Posts
Comments
If you have the power to change the Google form, by the way, one of its questions is "What dose did you take (in mg)? 0.5, 1, 1.5, 2+"
Presumably this should read "(in g)", and it would also help if it were explicitly stated as being "per day".
This isn't necessarily something you have to be tricked by a third party into. Be more Gwern! If there are two brands of cola you've not tried before, one stevia and one not, you can do a blinded trial by decanting them or similar. It'll certainly be easier with a third party, but one could do this solo.
There's an analogy with the notion of "toil" which is popular in the Site Reliability Engineering subfield of software engineering. Toil in this context is work which is necessary to keep the lights on, but which doesn't actually improve anything. In some sense, the job of an SRE is to reduce toil; they must certainly be psychologically able to deal with it, because it's the stuff with which they work! I'll just talk a bit about it here in a fairly undirected way, in case any of it gives you ideas. The SRE Handbook is well worth reading if you're a software engineer, by the way.
The SRE's (imperfectly-aligned-to-your-problem) answer to the problem "we're being buried in toil" is to track the proportion of time spent on toil versus "productive" work. If the toil becomes greater than some proportion, the response is to divert resources from feature work towards reducing the toil (e.g. by automating it, or addressing the root causes of the issues that you're spending time fighting). An extremely simple example of such automation is setting up direct debits to pay bills, or repeat online orders for groceries. An SRE performing any particular piece of toil would at least spend a moment to think about whether it could be automated instead.
Runbooks (lists of triggers and responses to guide you through operations) are a standard SRE-style tool for making the toil less error-prone and stressful. To know when you should be performing some piece of toil, it's standard to identify and set up alerts, so that you have a specific trigger. ("I just got a Slack alert saying that the database has reached 70% capacity; the alert pointed me to this wiki page telling me step-by-step how to bring the database offline safely and perform a vacuum to release space", or "my washing basket is 3/4 full; that means this evening I will be putting on a load of laundry".)
It's also standard to batch up the toil. A team of people will usually have a rota, so that any given person's time is mostly spent doing productive work, and the toil is the responsibility of the people on duty. That way, you only get a small amount of relative hell before you rotate onto better work. The toil necessary to maintain a human life is generally not that urgent and is hence very amenable to batching, except for the most basic biological things like using the toilet or putting food into your mouth (note: preparing food is not an urgent biological need unless your planning procedures have failed!). You can batch up a lot of it: e.g. you spend half of Saturday preparing meals for the week, or otherwise arranging so that the daily time spent preparing and putting food into your mouth is as low as possible, and you can declare that one Sunday every two months is paperwork.
It would be nice to have example GPT4 outputs for each demonstrating the wrongness, because I tried "Continue the sequences: 5, 8, 13," expecting the answer 21, and for me it did indeed explain along the lines "21, because Fibonacci". As you say, this dataset is inherently unstable over time, so it would be nice to snapshot it. (One obvious way would be to convert from a list of strings to a dictionary of `{ "prompt": ["response1", "response2", …] }`; the current schema injects into this by setting all those lists to be empty.)
suffering is bad because anyone who suffering is objectively in negative state of being.
I believe this sentence reifies a thought that contains either a type error or a circular definition. I could tell you which if you tabooed the words "suffering" and "negative state of being", but as it stands, your actual belief is so unclear as to be impossible to discuss. I suspect the main problem is that something being objectively true does not mean anyone has to care about it. More concretely, is the problem with psychopaths really that they're just not smart enough to know that people don't want to be in pain?
By the way, you're making an awful lot of extremely strong and very common points with no evidence here ("ChaosGPT is aligned", "we know how to ensure alignment", "the AI understanding that you don't want it to destroy humanity implies that it will not want to destroy humanity", "the AI will refuse to cooperate with people who have ill intentions", "a system that optimises a loss function and approximates a data generation function will highly value human life by default", "a slight misalignment is far from doomsday", "an entity that is built to maximise something might doubt its mission"), as well as the standard "it's better to focus on X than Y" in an area where almost nobody is focusing on Y anyway. What's your background, so that we can recommend the appropriate reading material? For example, have you read the Sequences, or Bostrom's Superintelligence?
(Or more concretely, Grand Central Station wasn't a Schelling point in New York before it was built. Before that time, presumably there were different Schelling points.)
Fittingly, I… don't think those words actually identify sazen :P I claim that "the thing you get if you do not take inferential distance into account" for most people would be baffled non-comprehension, not active misunderstanding.
Wonderful, thanks! Recording the quote for posterity:
Nothing can be soundly understood
If daylight itself needs proof.
(Imam al-Haddad, The Sublime Treasures)
Indeed, this is what I use. It feels much more natural to me in the following case, where obviously our statement is not a question:
Dr Johnson kicked a large rock, and said, as his foot rebounded, "Do I refute it thus?".
