What is the null hypothesis here? That Microsoft is, through light fine-tuning, optimizing the response of journalists and AI Safety researchers/commentators. The model is designed to give weird responses, so as to make people talk about it.
A Flood of Ideas: The Null Hypothesis of AI Safety with respect to Bing Chat
Definitely. Scott Alexander writes eloquently about this in “Rule Thinkers In, Not Out” (https://www.lesswrong.com/posts/u8GMcpEN9Z6aQiCvp/rule-thinkers-in-not-out) - in exchange for accuracy by a thinker in some domain, you have to put up with a lot of variance, bias, or both.
To be at first glance but not actually contrarian: we are not putting enough people (who should be there) into psychiatric hospitals. This is not exclusionary of the idea that people are winding up in mental asylum who should not be there. Rather, the two are complimentary: poor diagnostics and lack of accountability, as well as limited resources, ensure that there are likely large amounts of both “should not be institutionalized but are institutionalized” and “should be institonalized but are not.”
I do not live in the United States, but in a similar Western democracy and worked for a year in a health bureaucracy that contracted psychiatrists to work in mental hospitals. The doctors were under great pressure, due to lack of funding, to release people as soon as possible, whether or not mental health support was available in the community they would be returning to (in most cases, this meant “the street/homeless shelter”). As a result, relapse/readmission was near constant.
As to what prevents them from becoming a constituency for the modern left—even when many of them are ethnic minorities—the reasons identified below are cogent, and I would add, to judge from the people we dealt with in my job, the mentally ill are unpleasant/unsettling to deal with. I know many people who will talk about their grad school depressions as a major mental illness, but they have nothing on the cases the psychiatrists I worked with handled. Serial arsonists, people who habitually ate glass and metal, others convinced that all indigenous people were possessed by the devils and lashed out at them at every opportunity.
The mentally ill are also often the cause of many other (politically preferred) social problems. Recently, in my city, there were a string of verbal assaults and physical intimidations of Muslim women at transit stations. After widespread condemnation of these attacks and community groups talking about the need to “raise awareness” , with an implicit assumption that the actors were many, the police figured out the attacks were the work of two mentally ill men with long histories of on again/off again institutionalization. The attacks stopped and the issue disappeares from the news. To be blunt, no one gains anything from directing attention at the mentally ill.
As a difference between rates of growth, 3% is 1.5 greater than 2%. The question is a trick one and plays on public neglect of the nature of compounding growth.
Taking an economy of size 100 in Year Zero (Y0). At Y1:
2% growth yields an economy of size 102 3% growth yields an economy of size 103
Not very impressive.
But at Y10:
2% = 121.9
3% = 134.3
And at Y20:
2% = 148.6
3% = 180.6
All else being equal, you’re substantially better off with 3% growth than 2%, and increasingly better off over time. I believe we are better off with voters who understand that and elect politicians accordingly.
(The example comes from George Will, who in an EconTalk interview voiced his despair that “Washington is full of people who think the difference between 2% GDP growth per year and 3% GDP growth per year is only 1%”)
Implementation problems are definitely a problem with Brennan’s Knowledge Test To Vote idea and consist of two parts:
(1) getting the present voters to agree to it (2) setting a test that is discriminatory in the right rather than the wrong ways.
One would hope a good answer to (2) would help with (1), though convincing people to give up the vote would be very hard.
I have been thinking a fair bit lately about the content of a Voting Test. Presumably one would want tests of knowledge that are proxies for being what Brennan calls a Vulcan—an informed Non-partisan voter who considers things like evidence—rather than a Hooligan—informed partisan—or Hobbit—uninformed and nonpartisan. Brennan’s idea to test for basic knowledge about government is a good start—how does a bill become law, how do the different branches of government work, how much does your country spend on foreign aid as a percentage of government expenditures (the latter being something surveyed voters consistently and overwhelmingly get wrong).
I would add to such a test sections for basic probability, statistics, and economics as these are vital for understanding public policy issues. Anyone who thinks the difference between 2% annual GDP growth and 3% annual GDP growth is 1% has next to nothing to contribute to public discourse.
Brennan considers the question at length in his book, precisely because of unreasonable restrictions of suffrage in the past. The level of knowledge he is seeking is not high—knowing the distinction between the legislative, executive, and judicial branches, or the outline of how a congressional bill becomes law, fundamental questions of fact about how the current government works rather than contested questions about history. Shockingly, the majority of the eligible voters in all countries surveyed are unable to achieve better than 50% on basic knowledge tests (relative to their own country—it makes no sense to quiz Swedes about Australian parliamentary procedure).
