Posts
Comments
It can’t get through metal, but this fentanyl-detecting machine can detect fentanyl in packages:
You might want to rethink the definition of "hardest problem" because I wouldn't expect the IMO committee to care particularly much that their problems are hard for machines as well as humans.
For example, if you look at the 2005 IMO, but suppose that problem 6 was geometry, so that the problem 3 was the "hardest", then you're in trouble. Problem 3 has a one-line proof "by sum_of_squares", assuming that someone has written a sum-of-squares tactic in time and that the sum-of-squares certificate is small enough to find within the time limit.
Alternatively, the IMO committee might set a problem without realizing that it's close to a special case of something already in mathlib.
As defined, I would put higher probability on solving the "hardest problem" than getting a gold medal and would bet on that.
Data from NSW, Australia. Delta was stable at 150-300 cases/day, so recent spike is likely omicron. 93% of people aged 16 or over are double-vaccinated. https://covidlive.com.au/report/daily-cases/nsw
Re. the NBA study, say I go get my antibody levels tested (which is easy here in Australia), do we know what counts as a high result?
Reusable respirator guide
- In scope: information about half-facepiece reusable respirators. Out of scope: whether you should wear a mask, etc.
- Reasons to prefer a reusable respirator over a disposable one: they're (i) more comfortable and (ii) prosocial, preserving disposable masks for medical personnel.
- You need to buy two components: the mask and filters.
- The mask. The wirecutter recommends 3M's 6500 series. I have the next model up, from the 7500 series. The 7500 comes in three sizes, small (7501), medium (7502) and large (7503). I think the idea of the sizing is: medium fits most, but there are also small and large (evidence: some stores only stock mediums). If you are familiar with the size of my head, I'm a medium. If outside the US, say in Australia, 3M masks might be hard to find, so you might try other brands like Sundstrom.
- To disenfect the mask periodically, you can soak it in disenfectant.
- The filters. These come in a vast number of options and it's all quite confusing. A good source of information is this 3M guide. For viruses, we're interested in particulate filtration. Summarizing page 3 of the 3M guide, a filter is described by a minimum efficiency (95%; 99%; or 99.97% \approx 100%) and a letter that describes whether the filter efficiency degrades in the presence of oil-based aerosols (N, R, P), which is not relevant for non-industrial use. You combine a letter with the minimum efficiency, e.g., N95 = minimum efficiency 95%, not suitable for use in the presence of oil-based aerosols, or P100 = minimum efficiency 99.97%, will not degrade in the presence of oil-based aerosols. Some brands will use European nomenclature: P2 = P95, P3 = P100.
- A filter with higher minimum efficiency will be harder to breathe through. That said, I personally find it quite comfortable to breathe through a P100 filter (e.g., I find it easier to breathe through a respirator with a P100 filter than breathing through an N95 disposable mask), so I would recommend using a 99.97% efficiency filter. Maybe the trade-off between filter quality and how easy it is to breathe will be different for others.
- Filters are sold in pairs. In terms of specific 3M P100 filter recommendations, the 2291 is the best (it's more expensive but designed to be easier to breathe through), then the 2091 is next best, and the 2296, 2297, 2096, and 2097 are fine too -- they have an extra layer of carbon filter for relief against nuisance levels of certain gases, which you don't need for virus filtration (but if you're walking around a busy city with lots of traffic, the 2297 or 2097 might be preferable). All these are fine though so get whatever is available. All of them are compatible with 3M 6500 or 7500 series masks.
- I'm unsure how long a particulate filter lasts. I think the answer might be be months if you're not working in dusty environments. They become more efficient over time but harder to breath through. This 3M guide doesn't give a timeframe to expiry, it just says to replace them when they become dirty or difficult to breathe through.
- Does the surface of a filter become contaminated? I don't know. Presumably? Maybe don't touch it. If you get a Sundstrom respirator, the filter is entirely encased in plastic, so you could regularly disenfect the parts of the mask that you can touch.
