Amateur Hour

A little light on the reading the last couple of months, thanks to the holidays.

Books

I spent some time reading ethnographies over the last few months. This was mostly new to me – although I had several friends from college in anthropology / sociology who talked about ethnographies, this was the first time I actually sat down to read them myself. I was pleasantly surprised by how readable they were – I was expecting things to be denser and more jargony, but overall I think the authors actually did a good job explaining everything. There were definitely a lot of references to previous ethnographies and to various social theorists that I didn’t really understand, but even without that knowledge I was able to follow the main argument.

(As a side note – I wonder if the titling scheme is standard somehow? They all follow a schema of <Generic Name>: <the actual subject matter>.

The Charisma Machine: The Life, Death, and Legacy of One Laptop per Child

https://mitpress.mit.edu/9780262537445/the-charisma-machine/

This was a great ethnography about OLPC, covering both the work going on at the MIT Media Lab but also work in the field in Paraguay. I vaguely remembered OLPC being a “hot topic” when I was on middle / early high school, but I haven’t heard from it for years and I wanted to know what happened. The short answer is that it pretty much failed, and this book goes into some of the reasons why. Some things I particularly enjoyed about this book:

• It would’ve been easy to dismiss the entire project as just another techno-utopian failure, but the author really showed how the ideological commitments (e.g. the idea that students would teach themselves if only they had the chance) directly led to some of the implementation problems (e.g. no pedagogy for the teachers, and no tools for the teachers to actually manage 20+ laptops at once).
• I loved the concept of “the technologically precocious boy”, which refers to a very specific vision of (male) youth centered around creativity, exploration, and some amount of mischievousness. It’s a label that I definitely benefited, even during my early 20s and the extended adolescence of college.
  • There’s probably some fertile ground for comparing “technologically precocious” boy vs “the artistically genius boy” or the “mathematically advanced boy”, which I think are variants of the same general theme.
• I’ve always been a little suspicious of Papert’s theory of constructionist learning, and was gratified to see some discussion about its actual validity. The author basically claims that constructionist learning never had much empirical support (and that Papert’s rollout of Logo to “normal” schools basically was total failure), and that it mostly survives because of a kind of nostalgia about how the researchers felt about their childhood learnings.
  • This more-or-less matches my impressions – I don’t think I’ve ever heard of a rigorous evaluation of whether, say, Scratch actually is good at teaching people. Instead, I hear a lot of “well, I learned a lot by doing X, Y, Z” or “this other famous person used it, so it must be good”, but very little concern about whether it works for non cherry-picked populations.
• The general description of a “charismatic” hype-machine drawing attention away from practical and sustainable improvements to education seems all too familiar for me. See also: The Five Levels of Hype

Of Two Minds: An Anthropologist Looks at American Psychiatry

https://www.penguinrandomhouse.com/books/104443/of-two-minds-by-t-m-luhrmann/

I’ve been interested in the intellectual underpinnings of therapy for quite some time now. Therapy is very common in my peer group (to the point where I assume people are in therapy, even if they never bring it up), and “therapy-speak” has become ubiquitous in our culture (e.g. “trauma”, “gaslight”, “toxic”, etc). Personally, though, I had mildly negative experience when I tried therapy in college, and when I looked into the research literature on therapy I basically ran into a brick wall. It seems like there are dozens of different “schools of therapeutic thought” (e.g. Freudian Psychoanalysis¹? Cognitive Behavioral Therapy? Gestalt Therapy?), with conflicting theories of how the mind works and how treatments should go forward. It also seemed like the school of thought was basically irrelevant – most of the differences in treatment effectiveness seemed either to be from random chance or from the skill of the practitioner, not the type of therapy they were giving. It seems like a case where therapy is a pre-paradigm world, where things seem to work but no one really knows why (and the explanations don’t have much actual explanatory power).

I mentioned this interest (probably several times) to some friends, and one of them gifted me this book in response for my birthday. Now, 7+ months later I have finally cleared it off my todo list, giving me some hope that I can actually make forward progress on that list.

As the title says, this is an ethnography about American Psychiatry, and how psychopharmacology and psychotherapy (the two minds) interact during psychiatric training. It didn’t quite answer my question about the intellectual underpinnings, but the details about the training process was absolutely fascinating. The book cited a ton of of prior work on medical training, so I’m sure I lost a lot of nuance about what was specific to psychiatry and what was more general to the American medical system. Some quick thoughts:

• I enjoyed the discussion of how “real” mental illness categories are (e.g. how do you know its schizophrenia vs bipolar disorder). The way doctors talked about it – as very “real” concepts that crystallize given enough experience, even if the actual DSM definitions are quite vague and unhelpful – matches how I think about certain concepts (e.g. “coupling” in software engineering)
• I found the politics of the psychotherapy training to be fascinating. Learning how to give psychotherapy requires being in psychotherapy to see the other side (so-to-speak), but this creates all kinds of weird power dynamics between the trainees and their trainers (who are not their direct supervisors, but often supervise other trainees).
  • The degree to which trainees started using therapeutic concepts to analyze other people in their lives (and especially their fellow trainees) was really fascinating, and maybe a pre-cursor to today’s ubiquitous therapy-talk.
• The economics of who pays for mental health treatment was also a super interesting discussion. Residents are mostly paid by Medicaid, which means they often see the sickest and most acutely mentally ill (e.g. those with active psychosis), while established therapists often only see upper middle class people who can afford their fees (and thus can interact with the world at some baseline level). This also leads to a lot of diverging viewpoints on proper psychiatric treatment.

