The commenting guidelines allows users to set their own norms of communication for their own private posts. This lets us experiment with different norms to see which work better, and also allows the LessWrong community to diversify into different subcommunities should there be interest. It says habryka’s guidelines because that’s who posted this post; if you go back through the other open threads, you will see other people posted many of them, and different commenting guidelines here and there. I think the posts that speak to this the most are:
[Meta] New moderation tools and moderation guidelines (by habryka)
Meta-tations on Moderation: Towards Public Archipelago (by Raemon)
There’s a post somewhere in the rationalsphere that I can’t relocate for the life of me. Can anybody help?
The point was communication. The example given was the difference between a lecture and a sermon. The distinction the author made was something like a professor talking to students in class, each of whom then goes home and does homework by themselves, versus a preacher who gives his sermon to the congregation, with the expectation that they will break off into groups and discuss the sermon among themselves.
I have a vague memory that there were graphics involved.
I have tried local search on LessWrong, site search of LessWrong, and browsing a few post histories that seemed like they might be the author based on a vague sense of aesthetic similarity. I was sure it was here, but now I fear it may have been elsewhere or it is hidden in some other kind of post.
I really liked this essay.
And as hacking bad tests shrinks in importance, education will evolve to stop training us to do it.
This, however, is entirely excessive optimism.
I get all the normal pain/temperature/pressure/friction feedback just fine. It is only the problem of knowing where they are in space without looking at them.
I don’t know what the procedure for this is, but it occurs to me that if we can specify information about an environment via differential equations inside the neural network, then we can also compare this network’s output to one that doesn’t have the same information.
In the name of learning more about how to interpret the models, we could try something like:
1) Construct an artificial environment which we can completely specify via a set of differential equations.
2) Run a neural network to learn that environment with every combination of those differential equations.
3) Compare all of these to several control cases of not providing any differential equations.
It seems like how the control case differs from each of the cases-with-structural-information should give us some information about how the network learns the environmental structure.
I can vouch for sudden and significant gains in comfort and functionality by focusing on improving your posture. The method I used was less thorough than here—instead I just used an exercise band and a few stretching exercises to improve the shoulder position. This provided improved comfort immediately, and significant reduction in the fragility of my back in a matter of days.
I just discovered this sequence, and I am pleased and impressed. The subject of this post is something I have been looking at learning more about a lot recently, because I have a problem in the area.
Specifically, I never know where my feet are positioned. I can infer it, and I can confirm it, but I simply don’t feel the position of my feet in relation to the rest of my body. Even when I am trying to focus on it.
By contrast, I do feel where my calves are in space. Most of the time when I need to place my feet precisely, I am actually just aiming my calves at that point and relying on the fact that my feet are on the end of my calves.
This doesn’t strike directly at the sampling question, but it is related to several of your ideas about incorporating the differentiable function: Neural Ordinary Differential Equations.
This is being exploited most heavily in the Julia community. The broader pitch is that they have formalized the relationship between differential equations and neural networks. This allows things like:
applying differential equation tricks to computing the outputs of neural networks
using neural networks to solve pieces of differential equations
using differential equations to specify the weighting of information
The last one is the most intriguing to me, mostly because it solves the problem of machine learning models having to start from scratch even in environments where information about the environment’s structure is known. For example, you can provide it with Maxwell’s Equations and then it “knows” electromagnetism.
There is a blog post about the paper and using it with the DifferentialEquations.jl and Flux.jl libraries. There is also a good talk by Christopher Rackauckas about the approach.
It is mostly about using ML in the physical sciences, which seems to be going by the name Scientific ML now.
Strong upvote, this is amazing to me. On the post:
Another example of explaining the intuitions for formal results less formally. I strongly support this as a norm.
I found the graphics helpful, both in style and content.
Some thoughts on the results:
This strikes at the heart of AI risk, and to my inexpert eyes the lack of anything rigorous to build on or criticize as a mechanism for the flashiest concerns has been a big factor in how difficult it was and is to get engagement from the rest of the AI field. Even if the formalism fails due to a critical flaw, the ability to spot such a flaw is a big step forward.
The formalism of average attainable utility, and the explicit distinction from number of possibilities, provides powerful intuition even outside the field. This includes areas like warfare and business. I realize it isn’t the goal, but I have always considered applicability outside the field as an important test because it would be deeply concerning for thinking about goal-directed behavior to mysteriously fail when applied to the only extant things which pursue goals.
I find the result aesthetically pleasing. This is not important, but I thought I would mention it.
I feel like this was rendered its own explicit meme in the form of The Game.
They ask whether TFP and related measures undervalue the tech sector. They conclude no:
Countries with smaller tech sectors than the US see a similar productivity slowdown.
Even if undervalued, the tech sector is not big enough to explain the whole slowdown in the US.
The slowdown begins in 1973, predating the tech sector.
They agree, and even raise approximately the same point:
To consider a simple example, imagine that an American company is inefficient, and then a management consultant comes along and teaches that company better personnel management practices, thereby boosting productivity. Does this count as an improvement in TFP or not? Or is it simply an increase in labor supply, namely that of the consultant? On one hand, some hitherto-neglected idea is introduced into the production process. That might militate in favor of counting it as TFP. On the other hand, the introduced idea is not a new one, and arguably the business firm in question is simply engaged in “catch up” economic growth, relative to more technologically sophisticated firms.
I am confused by their distinction between “catch up” growth and regular growth; it seems to me it should not matter how long it takes for an idea to diffuse when counting its value. Consider if each idea were like a corporation: it’s not like anyone dismisses the growth that happened after the iPod came out as “catch up” value, and only gains during Jobs’ original tenure count as “real” value.