- In the case of a shortage, there are a number of combination cartridges that filter against various gases as well as particles. For example, all those on page 18 of the first 3M guide. These will work for particulate filtration, you're just also paying for the gas filtration (in money, heaviness, and possibly in that it is more difficult to breathe through them).
- Be wary of counterfeit filters. These are considerably cheaper.
Thank you very much for producing these. As someone who's rather time poor but trying to become more informed, they are very helpful.
Here is some alternative code for building an HN clone: https://github.com/jcs/lobsters (see https://lobste.rs/about for differences to HN).
It's also one night before full moon (which is at 4:50am on June 15), which should make the sky quite bright.
On a related note, consider what the moon looks like one night before it's full. Would you describe this as "over three-quarters full"? While that's technically correct, I wouldn't. I'd maybe describe a June 11-12 moon as "over three-quarters full" but I'd say a June 13-14 moon is "almost full". So we should up the probability that we're in a story/simulation/mirror.
Observation: If the purpose of this exercise is to run an AI box experiment, with EY as gatekeeper and the internet hivemind as the AI, then the ability to speak in parseltongue is problematic: It appears to make the game easier for the AI, thereby preventing the results from being generalized to a standard AI box experiment.
So why did Eliezer include the parseltongue constraint?
Maybe parseltongue is meant to introduce the concept of provability in a way that everyone can understand. To speak in parseltongue in real life, you just speak in logic statements and supply a proof with any statement you make. It seems reasonable (modulo computational complexity and provability concerns) for an AI to be able and/or required to supply proofs in an AI box experiment and parseltongue enables that in version of the game in the story.
I don't understand the constraint to speak only in parseltongue. Is that there to force us to focus on a solution set that is somehow of interest for friendly AI research?
Here's a flawed solution, but maybe someone can fix it.
Harry performs partial transfiguration on his brain, to transform it into a state where he thinks that he's booby-trapped the universe (for example, by transfiguring some strangelets along with a confinement field that will expire before the strangelets do). Then he just explains honestly to Voldemort why the universe will end if he dies.
The mirror and efficient simulation
Until the mirror appeared, the HPMOR universe could be simulated efficiently, at least as far as we knew. Time travel is limited to a six-hour cache; you can't transfigure arbitrary things, and Harry's attempts to use time travel to solve computational problems failed. This is likely to be deliberate.
So, how does the mirror exist? According to the inscription on the back, the mirror shows the actor's coherent extrapolated volition (CEV). Is this possible to compute efficiently from an actor's source code? I would guess not. (Is this right? Is there a hardness proof?)
Here are some possibilities to preserve efficient simulation:
- The mirror is a trick and doesn't do what it says it does
- Perhaps the NPCs are programmed to have known CEVs
- Causality goes in the other direction. Computing the CEV from the source is hard, but constraining an actor to act in a way consistent with their CEV might not be. Perhaps everyone's CEV is set when they first look in the mirror, and then they are constrained to act in accordance with this. (This is somewhat akin to the way of avoiding paradoxes in a quantum universe with time travel, whereby if you go back in time and shoot your grandmother, the bullet diffracts around her.)
- Maybe the previous item is only true of the NPCs. Is the mirror blank at the end of 110 because Harry is a PC (alternatively: simulating him is the purpose of simulation) and computing his CEV is too hard?
This will be the last meetup at my apartment.
I sadly can't make this one.
We plan to discuss metaethics. It might be useful to narrow this down.
Yes. My house, but I won't be there until late.
And upvote Richard's comment too, to thank him for hosting!
I can't make it to this one.
The workshop was invigorating because nearly everyone seemed confused.
Paleo + intermittent fasting + read Kevin's posts on supplements.
Let me expand on my comment from the Hacker News thread.
