Open Thread July 2019

Welcome! :)

I haven't thought about doing this with open problems, but I really like the idea and feel like I've done it implicitly with a number of problems important to me (coordination issues, moloch, etc).

I do however, do this explicitily when I learn new solutions, making sure I integrate them into my world model, can instinctively frame problems using those models, and tie them to other potential solutions.

It feels like beginning to explicitly to do this with problems is a large next step that could really take effectiveness to the next level. Thanks!

Short mashup from two sources:

Nielson proposes an informal model of mastery:

...their prior learning has given them better chunking abilities, and so situations most people would see as complex they see as simple, and they find it much easier to reason about.

...

In other words, having more chunks memorized in some domain is somewhat like an effective boost to a person's IQ in that domain.

where the chunks in question fit into the 7+/-2 of working memory. Relatedly, there is Alan Kay's quip:

"A change in perspective is worth 80 IQ points."

Which is to say the new perspective provides a better way to chunk complex information. In retrospect this feels obvious, but my model for multiple perspectives beforehand mostly a matter of eliminating blind spots they might have. I'll have to integrate the contains-better-chunks possibility, which basically means that seeking out new perspectives is more valuable than I previously thought.

Sounds great. Welcome!

It seems to me clear that, if Commenters have a policy of reporting the maximum of their perception and 10, then there is now less mutual information between the commenter's report and the actual post quality than there was previously. In particular, you now can't distinguish between a post of quality 2 and a post of quality 5 given any number of comments, whereas you could previously.

There's no official, endorsed CFAR handbook that's publicly available for download. The CFAR handbook from summer 2016, which I found on libgen, warns

While you may be tempted to read ahead, be forewarned - we've often found that participants have a harder time grasping a given technique if they've already anchored themselves on an incomplete understanding. Many of the explanations here are intentionally approximate or incomplete, because we believe this content is best transmitted in person. It helps to think of this handbook as a companion to the workshop, rather than as a standalone resource.

which I think is still their view on the matter.

I have heard that they would be more comfortable with people learning rationality techniques in-person from a friend, so if you know any CFAR alumni you could ask them (they'd probably also have a better answer to your question).

Welcome limerott! I hope your time on the site will be well-spent and feel free to ask any questions here in the Open Thread, or via Intercom in the bottom right corner.

Ruby can probably get us the answer to whether the premise is true next week. Seems likely true, but I would only give it 75%.

I'll suggest the main change is open threads get fewer comments because the system makes open threads less conspicuous, and provides alternatives, such as questions, that are decent substitutes for comments.

Upvoted for testability.

Constructivist learning theory is a relevant keyword; its premise is pretty much directly the same as in your quote (my emphasis added):

An important restriction of education is that teachers cannot simply transmit knowledge to students, but students need to actively construct knowledge in their own minds. That is, they discover and transform information, check new information against old, and revise rules when they do not longer apply. This constructivist view of learning considers the learner as an active agent in the process of knowledge acquisition. Constructivist conceptions of learning have their historical roots in the work of Dewey (192 9), Bruner (1961), Vygotsky (1962), and Piaget (1980). Bednar, Cunningham, Duffy, and Perry (1992) and von Glasersfeld (1995) have proposed several implications of constructivist theory for instructional developers stressing that learning outcomes should focus on the knowledge construction process and that learning goals should be determined from authentic tasks with specific objectives. Similarly, von Glasersfeld (1995) states that learning is not a stimulus-response phenomenon, but a process that requires self-regulation and the development of conceptual structures through reflection and abstraction. It is important to note, in this respect, that constructivism is embodied in numerous ways and that these different views share important overlaps, but also c ontain major differences.

Constructivism is an approach to teaching and learning based on the premise that cognition (learning) is the result of "mental construction." In other words, students learn by fitting new information together with what they already know.

It's a big topic in educational research, so there's a lot of stuff about it out there. E.g. "How Learning Works: Seven Research-Based Principles for Smart Teaching" summarizes some research on it:

Students connect what they learn to what they already know, interpreting incoming information, and even sensory perception, through the lens of their existing knowledge, beliefs, and assumptions (Vygotsky, 1978 ; National Research Council, 2000 ). In fact, there is widespread agreement among researchers that students must connect new knowledge to previous knowledge in order to learn (Bransford & Johnson, 1972 ; Resnick, 1983 ). However, the extent to which students are able to draw on prior knowledge to effectively construct new knowledge depends on the nature of their prior knowledge, as well as the instructor's ability to harness it. In the following sections, we discuss research that investigates the effects of various kinds of prior knowledge on student learning and explore its implications for teaching.

