GuySrinivasan

Karma: 2,143

GuySrinivasan 15 Nov 2021 1:07 UTC
10 points
in reply to: aphyer’s comment on: Improving on the Karma System
I am a huge fan of tiered-complexity views on complex underlying systems. The description to new users would be:
- Ratings are a magic median-like combination of how users rated a post. Click through for more details...
- Displayed ratings are the median of how users have rated the post/comment. Smoothed. Weighted by how LessWrongy the rater has been. Your own rating will have more effect when your historical ratings are good predictions of how trusted moderators end up rating. Click through for more details...
- Sometimes mods will rate posts/comments, after careful reflection of how they want LessWrong in general to rate. When they do, everyone who previously rated will be awarded additional weight to their future votes if their ratings were similar to what the mod decided, or penalized with less future vote weight if their ratings were pretty far off. That’s how the weights are determined when aggregating people’s votes on comments. Of course, it’s more complicated than that. Folks were grandfathered in. New folks [behavior]. Mods who are regularly different than other mods and high-weight voters trigger investigation into whether they should be mods anymore, or whether everyone is getting something wrong. Multiple mod votes are a thing, as is voting similar to high-weight voters (?? maybe ?? is it ??), as is promoting high-weight voters to mods, as is etc etc. Click through for more details, including math...
- [treatise]
- [link to the documented code]

GuySrinivasan 14 Nov 2021 22:46 UTC
17 points
on: Improving on the Karma System
Very well thought out. I think the two biggest things missing from your analysis are:
- 5-star ratings have become corrupted in the wild. Small-time authors get legitimately angered when a fan rates their work as 4-stars on Amazon, because anything other than 5 stars is very damaging. We don’t want to port this behavior/intuition to LW, but by default that’s what we’d do. jimrandomh mentions this in their comment. I don’t know how to overcome this problem while retaining 5-star ratings.
- Users like to “correct” a post/comment’s rating. Personally I hate this behavior, but after ranting about it several times over the years I’ve learned that I don’t represent everyone. :D So if they see a comment with average 2-stars, which they think should be 3.5, they will not want to rate it 3.5. Instead they will want to rate it 5.0 to “make up for” the other “wrongheaded” views. Maybe one way to overcome this problem is to allow a lightweight way to say “I think this is 3.5. I want to use my QJR to fix the current rating so I’m voting 5.0. Please increase the chance a mod sees this, double down on my bet, and maybe even change my rating back to 3.5 if the post becomes 3.5?? And also if a mod rates a post maybe this should drastically reduce the effective QJR of people contradicting the mod… or something, if a mod says 3.5 and the weighted avg is still only 2.5 I won’t be happy”.
What links here?
- Improving on the Karma System by Raelifin (14 Nov 2021 18:01 UTC; 88 points)

GuySrinivasan 13 Nov 2021 7:29 UTC
2 points
on: Average probabilities, not log odds
My heuristic for deciding what heuristic to use, when you’re going to do something quick-n-dirty: figure out what quantity you’re actually interested in, use means in its natural scale for point estimates, and transform back to your inputs.
How does this apply to some examples?
In your post, you’re talking quite a lot about bets. To a first approximation marginal utility is linear in marginal wealth, so usually this means the quantity we are actually interested in is linear in probability, and the correct heuristic is “arithmetic mean” (of probability).
In SimonM’s comment, we’re talking about probabilities directly. Forecasting. Usually that means what we care about is calibration or a proper scoring rule, so the natural scale is [0,1] or log-odds. Now the correct heuristic is “arithmetic mean” (of log-odds of probability).
What about your difficult examples?
Expert 1 gives you a probability distribution 50% A, 25% B, 25% C, expert 2 gives you a probability distribution 25% A, 50% B, 25% C, and expert 3 gives you a probability distribution 25% A, 25% B, 50% C.
We’re already doing quick-n-dirty things rather than anything rigorous. What I usually do when I want constraints and my summary statistics don’t satisfy them is to just go ahead and normalize, after which of course we get ¹⁄₃, ¹⁄₃, ¹⁄₃ by symmetry after any estimation.
that depends on the prior
Once we’re no longer just doing something quick-n-dirty, all, uh, bets, are off. But my heuristic is still to transform to the domain where what you care about is linear, do your somewhat-more-sophisticated dirty work, and transform back to get your point estimate.