I went to the July workshop. I think it was probably the most useful week of my life in terms of exposure to things I could be doing to be more productive and effective. Since then, I've mainly been trying to incorporate the low-hanging fruit---the obviously good simple ideas---into my life. Some examples:
- At work, I realized I wasn't doing anywhere near enough planning. My employees were spending time on the wrong things because I hadn't planned things out in enough detail to make it clear what was the most important thing to do next. I fixed this immediately after the camp.
- Again at work, I now document most discussions as they happen. It slows down the discussion a bit, but it's a net win because next time you can just have the delta of the discussion and avoid rehashing things.
- Val explained to me how he changed his diet in order to decouple the relationship between his energy levels and when he last ate. I hadn't actually realized such a thing was possible. It also works for me.
- I realized that it's a waste of time to learn things/watch video lectures and not put what I want to remember into anki. I make the cards while I watch the lectures, and have trained myself to review them on my phone when waiting in lines, walking down the street etc.
- I'm training myself to no longer think things like "I should go to the gym tomorrow." Instead I decide to go at X o'clock and set an alarm on my phone to remind me. (I apparently haven't got this one completely down yet, because yesterday I thought "I should go to the gym tomorrow" and didn't catch myself: it's now 11pm and I never went.)
I feel like I'm still a beginner at instrumental rationality and have an enormous amount to work on (I haven't even started trying to adopt most of the actual curriculum!). Since attending, I've been getting considerably more work done: my brother, who is also my business partner, has noticed too.
When I got back I was super busy and have been really bad at staying in touch with the instructors and other attendees, but there are a few people I met at the camp who I would really like to have more contact with.
We'll organize something. I'll talk to people about it at this coming meetup. There are some other non-local Less Wrong people in Melbourne right now too.
We could do! Will you be here? We'll at least repeat this meetup on December 21.
Notes from discussion about finding better metrics to measure CfAR's effectiveness
[Discussion about what CfAR is trying to teach.]
Have to avoid CfAR just measuring how much people are paying attention in their courses
Beforehand: write down goals
Six months later: measure them against their pre-camp goals
Control group
Just attending a camp might make it feel like you're making more progress towards your goals.
Some goals hard to measure
How much throughput are they expecting? 3 camps/month.
Measuring income: a lot of people might decide during the camp to change career, or devote their life to effective charity, so income may well drop when your rationality goes up.
Also income has a long lag before interventions take effect
epicwin: gamify your life.
We're going to stop proposing solutions for a while
Suggestion to generate ideas individually.
Desirable qualities: low stat noise, objective, quantitative etc.
What sorts of measures: e.g. based on other's opinions, performance at a task, other classes
Could talk about what goals people have? Then we could think about how to measure those particular goals. Money, happier, healthier, etc.
Whose perspective are we taking? The consumer of the minicamps.
In self-help field, typically sell books or lecture tours based on how good your stuff sounds. Most of it doesn't come with metrics. If there's some of that works (e.g. GTD), then CfAR should be teaching it.
[Tangent about quality of scientific literature.] Problem with finding material to teach by looking for stuff whose effectiveness has already been well measured in the literature is that you're setting the bar too high. We don't need to published, so can make do with messier non-laboratory non-ethics-committee-approved measures.
Come up with a test you like. Try to teach to that.
Break up to discuss:
Desirable qualities of metrics
Types of test
Goals that people might have
[We break into groups]
I'm with yurifury (in hammock) + patrick (in bean bag)
- Types of tests
Maybe have an interactive test that heaps you make decisions (like it includes a checklist of things to think about and biases to avoid) and it would generate metrics along the way. An iterative process, so you could track how people's goals change. And if it's a process to help people, people more likely to use it. Blurring line between self-help tool and pure metric. Or not a tool just for significant decisions, but a weekly thing. People's goals might change from month to month and a tool that let you say "i'm doing well on this one" "i don't care about this one anymore" would be helpful and would generate metrics.