One related example of it that I particularly like is in the paper Building Islands of Expertise in Everyday Family Activity. It discusses how a boy who's interested in trains initially learns stuff about trains, which then helps him learn more about other stuff as well:

By the time the boy turns 3-years old, he has developed an island of expertise around trains. His vocabulary, declarative knowledge, conceptual knowledge, schemas, and personal memories related to trains are numerous, well-organized, and flexible. Perhaps more importantly, the boy and his parents have developed a relatively sophisticated conversational space for trains. Their shared knowledge and experience allow their talk to move to deeper levels than is typically possible in a domain where the boy is a relative novice. For example, as the mother is making tea one afternoon, the boy notices the steam rushing out of the kettle and says: “That’s just like a train!” The mother might laugh and then unpack the similarity to hammer the point home: “Yes it is like a train! When you boil water it turns into steam. That’s why they have boilers in locomotives. They heat up the water, turn it into steam, and then use the steam to push the drive wheels. Remember? We saw that at the museum.”

In contrast, when the family was watching football—a domain the boy does not yet know much about—he asked “Why did they knock that guy down?” The mother’s answer was short, simple, stripped of domain-specific vocabulary, and sketchy with respect to causal mechanisms—“Because that’s what you do when you play football.” Parents have a fairly good sense of what their children know and, often, they gear their answers to an appropriate level. When talking about one of the child’s islands of expertise, parents can draw on their shared knowledge base to construct a more elaborate, accurate, and meaningful explanations. This is a common characteristic of conversation in general: When we share domain-relevant experience with our audience we can use accurate terminology, construct better analogies, and rely on mutually held domain- appropriate schema as a template through which we can scribe new causal connections.

As this chapter is being written, the boy in this story is now well on his way to 4- years old. Although he still likes trains and still knows a lot about them, he is developing other islands of expertise as well. As his interests expand, the boy may engage less and less often in activities and conversations centered around trains and some of his current domain-specific knowledge will atrophy and eventually be lost. But as that occurs, the domain-general knowledge that connected the train domain to broader principles, mechanisms, and schemas will probably remain. For example, when responding to the boy’s comment about the tea kettle, the mother used the train domain as a platform to talk about the more general phenomenon of steam.

Trains were platforms for other concepts as well, in science and in other domains. Conversations about mechanisms of locomotion have served as a platform for a more general understanding of mechanical causality. Conversations about the motivation of characters in the Thomas the Tank Engine stories have served as platforms for learning about interpersonal relationships and, for that matter, about the structure of narratives. Conversations about the time when downtown Pittsburgh was threaded with train tracks and heavy-duty railroad bridges served as a platform for learning about historical time and historical change. These broader themes emerged for the boy for the first time in the context of train conversations with his parents. Even as the boy loses interest in trains and moves on to other things, these broader themes remain and expand outward to connect with other domains he encounters as he moves through his everyday life.

Expertise research is also relevant; it talks about how people build up increasingly detailed mental representations of a domain they are learning, and which guide them when they decide what actions to take. The representations start out a coarse, but get increasingly detailed over time. This is an excerpt from the book Peak: Secrets from the New Science of Expertise, which notes a "chicken and egg" problem that requires grounding any skill in some pre-existing knowledge first:

As we’ve just seen from several studies, musicians rely on mental representations to improve both the physical and cognitive aspects of their specialties. And mental representations are essential to activities we see as almost purely physical. Indeed, any expert in any field can be rightly seen as a high-achieving intellectual where that field is concerned. This applies to pretty much any activity in which the positioning and movement of a person’s body is evaluated for artistic expression by human judges. Think of gymnastics, diving, figure skating, or dancing. Performers in these areas must develop clear mental representations of how their bodies are supposed to move to generate the artistic appearance of their performance routines. But even in areas where artistic form is not explicitly judged, it is still important to train the body to move in particularly efficient ways. Swimmers learn to perform their strokes in ways that maximize thrust and minimize drag. Runners learn to stride in ways that maximize speed and endurance while conserving energy. Pole-vaulters, tennis players, martial artists, golfers, hitters in baseball, three-point shooters in basketball, weightlifters, skeet shooters, and downhill skiers—for all of these athletes proper form is key to good performance, and the performers with the best mental representations will have an advantage over the rest. 

In these areas too, the virtuous circle rules: honing the skill improves mental representation, and mental representation helps hone the skill. There is a bit of a chicken-and-egg component to this. Take figure skating: it’s hard to have a clear mental representation of what a double axel feels like until you’ve done it, and, likewise, it is difficult to do a clean double axel without a good mental representation of one. That sounds paradoxical, but it isn’t really. You work up to a double axel bit by bit, assembling the mental representations as you go. 

