Posts made by members of the LessWrong 2.0 team are typically made from the perspective of the individual - even when they are writing in their capacity as LessWrong team members. My (Ruby’s) model of the team’s reason for this is that even if there are collective decisions, there are no collective models. Not real models.

When the team agrees to do something, it is only because enough of the individual team members individually have models which indicate it is the right thing to do. Our models might be roughly the same at a high-level such that you can describe a “common denominator” model, but this isn’t an actual model held by an actual person. I think such “common denominator” group models are undesirable for at least the following reasons:

  1. Pressure to form consensus reduces the diversity of models, sometimes going from multiple models per person to only a single model for the group. This can then result in overconfidence in the surviving model.
  2. There might be no group model. The group might have agreed on a decision, but they never reached consensus on the reasons for it.
  3. It is costly to describe group models. Either someone has to draft the model, get feedback, make revisions, repeat, until eventually it is “good enough” or someone describes a model putatively held by the group, but which is not actually representative of the group's thinking.
  4. In fact, no individual might endorse the group model as being their own.
  5. The person describing the group model doesn’t necessarily understand things they’re including which came from others.
  6. In averaging multiple models, detail is lost and you no longer have a model which can usefully generate predictions.
  7. No individual owns the model, making it hard for any one person to elucidate, defend, or be held accountable for it.
  8. Reluctance to speak on the behalf of others means that very little gets said.

Crucially, group models which get shared externally are very often not the models which were used to make decisions. If you want to understand a decision, you want the actual model which generated it.

Given the goal of sharing our actual true thinking with the outside world (rather than nicely curated PR announcements), the LessWrong team has the rough policy that we speak from our personal point of view and avoid committing to an impersonal, authoritative, “official view of LessWrong.”

I suspect (and I believe the team generally agrees) that individual team members posting from their own points of view will ultimately result in the outside world having a better understanding of our thinking (individual and collective) than if we attempt to aggregate our individual models into the “organization’s models”. Organizations don’t have models, people do.

That said, we talk a lot to each other and our models are correlated. We tend to agree on the broad outline of things, e.g. we agree at the crudest level that LessWrong is about rationality and intellectual progress, even if we don’t agree on more detailed framings and relative emphasis. We think roughly like each other, but don’t be surprised if a different team member says about a high-level vision post I wrote that it’s not their model of it or that they don't agree with every detail.

Seemingly, this communication policy might allow us (the LessWrong team) to weasel out of our public statements. “Oh, that’s just what Ruby said - the rest of us never said that.” This is far from the intention. This policy is focused on how we communicate our reasons for doing things rather than statements about our commitments or actions. If a LessWrong team member says the LessWrong team plans to do something (especially major directions), it’s fair game to hold the entire team accountable for doing that thing.

New to LessWrong?

New Comment
20 comments, sorted by Click to highlight new comments since: Today at 11:29 AM

(This is an unrelated question about LW that I'd like the LW team to see, but don't think needs its own post, so I'm posting it here.) I want to mention that it remains frustrating when someone says something like "I'm open to argument" or we're already in the middle of a debate, I give them an argument, and just hear nothing back. I've actually kind of gotten used to it a bit, and don't feel as frustrated as I used to feel but it's still pretty much the strongest negative emotion I ever experience when participating on LW.

I believe there are good reasons to address this aside from my personal feelings, but I'm not sure if I'm being objective about that. So I'm interested to know whether this is something that's on the LW team's radar as a problem that could potentially be solved/ameliorated, or if they think it's not worth solving or probably can't be solved or it's more of a personal problem than a community problem. (See this old feature suggestion which I believe I've also re-submitted more recently, which might be one way to try to address the problem.)

Ray has recently been advocating for a more general tagging system (kinda like Github and Discord but with tags optimized for LW-reactions like the ones in your feature suggestion), and the LW team has been more seriously exploring the idea of breaking voting down into two dimensions "agree/disagree" + "approve/disapprove". My guess is that both of those would help a bit, though for your use-case it also seems important that you know the identity of who reacted what way.

I think both agree/disagree and approve/disapprove are toxic dimensions for evaluating quality discussions. Useful communication is about explaining and understanding relevant things, real-world truth and preference are secondary distractions. So lucid/confused (as opposed to clear/unclear) and relevant/misleading (as opposed to interesting/off-topic) seem like better choices.

*disagrees with and approves of this relevant, interesting, and non-confused comment*

I think both agree/disagree and approve/disapprove are toxic dimensions for evaluating quality discussions.

Hmm, but are they more toxic than whatever "upvote/downvote" currently means? The big constraining factor on things like this seems to me to be complexity and inferential distance of what the voting means. I would be worried that it would be much harder to get people to understand "lucid/confused" and "relevant/misleading" though I am not confident.

Within the hypothetical where the dimensions I suggest are better, fuzziness of upvote/downvote is better in the same way as uncertainty about facts is better than incorrect knowledge, even when the latter is easier to embrace than correct knowledge. In that hypothetical, moving from upvote/downvote to agree/disagree is a step in the wrong direction, even if the step in the right direction is too unwieldy to be worth making.

Whose model is the boss model, though?

Edit: Answered elsethread:

habryka is main decision maker

[E]ven if there are collective decisions, there are no collective models. Not real models.

When the team agrees to do something, it is only because enough of the individual team members individually have models which indicate it is the right thing to do.