GuySrinivasan 29 Oct 2021 19:14 UTC
2 points
in reply to: warrenjordan’s comment on: How to generate idea/solutions to solve a problem?
This was all right -ish, replacing the very physical principles. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2432096 Not inspired but fine to try.
This is a copy of a format I used with my team: https://docs.google.com/document/d/1guFRreqNUS6JUauBpWumCYoWilWorbkcnpBIg-yVJKQ/edit?usp=sharing It worked decently well for generating very new ideas.
I’m not sure I recommend either of them, but they’re fine.

GuySrinivasan 29 Oct 2021 4:07 UTC
2 points
on: How to generate idea/solutions to solve a problem?
TRIZ—Wikipedia
This is a pointer to a large categorization of techniques used for physical innovation. Elsewhere on the internet it has been adapted for software innovation. Try picking techniques randomly to see if you can creatively interpret them to provide a novel solution to your problem, repeat until inspired.

GuySrinivasan 28 Oct 2021 17:56 UTC
6 points
on: Recommending Understand, a Game about Discerning the Rules
I have been independently recommended this game by three sources within 24 hours, so yeah I’ll play it.
I highly recommend well-facilitated Zendo for “training” and fun; this seems like it probably hits the same notes. Stretch goal: be as efficient as possible, attempting to learn the sorts of rules the creator would make and using those priors to win unreasonably effectively. I have no idea if that will work in Understanding, having not yet played, but if several people play Zendo together and get good at it (and don’t try to change their “GMing” based on the fact that people will try to model the creators), you can see some absolute magic happen as folks guess rules correctly with ridiculously small amounts of data.

GuySrinivasan 17 Oct 2021 7:53 UTC
8 points
in reply to: dxu’s comment on: Lies, Damn Lies, and Fabricated Options
Yes, and nowadays I do recognize that I’m really choosing between going to sleep now or playing some number, not at all necessarily just one, “one more turn”s, and gauge what that distribution looks like and then choose.
#TOOLS
Recognize that your decisions are evidence of what sorts of decisions you’re likely to make when deciding whether to do something; apply multipliers as needed.

GuySrinivasan 17 Oct 2021 5:22 UTC
15 points
on: Lies, Damn Lies, and Fabricated Options
#EXAMPLE
I always try to go to bed at right around 11 PM. I finish my game at 10:45 PM. I can stop playing now and go to sleep, or I can play one more 20-minute game and then go to sleep. The latter is pretty clearly better, games are short and I’m not all that tired. Three hours later I go to sleep.

GuySrinivasan 15 Oct 2021 19:25 UTC
5 points
in reply to: abstractapplic’s comment on: Creating a truly formidable Art
As a most-of-the-time participant, I can definitely agree. What’s going on behind the scenes, at least for me, is much more than what you’ll see in the comments. Most frequently practiced, I think:
- Noticing Confusion—I will often see a curve or scatterplot that inspires a feeling of that’s different. It’s great to get practice noticing that over and over, and then digging into how well that intuition tracks truth.
- Curiosity—almost of course. But it’s good practice at finding things to be curious about. And then also good practice knowing that you won’t actually spend 20 hours on this project so evaluating: I’m curious; is it worth my time to follow up on this, or should I pursue even more promising avenues?

GuySrinivasan 7 Oct 2021 1:01 UTC
5 points
in reply to: elspood’s comment on: 2021 Darwin Game—Everywhere Else

False. You could make it true by deciding to count species that won in multiple biomes once for each biome they won in, but I’m not going to do that.

I’m giving myself credit for a much better score than that. ;)

Edit: as per elspood’s simulations, I am revoking all of my supposed credit.

GuySrinivasan 5 Oct 2021 21:52 UTC
2 points
in reply to: GuySrinivasan’s comment on: D&D.Sci 4th Edition: League of Defenders of the Storm Evaluation & Ruleset
Feedback: I liked this one a lot, in theory. I just found myself with less time than I wanted to actually engage. It had a latent structure that was definitely discoverable, but also definitely plenty obfuscated. Even with time constraints the barrier-to-entry was low enough that I could get something nice in pretty quickly. And the differences between individual, team, vs-that-team, and vs-any-team were interesting to mull over.