And "obviously" the full stop should go outside, because of:
Dr Johnson kicked a large rock, and said, as his foot rebounded, "Do I refute it thus?", howling with pain.
And there's nothing special about a question mark, so this rule should be identical if a full stop is substituted.
I will pick a rather large nit: "for example a web server definitely doesn't halt" is true, but for this to be surprising or interesting or a problem for Turing reasons, it just means you are modelling it incorrectly. Agda solves this using corecursion, and the idea is to use a data type that represents a computation that provably never halts. Think infinite streams, defined as "an infinite stream is a pair, $S_0 = (x, S_1)$, where $S_1$ is an infinite stream". This data type will provably keep producing values forever (it is "productive"), and that's what you want for a web server.
I'm pretty sure it's not schools, unless private schools somehow have a massive impact. The case rates were already dropping on July 21st, which is presumably a couple of days after The Event anyway; the summer holidays for state schools (i.e. the vast majority of children) started on the 25th.
Irrelevant nit: the archaic second-person singular of "do" is "dost", as in "dost thou not know". "Doth" is the third-person form, as in "the lady doth protest too much".
For some reason I can't find any relevant hits with Google, but I've heard "support vs advice" described as "sympathy or fascism" before. "I want to moan at you" vs "I want you to take over and solve my problem".
For some years now I have had a Panasonic breadmaker, model SD-ZB2512. It takes less than five minutes in the evening, generating no mess and no washing up (if you use olive oil instead of butter, so as to avoid generating a fatty knife), and you can have hot fresh bread ready-baked as you wake up. The only downside to bread made this way is that you have to slice it. It tastes dramatically better than all but the most expensive shop-bought bread, and the ingredients store in a cupboard for literally months so it's even highly pandemic-proof. Bread that is still hot from the breadmaker is really one of the best foods I know. The breadmaker has literally no cost to upkeep: you don't even need to clean it, as it's basically an oven in a pot.
(Posting this in a spirit of self-congratulation: I wrote up a spiel about what I found confusing, and then realised that I'm confused on a much more fundamental level about the nature of the various explanations and how they relate to each other, and am now going back to reread the various sources rather than writing something unhelpfully confusing about a confused confusion.)
Strong +1 to the idea; I'll be on a different team, but I strongly encourage people to give it a try. I think Hunt 2019 was quite possibly the most fun I have ever had.
My immediate reaction is that I remember hating it very much at school when a teacher punished the entire class for the transgression of an unidentifiable person!
Nitpick: I think there's a minor transcription error, in that "biological-esque risk" should read "biological X-risk".
You're thinking of "Glomarisation" (https://en.wikipedia.org/wiki/Glomarization).
See, for example, https://www.lesswrong.com/posts/xdwbX9pFEr7Pomaxv/meta-honesty-firming-up-honesty-around-its-edge-cases and https://www.lesswrong.com/posts/bP5sbhARMSKiDiq7r/consistent-glomarization-should-be-feasible.
I'm a big believer in "the types should constrain the semantics of my program so hard that there is only one possible program I could write, and it is correct". Of course we have to sacrifice some safety for speed of programming; for many domains, being 80% sure that a feature is correct in 95% of the possible use cases is good enough to ship it. But in fact I find that I code *faster* with a type system, because it forces most of the thinking to happen at the level of the problem domain (where it's easy to think, because it's close to real life); and there are a number of ways one can extremely cheaply use the type system to make invalid states unrepresentable in such a way that you no longer have to test certain things (because there's no way even to phrase a program that could be incorrect in those ways).
For a super-cheap example, if you know that a list is going to be nonempty, use a non-empty list structure to hold it. (A non-empty list can be implemented as a pair of a head and a list.) Then you can save all the time you might otherwise have spent on coding defensively against people giving you empty list inputs, as well as any time you might have spent testing against that particular corner case.
For another super-cheap example that is so totally uncontroversial that it probably sounds vacuous (but it is in fact the same idea of "represent what you know in the type system so that the language can help you"), don't store lists of (key, value); store a dictionary instead, if you know that keys are unique. This tells you via the type system that a) keys are definitely unique, and b) various algorithms like trees or hashmaps can be used for efficiency.
I believe the world is this way because of the following two facts:
- monads are very hard to get your head into;
- monads are extremely simple conceptually.
This means that everyone spends a long time thinking about monads from lots of different angles, and then one day an individual just happens to grok monads while reading their fiftieth tutorial, and so they believe that this fiftieth tutorial is The One, and the particular way they were thinking about monads at the time of the epiphany is The Way. So they write yet another tutorial about how Monads Are Really Simple They're Just Burritos, and meanwhile their only actual contribution to the Monad Exposition Problem is to have very slightly increased the number of paths which can lead an individual to comprehension.
