Reading through Backdoors as an analogy for deceptive alignment prompted me to think about a LW feature I might be interested in. I don't have much math background, and have always found it very effortful to parse math-heavy posts. I expect there are other people in a similar boat.

In modern programming IDEs it's common to have hoverovers for functions and variables, and I think it's sort of crazy that we don't have that for math. So, I'm considering a LessWrong feature that:

• takes in a post (i.e. when you save or go to publish a draft)
• identifies the LaTeX terms in the post
• creates a glossary for what each term means. (This should probably require confirmation by the author)
• makes a hoverover for each term so when you mouseover it reminds you.

On "Backdoors", I asked the LessWrong-integrated LLM: "what do the Latex terms here mean"?

It replied :

The LaTeX symbols in this passage represent mathematical notations. Let me explain each of them:

1. $F$ : This represents a class of functions. The curly F denotes that it's a set or collection of functions.
2. $f\in F$ : This means that $f$ is a function that belongs to (is an element of) the class $F$.
3. $f^{\ast }$ : The asterisk superscript typ

The “prompt shut down” clause seemed like one of the more important clauses in the SB 1047 bill. I was surprised other people I talked to didn't think seem to think it mattered that much, and wanted to argue/hear-arguments about it.

The clauses says AI developers, and compute-cluster operators, are required to have a plan for promptly shutting down large AI models.

People's objections were usually:

"It's not actually that hard to turn off an AI – it's maybe a few hours of running around pulling plugs out of server racks, and it's not like we're that likely to be in the sort of hard takeoff scenario where the differences in a couple hours of manually turning it off will make the difference."

I'm not sure if this is actually true, but, assuming it's true, it still seems to me like the shutdown clause is the one of the more uncomplicatedly-good parts of the bill.

Some reasons:

1. I think the ultimate end game for AI governance will require being able to quickly notice and shut down rogue AIs. That's what it means for the acute risk period to end. 

2. In the more nearterm, I expect the situation where we need to stop running an AI to be fairly murky. Shutting down an AI is going to be ve... (read more)

There was a particular mistake I made over in this thread. Noticing the mistake didn't change my overall position (and also my overall position was even weirder than I think people thought it was). But, seemed worth noting somewhere.

I think most folk morality (or at least my own folk morality), generally has the following crimes in ascending order of badness:

• Lying
• Stealing
• Killing
• Torturing people to death (I'm not sure if torture-without-death is generally considered better/worse/about-the-same-as killing)

But this is the conflation of a few different things. One axis I was ignoring was "morality as coordination tool" vs "morality as 'doing the right thing because I think it's right'." And these are actually quite different. And, importantly, you don't get to spend many resources on morality-as-doing-the-right-thing unless you have a solid foundation of the morality-as-coordination-tool.

There's actually a 4x3 matrix you can plot lying/stealing/killing/torture-killing into which are:

• harming the ingroup
• harming the outgroup (who you may benefit from trading with)
• harming powerless people who don't have the ability to trade or col

There's a skill of "quickly operationalizing a prediction, about a question that is cruxy for your decisionmaking."

And, it's dramatically better to be very fluent at this skill, rather than "merely pretty okay at it."

Fluency means you can actually use it day-to-day to help with whatever work is important to you. Day-to-day usage means you can actually get calibrated re: predictions in whatever domains you care about. Calibration means that your intuitions will be good, and _you'll know they're good_.

Fluency means you can do it _while you're in the middle of your thought process_, and then return to your thought process, rather than awkwardly bolting it on at the end.

I find this useful at multiple levels-of-strategy. i.e. for big picture 6 month planning, as well as for "what do I do in the next hour."

I'm working on this as a full blogpost but figured I would start getting pieces of it out here for now.

A lot of this skill is building off on CFAR's "inner simulator" framing. Andrew Critch recently framed this to me as "using your System 2 (conscious, deliberate intelligence) to generate questions for your System 1 (fast intuition) to answer." (Whereas previously, he'd known System 1 ... (read more)

A major goal I had for the LessWrong Review was to be "the intermediate metric that let me know if LW was accomplishing important things", which helped me steer.

I think it hasn't super succeeded at this.

I think one problem is that it just... feels like it generates stuff people liked reading, which is different from "stuff that turned out to be genuinely important."