GuySrinivasan 5 Oct 2021 20:18 UTC
2 points
on: D&D.Sci 4th Edition: League of Defenders of the Storm Evaluation & Ruleset
I was fairly surprised my PvE team did so well. Why did my heuristics basically work?
How well do individuals do? Multiply. How well do teams do, relative to the product of their members? How well do individuals on a team do against the known opponents, pairwise? Okay try all teams, combine their “vs opponents” scores with “team coherence” scores, pick the max.
Finding how well teams did relative to how their individuals did let me accidentally pick just from balanced teams; asking how well individuals did against the known opponents got rid of Null and biased towards the 6s, 5s, and 1s. I think Maelstrom (Water 3) probably got on there because the team did really well in the data despite Maelstrom’s inclusion, which looked the same to me as Maelstrom synergizing well with the rest of the team.
I weighted “individuals vs opposing team” equal to “team synergy”; I wonder if I had left the weights as each individual = each other individual = team synergy, whether that would have correctly axed 2-4s while still axing Null? [tries] Nope, that lets Null in quickly.

GuySrinivasan 5 Oct 2021 1:49 UTC
6 points
in reply to: Tofly’s comment on: 2021 Darwin Game—Tundra
IIRC every non-eaten organism gets 80% of its size, plus its nutrition, x10, in lottery tickets. Every lottery ticket has a 1/size/10 of being a new organism in the next generation.

GuySrinivasan 5 Oct 2021 0:39 UTC
2 points
in reply to: lsusr’s comment on: 2021 Darwin Game—Tundra
Thanks for looking! I am satisfied that the smol thing was viable. :D

GuySrinivasan 5 Oct 2021 0:17 UTC
3 points
in reply to: aphyer’s comment on: 2021 Darwin Game—Tundra
Oh yeah definitely but I was wondering why its name didn’t appear—that I forgot I changed it, or that it actually isn’t viable, or both.

GuySrinivasan 4 Oct 2021 23:56 UTC
3 points
on: 2021 Darwin Game—Tundra
Hrm, according to my notes I should have submitted a
LivingTrenchcoat
eats seeds
eats lichen
cold resistant
Tundra
herbivore, but it is not on the list of viable herbivores. Maybe I changed my mind and submitted something else?
From what I remember, an organism like that is technically viable even if it has 5.1 size by adding a point of e.g. weapons, right? If there was no competition, then one of it would eat the seeds, getting 72 entries into a 1.96% lottery, and N-1 of it would eat lichen, getting 48 entries into a 1.96% lottery, yielding an equilibrium population of 8 organisms, which might be large enough to avoid randomly fluctuating to zero. Or did I get the mechanics wrong?

GuySrinivasan 4 Oct 2021 17:22 UTC
2 points
in reply to: AllAmericanBreakfast’s comment on: The LessWrong Team is now Lightcone Infrastructure, come work with us!
US Senators and House representatives make a salary of $174,000/year. Doctors make $200,000/year
Do we think that the marginal increases in a Senator-making-174k’s wealth over time compared to a Doctor-making-200k look much like the marginal increases in a Doctor-making-174k’s wealth compared to a Doctor-making-200k?

GuySrinivasan 2 Oct 2021 5:22 UTC
4 points
on: D&D.Sci 4th Edition: League of Defenders of the Storm
For my upcoming Cloud Liquid Gaming team, I choose:
Blaze Boy
Greenery Giant
Maelstrom Mage
Phoenix Paladin
Rock-n-Roll Ranger
I’m not super happy about it, because I know nothing of how the matchups actually work. This is just a product of a kinda ad-hoc value function. How well do individuals do? Multiply. How well do teams do, relative to the product of their members? How well do individuals on a team do against the known opponents, pairwise? Okay try all teams, combine their “vs opponents” scores with “team coherence” scores, pick the max.
For my PvP team, I choose to PM our benevolent creator.

GuySrinivasan 27 Sep 2021 17:38 UTC
4 points
on: AISC5 Retrospective: Mechanisms for Avoiding Tragedy of the Commons in Common Pool Resource Problems
What would you do differently? What general procedural changes would have resulted in a better outcome? What procedures worked well such that you’d recommend they stay largely the same for next time?

GuySrinivasan 24 Sep 2021 22:58 UTC
6 points
on: The 2021 Less Wrong Darwin Game
I have submitted 10 entries. At least two are guaranteed to lose. At least two are elephants, because why not, c’mon. I have high hopes for two. I have extremely high hopes for two more but not because I think they’re likely to survive. The last two are unremarkable.
Thanks for putting this together Isusr!