the standard “risk via convergent instrumental goals” argument.
- NYT reports on a Horvitz-funded 100-year Stanford program studying social impacts of AI. The white paper makes clear that the superintelligence control problem is among the intended focus areas!
- Gwern reviews the only significant biography of Hugh Everett, inventor of the many worlds interpretation of quantum mechanics.
- Nick Bostrom discusses Superintelligence on EconTalk.
- GiveWell’s top charity updates for December 2014.
- Armstrong, “Why aim for the stars when galaxies are just as easy?” (TEDx talk)
Decent books finished in November:
- Strogatz, Sync
- Diamond, The Third Chimpanzee
- Dormehl, The Formula
- Hazen, The Story of Earth
- Schneier, Schneier on Security
- Diamond, Collapse
- de Queiroz, The Monkey’s Journey
- Zetter, Countdown to Zero Day
- Isaacson, Einstein
- Diamond, Guns, Germs, and Steel (re-read)
In case you hadn’t noticed, my choice of books is dominated by whether a book is available in audiobook, since the only time I have available to consume books these days is when my chronic insomnia is blocking me from sleeping anyway, so I put on my eye mask and open the Audible app on my phone. If I had more hours each month during which my eyeballs were still working, I’d be consuming a pretty different set of books, and I’d be reading them more carefully and critically rather than zooming through lots of easy material quickly.
Stuff I wrote elsewhere in November:
- Stuart Russell’s reply to “The Myth of AI” is very good, except that I don’t like the “provably aligned” phrase.
- Scale model of the solar system, on a web page that is (necessarily) half a mile long.
- Graham, “Mean People Fail.” More fodder for that article I swear I’m going to write one day about kindness as an epistemic virtue.
- Just FYI, I continue to update The Best Textbooks on Every Subject in response to new submissions.
- Kuperberg’s parable on motivated skepticism.
- Tenenbaum & Marcus debate the merits of the Bayesian cognition approach. (I come down much closer to Marcus, and I suspect the Tenenbaum tradition will end up being more productive for AI than for psychology.)
Earlier today I published a post called “Sometimes evil.” I’ve taken it down for editing, as the original draft was unclear and perhaps too tempting for trolls to misrepresent.
- The Naam-Hertling AI foom debate: one, two, three.
- Three versions of the talk Bostrom gave during his book tour for Superintelligence: at UC Berkeley, at Harvard, at Google.
- Which apps and tools actually keep your messages safe? The EFF secure messaging scorecard.
- Did Rutherford really think the idea of harnessing atomic power was “moonshine”? John Jenkin casts doubt.
- Which pieces of “philosophically interesting” science fiction do professional philosophers recommend?
- An index for John Danaher’s philosophical analyses of intelligence explosions and advanced robotics.
- Dewey, “Long-term strategies for ending existential risk from fast takeoff.”
- Hibbard, “Ethical artificial intelligence” (book draft, 165 pages, 34 figures).
- Johnson, How We Got to Now
- Bryson, At Home
- Kean, The Violinist’s Thumb
- Isaacson, The Innovators
- Munroe, What If?
- Schmidt & Rosenberg, How Google Works
- Harris, The Nurture Assumption (abridged version)
- Wu, The Master Switch
- Gawande, Being Mortal
- Buss & Meston, Why Women Have Sex
The Goal is the worst novel I’ve ever read, and the first long piece of fiction I’ve finished in the past 6 years. How is that possible?
The novel is actually an introduction to Goldratt’s theory of constraints, from operations research. The writing isn’t supposed to be good, it’s just supposed to drive home the principles of Goldratt’s theory clearly and efficiently. I think I was able to finish this one because the book wasn’t trying to do all the normal literary things that good novels try to do, and instead was clearly trying to teach me things, like a nonfiction book would. And because it was explained with a story, I’ll probably remember the core principles of The Goal better than I remember the key points of most nonfiction books I read.
The audiobook is especially amusing. Every character is played by a different voice actor, there are ambient background noises that fit the scene, and the musical queues are often hilarious, such as the vaguely romantic hold music that plays while the main character and his wife make up after a fight and speak corny romance dialogue.
Dweck’s Mindset — see this summary — was alternately decent and annoying. I do suspect there’s something to the growth/fixed mindset distinction, but Dweck downplays individual differences too much, glosses over conflicting studies, and waves suggestively in the direction of debunked blank slate hypotheses.
Pollan’s The Omnivore’s Dilemma is well-written but the arguments aren’t consistently compelling.
Harris’ 10% Happier was moderately enjoyable but doesn’t have much content.
- Google continues gobbling up AI talent.
- New-ish psychology study method: “preregistered adversarial collaboration.” From the abstract: “Prior to data collection, the [disagreeing researchers] reached consensus on an optimal research design, formulated their expectations, and agreed to submit the findings to an academic journal regardless of the outcome… [they also] set up a publicly available… agreement that detailed the proposed design and all forseeable aspects of the data analysis.”
- I just noticed the WTF, Evolution? book is out.
I like high-quality written debates for which (1) one ‘top thinker’ on the subject makes an argument, (2) four or more other top thinkers on the subject reply, and (3) then the first author writes a final reply to his or her critics. I think of these as “symposium-style debates,” as contrasted with e.g. the Oxford-style debates seen in Economist Debates and elsewhere. (Is there another name for them?)
- Boston Review‘s Forum
- Cato Unbound
- Brain and Behavioral Sciences
- Psychological Inquiry
- American Journal of Bioethics (almost: there’s no final reply in this case)
- Various journal special issues and academic edited volumes that serve as symposia for published books or invited papers/chapters
Know of other examples?
- Hsu, “Super-intelligent humans are coming.”
- Winikoff, “Assurance of agent systems: what role should formal verification play?“
- New Friendly AI open problem description: “Corrigibility.”