I'm now wondering "what if I built a power-tool that is designed for a single user to decide which posts seem to have mattered the most (according to them), and, then, figure out which intermediate posts played into them." What would the lightweight version of that look like?

Another thing is, like, I want to see what particular other individuals thought mattered, as opposed to a generate aggregate that doesn't any theory underlying it. Making the voting public veers towards some kind of "what did the cool people think?" contest, so I feel anxious about that, but, I do think the info is just pretty useful. But like, what if the output of the review is a series of individual takes on what-mattered-and-why, collectively, rather than an aggregate vote?

Something struck me recently, as I watched Kubo, and Coco - two animated movies that both deal with death, and highlight music and storytelling as mechanisms by which we can preserve people after they die.

Kubo begins "Don't blink - if you blink for even an instant, if you a miss a single thing, our hero will perish." This is not because there is something "important" that happens quickly that you might miss. Maybe there is, but it's not the point. The point is that Kubo is telling a story about people. Those people are now dead. And insofar as those people are able to be kept alive, it is by preserving as much of their personhood as possible - by remembering as much as possible from their life.

This is generally how I think about death.

Cryonics is an attempt at the ultimate form of preserving someone's pattern forever, but in a world pre-cryonics, the best you can reasonably hope for is for people to preserve you so thoroughly in story that a young person from the next generation can hear the story, and palpably feel the underlying character, rich with inner life. Can see the person so clearly that he or she comes to live inside them.

Realistical... (read more)

I started writing this a few weeks ago. By now I have other posts that make these points more cleanly in the works, and I'm in the process of thinking through some new thoughts that might revise bits of this.

But I think it's going to be awhile before I can articulate all that. So meanwhile, here's a quick summary of the overall thesis I'm building towards (with the "Rationalization" and "Sitting Bolt Upright in Alarm" post, and other posts and conversations that have been in the works).

(By now I've had fairly extensive chats with Jessicata and Benquo and I don't expect this to add anything that I didn't discuss there, so this is more for other people who're interested in staying up to speed. I'm separately working on a summary of my current epistemic state after those chats)

• The rationalsphere isn't great at applying rationality to its own internal politics
• We don't seem to do much better than average. This seems like something that's at least pretty sad, even if it's a true brute fact about the world.
• There have been some efforts to fix this fact, but most of it has seemed (to me) to be missing key

Conversation with Andrew Critch today, in light of a lot of the nonprofit legal work he's been involved with lately. I thought it was worth writing up:

"I've gained a lot of respect for the law in the last few years. Like, a lot of laws make a lot more sense than you'd think. I actually think looking into the IRS codes would actually be instructive in designing systems to align potentially unfriendly agents."

I said "Huh. How surprised are you by this? And curious if your brain was doing one particular pattern a few years ago that you can now see as wrong?"

"I think mostly the laws that were promoted to my attention were especially stupid, because that's what was worth telling outrage stories about. Also, in middle school I developed this general hatred for stupid rules that didn't make any sense and generalized this to 'people in power make stupid rules', or something. But, actually, maybe middle school teachers are just particularly bad at making rules. Most of the IRS tax code has seemed pretty reasonable to me."

Over in this thread, Said asked the reasonable question "who exactly is the target audience with this Best of 2018 book?"

By compiling the list, we are saying: “here is the best work done on Less Wrong in [time period]”. But to whom are we saying this? To ourselves, so to speak? Is this for internal consumption—as a guideline for future work, collectively decided on, and meant to be considered as a standard or bar to meet, by us, and anyone who joins us in the future? 

Or, is this meant for external consumption—a way of saying to others, “see what we have accomplished, and be impressed”, and also “here are the fruits of our labors; take them and make use of them”? Or something else? Or some combination of the above?

I'm working on a post that goes into a bit more detail about the Review Phase, and, to be quite honest, the whole process is a bit in flux – I expect us (the LW team as well as site participants) to learn, over the course of the review process, what aspects of it are most valuable. 

But, a quick "best guess" answer for now.

I see the overall review process as having two "major phases":

• Phase 1: Nomination/Review/Voting/Post-that-summarizes-the-voting
• Phase 2: Compila

A thing I might have maybe changed my mind about:

I used to think a primary job of a meetup/community organizer was to train their successor, and develop longterm sustainability of leadership.

