2024-06-07

• Saw this post on HN about declining birth rates in Japan. Most comments followed a similar pattern of being very concerned about the economic impact of declining birth rates (things like “how will we pay for Social Security???”). I don’t totally disagree with the concerns, but I also am interested in ideas that would challenge this narrative. And one came to me from another area that seems somewhat unrelated: bullshit jobs. The tl;dr here is that if there are enough bullshit jobs, then maybe the declining population isn’t that big a deal. We will shed the bullshit jobs and people will instead spend their time on more useful work and individual productivity will rise to counteract the population loss. I think this idea isn’t totally crazy. It was pretty interesting (in some sense) to see Twitter/X shed 80% of its staff and seemingly have very limited impact from that. Now, I wouldn’t guess that 80% of all jobs are bullshit jobs, but it does seem likely that a pretty good percentage are. And if, say, 30% of these “knowledge worker/professional services” jobs are the kind you see in Office Space, then we could conceivably weather a 30%-ish population decline without reduced output…or something like that. I don’t want to get hung up on the specific numbers. The point is, there seems reason to believe these are significant productivity gains left on the table in service of other, valid social goals. And while declining population would be a stress to the existing structure of things, there is probably lots of room to reorganize things to mitigate the downsides.
  
• Been making more AI summaries of my recent highlights. In previous post, I mentioned the idea that this would help avoid the “glossing over” problem associated with re-reading notes. Been doing it for a few days now, and I think it depends heavily on the structure of the resulting summary. It currently comes out as a huge blob of markdown (I’ve posted an example here). When it comes out as a huge blob, it still ends up suffering from the “glossing over” problem like 60% of the time for me. My thinking is that it’s because I can see it’s a huge blob and would take me many minutes to read through. And that conflicts with my goal of having the summary be a quick, efficient review of recent reading. Naturally, I am wondering whether presenting it in smaller chunks would be better/lead to more active engagement. Or maybe there’s a way to turn the summary into interesting question prompts that extend the summary/highlights with interesting info (e.g. in a section about off-CPU analysis of threads, you could have a prompt about what year/decade were threads invented and why, etc). The idea being that this info would be sufficiently “new” to be interesting and sufficiently “related” to have some benefits around reinforcing what you’ve read.
  
• Reflecting a bit on the above, one of the main goals (maybe the entire goal) is to make engaging with educational information fun and interesting. My model here of LLMs/gen AI tools is that they have strong potential to be an “interestingness engine.” Right now, I’m in the phase of trying to figure out, at a basic level, how to make something interesting that you’ve already seen before. The next level would seem to involve having a system dynamically adjust to meet the user where they are.
  
• Read this interesting blog post from Uber about counting how many rides drivers provide/riders take. Need to jump to a call with David to practice coding problems, but found it to have a reasonable discussion of how to achieve low latency for analytics queries on super high throughput data. Relies on Apache Pinot to do a lot of heavy lifting it looks like. Would be cool to know more about the structure of Pinot, the blog made it seem like it uses LSM trees with a columnar format.

---

---

---