I would be very curious to see how much of the book (which was published in 2001) is still true today.

Privilege: The Making of an Adolescent Elite at St. Paul’s School

https://press.princeton.edu/books/paperback/9780691156231/privilege

Shamus Khan was a professor at Columbia while I was an undergrad, and his book was repeatedly recommended to me by other students while I was there. It took a while, but I finally got around to reading it. I really enjoyed it – I found that it mostly resonated with my experience at Columbia (during the “extended adolescence” of college)

• The central argument (that “privilege” is characterized by a type of “ease” in moving through different spaces) definitely resonated with my experience at Columbia.
• The discussion of elite engagement with a broad array of cultural artifacts (e.g. Homer and Vergil, but also Taylor Swift and Megan Thee Stallion) also resonated, as did the shallowness of most of that engagement. A lot of my memories of Columbia was basically cultural posturing by people who didn’t know what they were talking about.

Although admittedly other parts of the St Paul’s experience seemed very foreign to me:

• St Paul’s seemed much more ritualistic than Columbia did (although to be fair, I actually think Columbia is sort of weird in how few rituals there are, which probably contributes to the weak group identity I feel with Columbia).
• St Paul’s also seemed to have this assumption that everyone was exception and destined for greatness in their field (e.g. obviously the artistic kid was going to be a genius painter). This was very different from my experience – honestly, I came away from Columbia thinking that a lot of people would probably end up just being “normal” bankers or software engineers or doctors: a good life but not something I’d call “exceptional”.

Academic Papers

• Occupancy Flow Fields for Motion Forecasting in Autonomous Driving: A short paper for a better loss function for vehicle predictions. Basically, you blend a flow field loss (predicting the velocities of objects) with an occupancy grid loss (predictions of what grid cells in space are occupied), along with a consistency loss to ensure the two predictions are “aligned”.
• A scalable and generic approach to range joins: Most join algorithms are specialized for equijoins like JOIN ON t1.col = t2.col, but sometimes you need to do a range join like JOIN ON t1.col > t2.col or JOIN ON ABS(t1.col - t2.col) < 5. This paper goes into some detail about how to solve those joins efficiently – the TLDR is to use implicit kd-trees for the ranges and hash joins for the equijoins. Another great paper from the Umbra group at Munich.
• Generalized intersection over union: A metric and a loss for bounding box regression: sort of an obvious extension on an IOU loss when the bounding boxes are totally non-overlapping. Given bounding boxes A and B, you find the smallest enclosing box C that contains both A and B. The generalized IOU is just the standard IOU minus the proportion of C that is not filled by A or B (i.e. |C - (A union B)| / |C|). Neat idea, although I think the evaluation wasn’t super convincing compared to a normal YOLO / R-CNN loss.

Industrial Policy

Since Industrial Policy seems back in vogue, I spent some time reading an assortment of papers on industrial policy and innovation:

• Reviving America’s Forgotten Innovation System: Fostering US Growth through Incremental Product and Process Innovation: this basically divides technologies up into “product innovation” (e.g. a new type of thing) vs “process innovation” (how to deliver roughly the same thing but better/cheaper/faster?). The US has lots of support for product innovation (e.g. patents, basic research funding, venture capital, etc) but has relatively little support for process innovation. Seems like a plausible argument, although the proposed solution was not as compelling.
• Knowledge, skills and organizational capabilities for structural transformation: Argues that organizational capability (“the effectiveness of an organization in coordinating and optimizing [collective activities]") is a crucial component needed for economic development, independent of the individual skills or machinery that a firm has. The concept seems relevant if a bit ill-defined (I guess Harvard Business Review would call this “organizational culture” or something?).
• How Technology Grows: a possible answer to “why should we (the US) care about manufacturing?” The author basically argues that active manufacturing is needed to sustain “process knowledge” and “communities of engineering practice”. Without these, it becomes almost impossible to continue innovating and staying on the technological frontier.
• Complex, Established “Legacy” Sectors: The Technology Revolutions That Do Not Happen: An argument about why technological innovations has a hard time getting traction in “legacy” sectors (healthcare, mass transit infrastructure, electrical generation), even though the opportunities are so huge. Honestly, I found the piece kind of sloppy, but I did like some of the examples of market structure problems (e.g. collective action problems where everyone has to switch over, or “non-appropriability” issues where the person paying for the new tech isn’t the one who benefits).

Random things on the Internet

• You Might as Well Timestamp It: a nice piece of software engineering wisdom.
• How a Chinese American Gangster Transformed Money Laundering for Drug Cartels: a really fascinating look at how global organized crime. A little conspiratorial at times (e.g. is fentanyl really a Chinese plot threatening American stability?), but quite good otherwise. The basic plot is that Mexican cartels were having trouble money laundering all their money, while Wealthy Chinese folks desperately wanted access to dollars due to currency controls preventing them converting enough RMB). Connect the two and profit.
• The Daily Grind: a great reminder of just how labor intensive life was before the industrial age.
• Mushtaq Khan Interview w/ 80000 Hours: I really enjoyed this interview – it was a super interesting view on some topics I hadn’t thought deeply about before. I don’t love a lot of the 80000 Hours podcasts (I find they’re often too focused on weird philosophical concerns or on AI risks), but this one was very down-to-earth and informative.