I still hold out for that dream. But, it seems like a pattern is:

1) community organizer with passion and vision founds a community

2) they eventually move on, and pass it on to one successor who's pretty closely aligned and competent

3) then the First Successor has to move on to, and then... there isn't anyone obvious to take the reins, but if no one does the community dies, so some people reluctantly step up. and....

...then forever after it's a pale shadow of its original self.

For semi-branded communities (such as EA, or Rationality), this also means that if someone new with energy/vision shows up in the area, they'll see a meetup, they'll show up, they'll feel like the meetup isn't all that good, and then move on. Wherein they (maybe??) might have founded a new one that they got to shape the direction of more.

I think this also applies to non-community organizations (i.e. founder hands the reins to a new CEO who hands the reins to a new CEO who doesn't quite know what to do)

So... I'm kinda wonde... (read more)

More in neat/scary things Ray noticed about himself.

I set aside this week to learn about Machine Learning, because it seemed like an important thing to understand. One thing I knew, going in, is that I had a self-image as a "non technical person." (Or at least, non-technical relative to rationality-folk). I'm the community/ritual guy, who happens to have specialized in web development as my day job but that's something I did out of necessity rather than a deep love.

So part of the point of this week was to "get over myself, and start being the sort of person who can learn technical things in domains I'm not already familiar with."

And that went pretty fine.

As it turned out, after talking to some folk I ended up deciding that re-learning Calculus was the right thing to do this week. I'd learned in college, but not in a way that connected to anything and gave me a sense of it's usefulness.

And it turned out I had a separate image of myself as a "person who doesn't know Calculus", in addition to "not a technical person". This was fairly easy to overcome since I had already given myself a bunch of space to explore and change this week, and I'd spent the past few months transitioning into being ready for it. But if this had been at an earlier stage of my life and if I hadn't carved out a week for it, it would have been harder to overcome.

Man. Identities. Keep that shit small yo.

High Stakes Value and the Epistemic Commons

I've had this in my drafts for a year. I don't feel like the current version of it is saying something either novel or crisp enough to quite make sense as a top-level post, but wanted to get it out at least as a shortform for now.

There's a really tough situation I think about a lot, from my perspective as a LessWrong moderator. These are my personal thoughts on it.

The problem, in short: 

Sometimes a problem is epistemically confusing, and there are probably political ramifications of it, such that the most qualified people to debate it are also in conflict with billions of dollars on the line and the situation is really high stakes (i.e. the extinction of humanity) such that it really matters we get the question right.

Political conflict + epistemic murkiness means that it's not clear what "thinking and communicating sanely" about the problem look like, and people have (possibly legitimate) reasons to be suspicious of each other's reasoning.

High Stakes means that we can't ignore the problem.

I don't feel like our current level of rationalist discourse patterns are sufficient for this combo of high stakes, political conflict, and epistemi... (read more)

Seems like different AI alignment perspectives sometimes are about "which thing seems least impossible."

Straw MIRI researchers: "building AGI out of modern machine learning is automatically too messy and doomed. Much less impossible to try to build a robust theory of agency first."

Straw Paul Christiano: "trying to get a robust theory of agency that matters in time is doomed, timelines are too short. Much less impossible to try to build AGI that listens reasonably to me out of current-gen stuff."

(Not sure if either of these are fair, or if other camps fit this)

Using "cruxiness" instead of operationalization for predictions.

One problem with making predictions is "operationalization." A simple-seeming prediction can have endless edge cases.

For personal predictions, I often think it's basically not worth worrying about it. Write something rough down, and then say "I know what I meant." But, sometimes this is actually unclear, and you may be tempted to interpret a prediction in a favorable light. And at the very least it's a bit unsatisfying for people who just aren't actually sure what they meant.

One advantage of cruxy predictions (aside from "they're actually particularly useful in the first place), is that if you know what decision a prediction was a crux for, you can judge ambiguous resolution based on "would this actually have changed my mind about the decision?"

("Cruxiness instead of operationalization" is a bit overly click-baity. Realistically, you need at least some operationalization, to clarify for yourself what a prediction even means in the first place. But, I think maybe you can get away with more marginal fuzziness if you're clear on how the prediction was supposed to inform your decisionmaking)

⚖ A year from now, in the three months prior, will I have used "cruxiness-as-operationalization" on a prediction, and found it helpful. (Raymond Arnold: 50%)