An agile version opposed to the current waterfall 1-year-apart surveys pre- and post-minicamp.
Difference between quantified selfers, who measure things to detect correlations, and people who just measure things they're trying to improve.
They should run a course specifically for quantified selfers. Gwern could be the Matt of the course. Then the could observe what changes about them and they might get less obsessive people to measure those specific things.
- Goals
Big problem with motivation: how to get people to take the tests. 1. Have measurements be side-effects of something otherwise useful. 2. Social pressure.
X, what would you like to change about yourself? Answer: spend more time reading. Akrasia.
People who talk regularly following minicamp (already) could evaluate each other. People at meetups could evaluate each other. 5 minute quiz at meetups.
Have an award for the person who changes the most. <- hard to make honest
We seem to be generating more ideas/unit time in our small group, than the group as a whole. Same for other groups? Yes.
We'll focus on goals that are easiest to measure. Health. Measures like resting heart rate not BMI Facebook friends/klout. Happiness Grades <- not so good Memory/IQish tests
[We decide that the notes above are a really good example of how useless a 12-person group discussion can be. AVOID AVOID AVOID]
The table was observed to be red-shifted.
Good suggestions from other groups:
Possible tests:
Idealogical Turing test as a measure. People learning to model each other. Tests things including: are you able to model other people well, do you understand other people's ideas. rather labour intensive
Search skill: able to use google. measure how well people are able to find information (with time constraints). Apparently google has a daily challenge.
Get people to do a three minute video sales pitch.
Game playing. Useful metrics in how well you play poker, Wits and Wagers.
Might be worth having preprepared list of goals that are easy to achieve and/or measure and have people select goals from that list
Look at course material and try to come up with tests inspired by course material, but being sure to ask "Why are we delivering this material?" too
[Tangent/derailment]
[Splitting up again, given what we learned. Other group is dancing.]
Summary of what we want to tell CfAR:
Summarize desirable characteristics of goals. Need as common knowledge.
Directions to go in towards gathering metrics
What we learned about how to have this discussion. Talking about goals seemed too general and not useful; discussion about transferable skills seemed more useful. Things that people wanted that came up frequently: Calibrated, appraise evidence, socially competent, ability to persuade others, motivation esp. across different time scales.
Some suggestions of metrics (filter above)
Also we want to post in the meetup organizers group what we learned about group discussions. Benefit of small groups is it's okay to interrupt.
[At this point my battery about to run out.]
Quixey is incredibly successful. Also, LessWrong is still young. Give it time! There may be a bunch of startups out there we haven't heard of yet. For example, I'm doing a startup with 3 other LWers, but we need a little longer before we're successful ;-)
I can come and I'd be happy to do the exercise I suggested last month
http://lesswrong.com/lw/bym/meetup_melbourne_practical_rationality/6gcz
but I'm thinking we'd probably prefer to hear from James and Scott about the minicamp (if they're willing).
Edit: Let's save the updating activity below for another night or for next month, when Aubrey de Grey isn't speaking.
We decided that we'd take it in turns to prepare activities for the practical rationality meetups, and I volunteered for this meetup.
Here's the plan. I have a bunch of trivia questions with numerical answers. For each question: Everyone records either 50% or 90% confidence intervals. Open discussion. Everyone may alter their intervals. Record initial and updated calibration. The aim is to test our calibration (is it true that approximately half your answers lie within your 50% confidence intervals?), ability to update (are your post-discussion guesses better?), and ability to efficiently persuade a group about something you are relatively certain about.
I propose we do both the initial estimation and discussion under moderate time pressure (i) because when you need to update in real life you often don't have much time, and (ii) so the whole thing doesn't take too long. We'll have to work out exactly how much time that should be but let's start with say 30 seconds to write down intervals and 1 min for discussion.
I'll reveal each answer immediately after and we can then take as long as we need to discuss how we're doing, so hopefully we get better as we go along. We'll be able to measure this, so at the end of the night we'll have data that will tell us if the activity was useful or not.