It’s like a staircase that you climb as you build it. Each step of your ascent puts you in a position to build the next step. Then you build that step, and you’re in a position to build the next one. And so on. Your existing mental representations guide your performance and allow you to both monitor and judge that performance. As you push yourself to do something new—to develop a new skill or sharpen an old one—you are also expanding and sharpening your mental representations, which will in turn make it possible for you to do more than you could before.

Finally, while it only touches on this topic occasionally, Josh Waitzkin had some nice "from the inside" descriptions of this gradual construction of an increasing level of understanding in his book The Art of Learning:

I practiced the Tai Chi meditative form diligently, many hours a day. At times I repeated segments of the form over and over, honing certain techniques while refining my body mechanics and deepening my sense of relaxation. I focused on small movements, sometimes spending hours moving my hand out a few inches, then releasing it back, energizing outwards, connecting my feet to my fingertips with less and less obstruction. Practicing in this manner, I was able to sharpen my feeling for Tai Chi. When through painstaking refinement of a small movement I had the improved feeling, I could translate it onto other parts of the form, and suddenly everything would start flowing at a higher level. The key was to recognize that the principles making one simple technique tick were the same fundamentals that fueled the whole expansive system of Tai Chi Chuan.

This method is similar to my early study of chess, where I explored endgame positions of reduced complexity—for example king and pawn against king, only three pieces on the board—in order to touch high-level principles such as the power of empty space, zugzwang (where any move of the opponent will destroy his position), tempo, or structural planning. Once I experienced these principles, I could apply them to complex positions because they were in my mental framework. However, if you study complicated chess openings and middlegames right off the bat, it is difficult to think in an abstract axiomatic language because all your energies are preoccupied with not blundering. It would be absurd to try to teach a new figure skater the principle of relaxation on the ice by launching straight into triple axels. She should begin with the fundamentals of gliding along the ice, turning, and skating backwards with deepening relaxation. Then, step by step, more and more complicated maneuvers can be absorbed, while she maintains the sense of ease that was initially experienced within the simplest skill set.

I was there before it was fully done. As a person with a strong interest in UX I found it quite exciting.

It definitely tries to be a modern Xerox Park or something like that, and it does really feel like it's doing a lot of really interesting things in the UI space. I have a really hard time telling whether any of the UI ideas they are experimenting with will actually turn out to be useful and widely adopted, but it definitely helped me think about UX in a better way.

I have been! I thought it was an interesting experiment, and I really hope they have another community day so I can visit again. I think there's probably a lot to learn from it, and I think there are things that you can only learn effectively by trying out weird experiments in real life to see "how this feels". But I don't really expect anything to directly come of it -- the project is pretty janky, and while it's a fantastic platform for tiny cute demos, I don't think any concrete part of it (other than the general sense of "this is an inspiration to try to go recreate this neat type of interaction in a more robust way") is really useful.

There are quite a few RSS feeds. You can construct them by clicking on the "Subscribe via RSS" button on the frontpage:

Only barely related, but Grassmann numbers are hilariously weird. Among other properties, their square is always zero (though they’re generally non-zero).

The front page features are very useful for getting up to speed. Recently Curated is newer posts the mods thought were important, From the Archives are old posts that were well received, and Continue Reading helps keep track of the sequences (the core content of the site) so you can consume them over time.

Welcome!

I would think they would just buy regular clothes. The same way that you cannot buy only one shoe of a pair of shoes.

The intrinsic connection is primarily that they arose out of the same broad community, and there is heavy overlap between personnel as a consequence.

I say this is not a worrying sign, because the comparison isn't between their shared methods and some better memeplex, it is between their shared methods and the status quo. That is to say, there's no reason to believe something else would be happening in place of this; more likely everyone would have scattered throughout what was already happening.

It's very important to distinguish those etceteras you listed, because those are three different worlds. In the world where AI risk is in fact low, HRAD can still be very successful in mitigating it further and also fruitful in thinking about similar risks. In the world where the underlying philosophical assumptions are wrong, demonstrating that wrongness is valuable in-and-of itself to the greater safety project. In the world where alternate safety approaches have higher tractability, how would we even tell without comparison to the challenges encountered in HRAD?

HRAD is also the product of specific investigations into tractability and philosophical soundness. I expect they will iterate on these very questions again in the future. If it winds up a dead end I expect the associated communities to notice and then to shift focus elsewhere.

To sum up, we have noticed the skulls. Hail, humanity! We who are about to die salute you.