There's something kind of worrying/sad about this. One would hope that with a small enough group, you'd be able to have discussion and Aumann magicconvergence lead to common models (and perhaps values?) being held by everybody. In this world, the process of making decisions is about gathering information from team members about the relevant considerations, and then a consensus emerges about what the right thing to do is, driven by consensus beliefs about the likely outcomes. When you can't do this, you end up in voting theory land, where even if each individual is rational, methods to aggregate group preferences about plans can lead to self-contradictory results.

I don't particularly have advice for you here - presumably you've already thought about the cost-benefit analysis of spending marginal time on belief communication - but the downside here felt worth pointing out.

I took the line written to mean that there are no "opinion leaders". In a system where people could vote but actually trust someone elses judgement the amount of votes doesn't reflect the amount of judgement processes employed.

I also think that in a system that requires a consensus it becomes tempting to produce a false consensus. This effect is strong enough that in all context where people bother with the concept of consensus there is enough basis to suspect that it doesn't form that there is a significant chance that all particular consensuses are false. By allowing a system of functioning to tolerate non-consensus it becomes practical to be the first one to break a consensus and the value of this is enough to see requiring consensus to be harmful.

All the while it being true that while opinions diverge there is real debate to be had.

This comment actually made our own policy clearer to me, thanks!

FWIW, we spend loads of time on belief-communication. This does mean (as Ruby says) that many of our beliefs are the same. But some are not, and sometimes the nuances matter.

In this world, the process of making decisions is about gathering information from team members about the relevant considerations, and then a consensus emerges about what the right thing to do is, driven by consensus beliefs about the likely outcomes.

This doesn't seem very different from what we do, we just skip the step where everyone's models necessarily converge. We still converge on a course of action. (habryka is main decision maker so in the event that consensus-about-the-relevant-details doesn't emerge, tends to default to his judgment, or [empirically] to delaying action).

Even if they do converge (which they do quite frequently in simpler cases), I think the correct model of the situation is to say "I believe X, as does everyone else on my team", which is a much better statement than "we believe X", because the phrase "we believe" is usually not straightforwardly interpreted as "everyone on the team believes that X is true" instead it usually means "via a complicated exchange of political capital we have agreed to act as if we all believe X is true".

FWIW, we spend loads of time on belief-communication.

To clarify, I didn't think otherwise (and also, right now, I'm not confident that you thought I did think otherwise).

We still converge on a course of action.

Sure - I now think that my comment overrated how much convergence was necessary for decision-making.

I second Ray's claim that we spend loads of time on belief communication. Something like the Aumann convergence to common models might be be "theoretically" doable, but I think it'd require more than 100% of our time to get there. This is indeed a bit sad and worrying for human-human communication.

This is indeed a bit sad and worrying for human-human communication.

Is it newly sad and worrying, though?

By contrast, I find it reassuring when someone explicitly notes the goal, and the gap between here and that goal, because we have rediscovered the motivation for the community. 10 years deep, and still on track.

Suck it, value drift!

Hmm, I think you must have misunderstood the above sentence/we failed to get the correct point across. This is a statement about epistemology that I think is pretty fundamental, and is not something that one can choose not to do.

In a system of mutual understanding, I have a model of your model, and you have a model of my model, but nevertheless any prediction about the world is a result of one of our two models (which might have converged, or at the very least include parts of one another). You can have systems that generate predictions and policies and actions that are not understood by any individual (as is common in many large organizations), but that is the exact state you want to avoid in a small team where you can invest the cost to have everything be driven by things at least one person on the team understands.

The thing described above is something you get to do if you can invest a lot of resources into communication, not something you have to do if you don't invest enough resources.

I get the sense that you don't understand me here.

In a system of mutual understanding, I have a model of your model, and you have a model of my model, but nevertheless any prediction about the world is a result of one of our two models (which might have converged, or at the very least include parts of one another).

We can choose to live in a world where the model in my head is the same as the model in your head, and that this is common knowledge. In this world, you could think about a prediction being made by either the model in my head or the model in your head, but it makes more sense to think about it as being made by our model, the one that results from all the information we both have (just like the integer 3 in my head is the same number as the integer 3 in your head, not two numbers that happen to coincide). If I believed that this was possible, I wouldn't talk about how official group models are going to be impoverished 'common denominator' models, or conclude a paragraph with a sentence like "Organizations don’t have models, people do."

In this world, you could think about a prediction being made by either the model in my head or the model in your head, but it makes more sense to think about it as being made by our model

I don’t think this actually makes sense. Models only make predictions when they’re instantiated, just as algorithms only generate output when run. And models can only be instantiated in someone’s head[1].

… the integer 3 in my head is the same number as the integer 3 in your head, not two numbers that happen to coincide …

This is a statement about philosophy of mathematics, and not exactly an uncontroversial one! As such, I hardly think it can support the sort of rhetorical weight you’re putting on it…


[1] Or, if the model is sufficiently formal, in a computer—but that is, of course, not the sort of model we’re discussing.

I think models can be run on computers and I think people passing papers can work as computers. I do think it's possible to have an organization that does informational work that none of it's human participants do. I do appriciate that such work is often very secondary to the work that actual individuals do. But I think that if someone aggressively tried to make a system that would survive a "bad faith" human actor it might be possible and even feasible.

I would phrase is that the number 3 in my head and the number 3 in your head both correspond to the number 3 "out there" or to the ""common social" number 3.

For example my number 3 might participate in being part of a input to a cached results of multiplication tables while I am not expecting everyone else to do so.

The old philosphical problem of whether the red I see the the same red that you see kind of highlights how the reds could plausibly be incomparable while the practical reality that color talk is possible is not in question.