Suggestions are welcome.
You just missed it.
It's not a zero-sum game!
Sadly I can't make it tonight. I find myself in the wrong timezone and I'll be asleep.
Thanks for organizing this Adam, I enjoyed it!
There was a strong consensus that we wanted to keep meeting, and I made a google group for us to plan activities, discuss what sorts of things we might like to do in the future, let each other know of interesting events, and so on.
At the moment the group is semi-private, meaning that only members can view content but anyone can join. There is no moderation. We can change this if there is a consensus to do so or as it becomes necessary.
I'll make AdamBell, Patrick (who organized the previous meetups) and matt admins, unless they object.
If you're in Melbourne, please join and perhaps introduce yourself!
I'll come but will probably be a bit late.
Many jobs, including almost all of those that people would do on a working holiday, have Award rates higher than the minimum wage. Effectively, in Australia the minimum wage depends on the job. $21 is probably the minimum allowed by whatever Award governs shokwave's employment, either the Fast Food Industry Award or KFC might have their own enterprise agreement.
If anyone cares, the place to learn about this is here.
I'll come.
I agree, start with no limit hold'em because there's an awful lot of good learning material about it and the games at low limits are pretty good, but at some point consider switching to pot limit omaha.
I think most online sites are 18+.
Most people use databases and heads-up displays, but to calculate and present statistics about your own and your opponents' play, not to calculate odds (calculating odds is easy). I like Poker Tracker.
Read some of the books published by Two Plus Two for solid beginner information that's mostly a little out of date; then sign up at a video training site (I like Deuces Cracked) for up-to-date information; finally go, e.g., here and accept one of the offers where they give you free money to play with and then use their money to practice at 1 cent/2 cent games.
Also, maybe ask this question at the Two Plus Two forums for a better response.
Go ahead! But it's hard.
I hear that everywhere too. It's a selection effect: most of the population aren't smart and rational enough to be long-term winning players and it's these people you hear complaining, while the good players go on quietly winning.
It's definitely true that the games are getting tougher every year, because the community is learning to play better, so the threshold of ability you need to be a winning player is constantly increasing. But it's not that high yet.
Now let's talk about your two bugbears, bots and collusion.
1. Bots
You never ever have to worry about bots. The goal in poker is to seek out and play against bad players, while tolerating the presence of good players. It's completely irrelevant whether these players are controlled by humans, machines, or some combination. (In practice, except possibly for heads-up limit hold'em, good players are still better than the best bots published in the academic literature anyway.)
2. Collusion
This is something you have to worry about, but in practice it's not that big a deal, especially if you play at low limits, where it's not going to be worth the bother for competent players to collude. There have been only a handful of times when I've suspected collusion online, in which case, the obvious response was to stop playing against those players. Sometimes collusion can be detected statistically, but if some collusion does go undetected, as long as you're winning, what does it matter?
If you're truly smart, truly rational, and with the goal function you describe in your post, an obvious answer is to play poker on the internet. But beware: if it turns out you're not actually as rational as most of us on Less Wrong think we are, it probably won't work out.
Thank you for organising this. I'll come. It's a good choice of venue.
Surely this is illegal in the US. First you're acting like a bank by taking deposits and paying interest, and probably you have to actually be a bank to do that; second you're setting yourself up as a bookmaker.
Design suggestion:
All the meta stuff associated with a comment, viz.
Posted by: Kaj_Sotala 02 March 2009 09:32:50AM 2 points Vote up | Vote down | Permalink | Parent | Report | Reply
takes up a lot of space and impedes readability of the discussion. Can all this stuff be made smaller and less prominent (maybe more like it is on Hacker News) and perhaps some of the links only be visible when you're in the actual comment's thread (like the "Flag" feature is on Hacker News)? (Also we don't really need to know the exact second that a post is made.)