Are you open to the idea of sailing right around this Scylla and Charybdis by discarding mass-participation democracy, or must the solution set be within the set of possible democracies?
Because if you are open to less-than-democratic solutions, restricting the voting franchise seems like a promising way forward. On this, try Jason Brennan’s “Against Democracy”, a critique of mass participation democracy, and an argument for a system where prospective voters must pass a knowledge test in order to vote.
Thank you for the positive review and good questions (and please forgive the lateness of this reply).
In reply:
there are certainly times where I am not in the mood for a specific random choice. When that happens, I allow myself one more random selection. Because I have set the sublists for maximal variety (i.e. There is only one “cosmic horror story” category) the next selection usually is something I find compelling and interesting enough to complete.
In the rare case where the second selection is still not interesting, I try to reflect on why, and ask myself whether what I want is actually a different kind of break or distraction—watching a short video, listening to music or a podcast, going for a walk, talking with a friend.
I am uncertain of how it would work for longer works. If a short story is quite long, such as a novella, I will keep reading it at each reading break until I finish it. I do use the technique for picking novels to read though and have found it useful for that. For novels, I have a set time each day—the hour before I go to bed—that I use exclusively for reading novels.
As the other commenter have been saying, excellent post.
There is an additional reason to believe, at least given contemporary capabilities and strategies, that the X-risk of an actual nuclear conflict is small. A few years ago I wrote to Fred Kaplan, the author of the stellar military history book “The Wizards of Armageddon”*, a history of US nuclear war planning from 1945-1990. I asked Kaplan what he judged the present state of nuclear war planning was. He responded to me that his sources informed him that nuclear war plans, in the US and presumably the Russian Federation, had been shaped by the same changes that shaped conventional war strategy from the Gulf War onward. The focus, in both nuclear and conventional war, is blinding and decapitation of the other side—destroying their C3I (command-control-communication-intelligence) infrastructure and killing their national command authority. The idea is to render an enemy unable to communicate with and deploy their nuclear forces, rendering them inert. One of the more likely outcomes of a nuclear conflict is the two nations being leaderless but largely intact, with only a few dozen low yield devices (<1 MT) having been used, and the rest stranded and unusable. The problem is that, as documented by Daniel Ellsberg in “The Doomsday Machine: Confessions of a Nuclear War Planner”, many countries likely hedge against this strategy by pre-delegation, telling the commanders of their nuclear forces that they may, if contact is lost with civilian authority during a crisis, use their nuclear weapons at their own discretion. But even that is, for the reasons listed above, unlikely to yield an X-Risk scenario. A really, really, really awful situation to live through, but not an X-Risk.
*which I highly recommend. Kaplan is critical of the nuclear war planners, but I think most of the X-risk people on this forum and in academia would have fit right in at RAND and other strategic think tanks during the Cold War.
I agree. But...
Devil’s Advocate: many anti-features of consumer products are there to protect the 99% of users who are not power users like yourself, and are a positive benefit for them. For example, the iPhone OS and UI protects most users from accidentally disabling their phones in ways they cannot understand or fix, limiting the utility of the device for them. Schools and tutors serve the great majority who have difficulty teaching themselves (and provide strong educational signaling benefits in a way that autodidacticism regrettably does not). Psychoactive drugs are second best (or third-best, or at least for many users better than nothing) meditation practice, in an analogous way to how caffeine substitutes for sleep. Jobs, by packaging tasks together, limit the search cost for people trying to find an economic niche in which they can thrive—most people want 8 hours of structured work, after which they are free to do whatever they want and can leave thinking about work at the office, as opposed to having to strive to consciously separate work and life the way entrepreneurs do.
If (a big “if”) ease of use and reduction of conscious, deliberate choice has been a net positive for consumers, these anti-features are the means to deliver them. You (and I, and many other users of this site) suffer because of that, but there are enough of us to support the flourishing open source ecosystem, so the results are net gains at the level of society.
Excellent scenario building! Like other commenters, I had been toying around with scenarios like this, and it’s good to see someone put so much effort into making a highly-detailed and plausible one.
Extra kudos for avoiding the Singleton flaw of most AI scenarios, where there is “one model to rule them all” instead of countless powerful actors working in alternately (and sometimes simultaneously) cooperative and competitive ways.