I'm interested in your comment about "using dynamic-untyped rather than well-typed because it helps you not worry about your own intelligence". I use well-typed languages religiously precisely for that reason: I'm not smart enough to program in an untyped language without making far too many mistakes, and the type system protects me from my idiocy.
You can buy good tomatoes (in the UK); they're just a bit expensive. Cheap tomatoes are nasty, but nice tomatoes are widely available; I get them from a company called Isle of Wight Tomatoes, and they're on Ocado.
I stopped taking the book seriously when I reached Walker's suggestion that teenagers might have a sleep cycle offset from adults because "wise Mother Nature" was giving them the chance to develop independence from the tribe, in a group of their peers, and that this was an important stage in societal development of a human.
If one *must* find an evo-psych explanation for this phenomenon, surely "we need people guarding the camp at more hours of the day" is simpler and less ridiculously tenuous. (Though this still has precisely the same "I could have explained anything with this" flavour that most popular evo-psych does.)
I've had experiences ranging from "great" to "terrible" when pairing. It's worked best for me when I'm paired with someone whose skills are complementary to mine. Concretely: I'm very much about rigour, type-safety, correctness; the person I have in mind here is a wizard at intuiting algorithms. The combination worked extremely well: the pairer generated algorithms, and I (at the keyboard) cast them into safe/correct forms.
However, when paired with someone who eclipses me in almost every dimension, I ended up feeling a bit bad that I was simply slowing us down; and conversely I've also experienced pairing with someone who I didn't feel was adding much to the enterprise, and it felt like I was coding through treacle (because the thoughts had to flow out to another person, rather than into the compiler).
In my experience, good pairs are really good, but also quite rare. You're looking for a certain kind of compatibility.
To answer your actual question: just try it! It's cheap to try, and you can find out very quickly if a certain pairing is not for you. (I would certainly start the exercise by making sure both parties know that "this pair isn't working out" is not a judgement on either party.)
I started Anki-ing everything. Previously, I've used Anki for very specific purposes (e.g. "learn the London Underground network" or "learn all the capitals of the world"). New decks this month, though, include "Jokes", "Legal Systems Very Different From Ours", "Tao Te Ching", and "Logical Induction". I'm pretty optimistic that "read something really worthwhile, Anki it up" is becoming a habit.
A formative experience in my attitude to magic was when I saw an excellent sleight-of-hand magician performing to my small group of friends (waiting in a line for an event). He was very convincing and great fun; but there was a moment in the middle of his series of tricks when my attention was caught by something else in the distance. When I looked back after five seconds of distraction, he was mid-trick; and I saw him matter-of-factly take a foam ball from his hand, put it into his pocket, and then open his hand to reveal no foam balls - to general astonishment. All his other tricks, before and after, I found completely convincing.
Accordingly, I grok that there's an entire art of doing incredibly obvious things in such a way that the viewer doesn't understand that something obvious has happened. That's one of the main things I want to learn from magic: how to perform trivial bullshit very convincingly (e.g. by knowing how to direct the viewer's attention).
Thanks for the tip about performing repeatedly to new groups. Now that you mention it, it's extremely obvious, but I don't think I'd have come up with that myself.
Thanks very much for this! I've written a lot of stuff on there (I'm the Patrick Stevens whose name is splatted all over the screenshot). I asked them a year ago (ish) whether I could have a data dump, and they said it was Too Difficult; and I didn't bother scraping it myself. I'm glad you actually went and did something about it!
On introductory non-standard analysis, Goldblatt's "Lectures on the hyperreals" from the Graduate Texts in Mathematics series. Goldblatt introduces the hyperreals using an ultrapower, then explores analysis and some rather complicated applications like Lebesgue measure.
Goldblatt is preferred to Robinson's "Non-standard analysis", which is highly in-depth about the specific logical constructions; Goldblatt doesn't waste too much time on that, but constructs a model, proves some stuff in it, then generalises quite early. Also preferred to Hurd and Loeb's "An introduction to non-standard real analysis", which I somehow just couldn't really get into. Its treatment of measure theory, for instance, is just much more difficult to understand than Goldblatt's.
True, though the decision of who is most cost-effective does remain for you to decide.
It's more of a tactic to make sure people don't think "hey, another crackpot organisation" if they haven't already heard about them. I'm hoping to raise GWWC to the level of "worth investigating for myself" in this post.
I do something similar. I consistently massively underestimate the inferential gaps when I'm talking about these things, and end up spending half an hour talking about tangential stuff the Sequences explain better and faster.
I'd frame it as "Nick Bostrom needs Jeeves. Are you Jeeves?" (After P.G. Wodehouse's Jeeves and Wooster.)