It does seem clear to me that the timing problem makes it very difficult to disentangle from other ideas at this high level.
Problems with the TFP:
For instance, all attempts to measure scientific progress through productivity come up against a timing problem: the innovation does not happen at the same time as it is adopted.
Comin, Diego, Bart Hobijn, and Emilie Rovito. Five facts you need to know about technology diffusion. No. w11928. National Bureau of Economic Research, 2006.
Undervalues enhancements to labor, capital, or land:
many scientific advances work through enabling a greater supply of labor, capital, and land, and those advances will be undervalued by a TFP metric. Let’s say someone invents a useful painkiller, and that makes it easier for many people to show up to work and be productive. Output will rise, yet that advance will show up as an increase in labor supply, rather than as an increase in technology or scientific knowledge.
Some ideas are counted as capital:
The more general problem is that many scientific and technological advances are embodied in concrete capital goods.
Some ideas are counted as labor:
If a worker generates and carries forward a new scientific idea for producing more with a given amount of labor, that measures the same way as the worker being taught greater conscientiousness and producing more.
It is not clear how consistent this is:
In defense of TFP measures, these problems are not always so serious if these biases are roughly constant over time. In that case, changes in TFP still would reflect changes in the rate of progress of science and technology. The absolute level of TFP could be biased by capital-embodied and labor-embodied technical change, but over time, for comparisons, the expected sign of that bias might be close to zero. Still, it is not obvious that the rate embodiment of new ideas into capital and labor, in percentage terms, should be constant over time.
They look at Total Factor Productivity instead:
Total factor productivity is an attempt to measure overall economic effectiveness: how much a society can do with the inputs it has. TFP, multi-factor productivity, or the Solow residual are all different names for this same concept. It refers to the amount of output growth left unexplained after accounting for all inputs, i.e. it is a residual and not something we measure directly. So if output grew by ten, and the contributions of land, labor, and capital each were judged at 3 (total of 9), TFP would be measured at one. As such, it is vulnerable to measurement errors in any of the main series that go into its calculation. Nonetheless the hope is that this variable measures the “left over” contribution of ideas to the process of production, and thereby helps us measure the efficacy of science. It won’t confuse progress in science with a country having a large stock of oil.
This seems to match the historical record better:
One advantage of TFP is that it seems to correspond to common intuitions as to when scientific progress was especially high. Robert H. Gordon, in his book The Rise and Decline of American Economic Growth, has argued at great length that scientific and technological progress reached a peak in the early part of the twentieth century. That was a time when fossil fuels, electrification, industrialization, nitrogen fertilizer, cars, radio, telephones, clean water, vaccines, and antibiotics all took on major roles in human lives in the wealthier countries. Within a matter of decades,, human life was transformed, in large part because of the extension and application of the earlier Industrial Revolution. Consistent with this picture, American TFP typically grew quickly in the 1920s and 1930s, ranging from between two to slightly over three percent per year. In more recent times, in contrast, TFP growth often has ranged between one and one and a half percent.
Disentangling science and GDP:
For instance, Norwegian GDP per capita is typically 20-30% above that of Sweden and Denmark, but Norway is accessing the same general stock of scientific knowledge. Similarly, US agricultural productivity per labourer and labour hour outperformed Europe during the 19th and 20th centuries, but some of that gap probably sprung from a high ratio of land to inhabitant, rather than an inherent technological advantage, and for some of that time American science may have been behind that of Europe.
The relevant footnote:
On agricultural productivity comparisons, see Broadberry, Stephen, and Mary O’Mahony. “Britain’s Twentieth-Century Productivity Performance in International Perspective.” Work and pay in the twentieth century (2007): 301-329, and Broadberry, Stephen N., and Douglas A. Irwin. “Labor productivity in the United States and the United Kingdom during the nineteenth century.” Explorations in Economic History 43, no. 2 (2006): 257-279.
The planning fallacy for garage projects is an interesting problem, because it doesn’t lend itself immediately to a reference class unless you have done a lot of projects before.
Still, next time you want to tackle a garage invention that you can just predict it will take as long as this one. It will be interesting to see how the difference between projects compares to the planning fallacy impact on a single project; in very big projects, the size of the project completely dominates any considerations of field or technology.
Come to think of it, does anyone know if there is a maker community somewhere that records its budgets and timelines? Maybe that could be used as a reference class for garage projects.
I nominate this post because it does a good job of succinctly describing different ways to approach one of the core problems of civilization, which is to say choosing policy (or choosing policy choosers). A lot of our activity here is about avoiding catastrophe somehow; we have spent comparatively little time on big-picture things that are incremental improvements.
Anecdotally, this post did a good job of jogging loose a funk I was in regarding the process of politics. Politics is a notoriously *ugh field* kind of endeavor in the broader culture, and a particular taboo in this community. And yet, the very act of seriously considering what the options are is a soothing balm when you are otherwise in a state of overwhelming disgust. It’s like the ritual of evitable dismay.
I nominate this post for two reasons.
One, it is an excellent example of providing supplemental writing about basic intuitions and thought processes, which is extremely helpful to me because I do not have a good enough command of the formal work to intuit them.
Two, it is one of the few examples of experimenting with different kinds of presentation. I feel like this is underappreciated and under-utilized; better ways of communicating seems like a strong baseline requirement of the rationality project, and this post pushes in that direction.
I have definitely linked this more than any other post. The key insight for me is that common knowledge is something which has specific costs and trade-offs. Previously I implicitly viewed common knowledge as an accident or a feature of the environment.
Mencius Moldbug is the pen name of Curtis Yarvin.