Great work with this newsletter, keep it up. It’s by far the best forecasting digest I’ve seen!
The major friction for me is that some of the formatting makes it feel overwhelming. Maybe use bold headings instead of bullet points for each new entry? Not sure.
jacobjacob
Karma: 2,775
Sunday August 9, 1pm (PDT) — talks by elityre, jacobjacob, Ruby
This post was presented as a brief talk during the LessWrong curated talks event on July 19.
Here is a transcription of the Q&A session following the talk.---
mr-hire: I’m curious if you’ve thought about what training this at scale would look like. When I was younger, I remember being trained a little bit on how to keep secrets, just with friends and stuff, but what would it look like if you had ways to train more deliberately?Raemon: I know some people who have actually thought about this a lot and have entire world views and practices oriented around it. For me, I noticed that I wasn’t good at keeping secrets, which was at times a problem for my relationships. So I actively thought about it for a while and increased my ability.
So the short answer is: do lots of thinking and practicing. To answer your question about training at scale, I have two ways that you can teach this.
First, you can teach people trigger-action plans for dealing with secrets. A key piece of such trigger-action plans is the ability to notice when you’re in a conversation that might create sensitive information or bear on information you already have that is confidential.
Being able to notice that is much of the battle, and then having follow up actions like “slow down and think before you say each new thing”, or “deflect the conversation in a new direction”. So that’s a skill I came up with on my own.
Another skill someone else pointed out, for situations where you find out confidential information about an organization, is to have different mental models for storing public and private information and keeping track of them differently.
And then, when having a conversation, live inside one of the two different models. This doesn’t work for me yet. I haven’t really tried to do it, but it’s a thing that seems to work for at least two people that I know of.
---
johnswentworth: So given that a big part of a secret is the fact that a secret exists, how do you ever trust that someone can keep secrets if they tell you that they have kept secrets before?
Raemon: Well, there technically was a whole second half of this talk. I have multiple blog posts coming that deal with a lot of like, “Hmm, this topic sucks. What do we do?” One of the problems is that people don’t even necessarily mean the same thing by secret. Sometimes it means they just don’t bring it up. Sometimes, it means they do not reveal any Bayesian information that can possibly inform people that the secret even exists.
And sometimes, it’s just like Alice saying, “Bob, just don’t tell Charlie. You can tell Dave. Just make sure it doesn’t get back to Charlie.” So what I try to do, noticing that we’re getting into a sensitive situation (or ideally if I have detected that the person I am speaking with has an ongoing, long-term relationship with the person where secrets are likely to come up), I try to have a meta-conversation about what secrecy means to them, discussing what the various parameters of secrecy are and which ones are the most important to them, before a specific secret comes up.
Sometimes, the duration of a secret matters and you need to keep it to your grave. Other times it could simply be a controversial thing happening in the next three months and I need you to keep it quiet until it is over.
Generally, I don’t think it is actually tractable to not give Bayesian information that you have any secrets. I think a slightly better equilibrium is where there’s some glomarization of “I can neither confirm nor deny that I have secrets relating to this thing”, and you just always say that whenever this conversation comes up. And then, you get into the meta a bit. A practice that I like to do while having this meta-conversation is avoid making eye contact. Not looking at each other while having the conversation ensures our micro-expressions aren’t betraying any information until we have built up a little bit of trust.
Ben Pace: I am also much more likely to accept a secret if it’s on a three-month scale rather than a four-year or permanent time scale.
---ricraz: I guess if we take a Hansonian perspective and we say, “Is keeping secrets really actually about keeping secrets...?”
Raemon: Oh my.
ricraz: A lot of people around here have quite high scrupulosity. I am fairly low on this. A lot of the time when I say to somebody, “Please keep it a secret,” it’s just mostly a social nicety. Maybe it reveals some information about them if they don’t keep it secret, and maybe I’d feel a bit annoyed… But it feels much more like it’s a standard part of the interaction I’m having with them. Maybe I’m ticking a box or something.
It feels fairly rare that I’m telling somebody information that it’s crucial for them to keep secret. And so, I wonder if actually outside a corporate context like, “Please don’t give away our secret product plan,” and stuff like that, how relevant is this actually in most social interactions?
Raemon: So another pet peeve is that, most of the time, I think people are doing something like what you just said—keep it on the downlow. Keep R0 of the secret <1 if you can.
The tricky bit is that’s what people need most of the time, but if the secret ends up causing them more damage than they expected, then they’re like, “No, you said you would keep it a secret,” and then retroactively judge you harsher than you might have assumed.
So one of the key things I want is transparency about what level of secret we’re talking about. So if people start telling me anything confidential, one of the first things I say is, “FYI, the default thing I am going to offer you is, I will try a little bit to keep R0 of the secret less than one, but I’m not going to try that hard. And if you want more than that, you are asking me a favor and I’m checking if you want to ask me that favor.”
And most of the time, they don’t actually care about that, the more high-level secret. I do think in the rational sphere and the ecosystem, there’s a lot of blurry lines between, “Ah, I’m just commenting about my friend,” and, “Oh, my comments about my friend actually directly inform whether some other person’s going to give that friend a grant,” which makes this all a bit trickier. So the main thing I want to get out of this is to have common knowledge of what the default norms are and of what favors you’re actually asking of people.
jacobjacob: So I heard you ask, Richard, if this really matters or if it matters only if you know some important intellectual property due to your work? And I think there’s a thing where a small secret with someone might be fine, but if you need to keep track of 10 or 100 small secrets across a large number of people, then whenever you say a sentence, you have to run this working-memory process which checks the sentence against all these secrets you have to keep track of, in various weird interactions. And you get to a point where this messes up your ability to have conversations or to think clearly, even if the individual secrets themselves are not super important.
Ben Pace: Yeah. I don’t enjoy how much obfuscating some people have to do in conversations with me when we’re trying to talk about the same thing, but they can’t say anything because it would betray that they know some information that I also know. I was expecting you to say something like “it’s also hard to tell if someone’s good at keeping secrets, and so, being able to keep secrets about small things is often the only way you can even check. If someone only keeps secrets about massive important things, it’s often hard to know about those things, because you’re not privy to them”. And so, it helps the person doing the right thing in smaller iterated games.
---George Lanetz: When I thought about this problem a while back, I decided it’s impossible to really keep my privacy. So I committed to a life that won’t require that. From your talks with other people, how often did you find that they actually care about their privacy? And how many people do that if they are not presidents or something like that?
Raemon: I run into social circles that are one or two steps removed from a lot of EA grants, and this comes up a lot there, where there’s informal situations that turn out to have fairly strong financial stakes. So I think I’ve probably interacted with 10 people, plus or minus a few, where some kind of common understanding of what secrecy meant mattered.
Ben Pace: I also have really strong feelings about the things George said. I think Vitalik Buterin has said something relevant here. Namely that the less privacy you have, the more Hansonian things get: the more signaling that you’re doing on all of your supposedly “private” conversations. I think the internet has generally done this a lot. People used to be able to have informal bonds of trust, but then suddenly it gets recorded and put online.
And then, a million people see it and you’re like, “I can’t live my life where a million people are watching all of my interactions, all of my moves, checking whether or not they’re good.” It would talk more time to make this point in full.
Gah, thanks, I’m too sloppy with these announcements. Will improve!
Sunday August 2, 12pm (PDT) — talks by jimrandomh, johnswenthworth, Daniel Filan, Jacobian
$1000 bounty for OpenAI to show whether GPT3 was “deliberately” pretending to be stupider than it is
You’ll have to file a request with the LessWrong Customer Service Team, with office hours 9am to 5pm Pacific time, Monday through Thursday, and a lunch break at 12-1pm.
Processing times for requests like this are usually around 4-5 days.
Thank you for your patience.
Woop!
Lol, fixed, thanks.
Lessons on AI Takeover from the conquistadors
Thanks, fixed.
Meta-preferences are weird
Sunday July 19, 1pm (PDT) — talks by Raemon, ricraz, mr-hire, Jameson Quinn
Thanks! David did most of the work. :)
Mazes and Duality
Fixed, thanks!
Thanks for letting us know lionhearted. You’re welcome back another week if you have a talk you feel better about! :)
To all attendees—the event will go ahead as planned with a replacement speaker (me!).