I've been using Xonsh for about four years now, and it continues to be a solid daily driver unix shell for me, despite some issues.

Before xonsh, it was extremely common for me to feel torn between writing a shell script and a Python script for a given small project: "well, this script does some complicated logic, which would be a massive pain in bash, but it also extensively deals with the kind of OS resources that are easy to handle in bash (files, text streams, process management, ...) but require a bunch of boilerplate in Python". Xonsh means you don't have to choose. Everything easy in Python is easy; and (almost) everything easy in bash is easy.

Day to day, I also love that I can have various Python utilities loaded directly into my shell - I use python's math capabilities constantly, and really appreciate the rich data types and utilities, both built in and from standard libraries like collections and itertools.

The biggest issue is that xonsh syntax is kind of "beholden" to Python syntax, but of course the python team doesn't see "this breaks xonsh scripts" as sufficient for a change to be "breaking". This sometimes causes stuff to break in your xonsh scripts if you were using clever mixtures of shell and Python syntax, which later gets parsed differently due to a Python update. I have a standing intention to figure out an easy way to peg my xonsh environment to a particular Python version to avoid this churn. Maybe I'll try nix for the fifth time.

The second biggest issue is that it's slow, compared to bash. This basically doesn't matter for me, since (a) when I'm doing anything performance critical I'm using Rust anyway, and (b) my usual workflow involves long lived sessions, so startup latency is acceptable.

Overall, it's definitely worth the trade-off for me vs all other shells I've tried (I've used bash, zsh, and fish for a year or more each, and tried others like nushell and elvish for at least a day); I'd recommend it if you often wish your shell had sane data types or powerful math baked in.

Observation: If I look at certain things I did when I was 15, I'm often extremely embarrassed by them, and have been since I was about 20. But I can't think of anything I've done since I was 20 that I feel similarly embarrassed about.

Possible explanations:

1. Personal growth/change rate stalled around 20; I stopped "advancing" in ways that would lead me to know enough to be embarrassed about more recent things. Or at least, I'm just less different from me-at-20 than me-at-20 was from me-at-15, and so more easily feel justified in post-20 actions (e.g. due to fundamental attribution error applying more cleanly)
2. Propensity to store embarrassing memories decreased around 20; I still do just as much fundamentally-embarrassing stuff but I'm less good at remembering it
3. I've hit some kind of "embarrassment critical threshold": a point in personal development where I outgrew or suppressed most impulses that would be embarrassing later. Also counts if I became better at predicting what would later be embarrassing.
4. Generically my life has had less variance per day after 20, or the moments in it have been differently-selected, resulting in fewer opportunities to do embarrassing things

1. Get some kind of bad respiratory sickness with a moderate fever
2. Test to see if it's COVID; it is
3. Get a remote doctor's appointment to see if you can get a paxlovid prescription.
4. Doctor is happy to prescribe paxlovid, and even emphasizes that you need to get it today, partly because you have risk factors for complications
5. [Feeling feverish] Unfortunately you live in the SF Bay area. There are only two pharmacies within 5 miles of your house, and both are known for having bad service, presumably because crime forced the other ones to close about five years ago
6. [Feeling even more feverish] CVS sends you a text saying that your insurance has declined the prescription and it will cost you $1777 for one week of paxlovid (not a typo)
7. You Google how this can be, and discover that Pfizer offers some kind of copay card that covers $1400 of it. You shrug and print out the card, wondering how it can be that this is a socially appropriate outcome; nobody told you about this program, you had to Google it yourself.
8. Put on a kf94; go to the closer CVS. Stand in line for 15 minutes. They tell you that, even though, yes, it's not closing time yet, they can't possibly fill your prescription in time; I guess they had assumed that no one would be stupid enough to actually want their $1777 prescription filled, and had cancelled it. Not to worry though, they say, just go to the other CVS across town: they're open two hours later. You protest that you don't actually have the prescription to hand to them, but they assure you that the other CVS will be able to "pull" the prescription.
9. [Still feeling feverish] Go to the other CVS. Wait another 15 minutes. "Oh, gosh, I'm sorry sir but actually there's nothing we can do because of the way they cancelled the prescription at the other CVS. You'll have to come back tomorrow."
10. Fall over and fucking die

The more I think about gradual disempowerment, the more I think I should be a lot less libertarian.

Like, imo the central problem with gradual disempowerment is almost identical to (one of) the central problem(s) of unregulated markets: resources go proportionally to those who produce more value-from-the-point-of-view-of-those-with-resources.

From the point of view of human welfare, this is pretty fine when market participants are all fairly similar in terms of their potential productivity. But when some humans have extremely low potential for productivity, those humans' welfare is totally ignored by the market except insofar as other market participants happen to care about it enough to spend resources on it directly / terminally. This is exactly the situation that all humans will probably be in soon.

And it's not like I know what should happen instead; I think I had believed that UBI would be sufficient to solve it in the human-only-economy case, since surely humans have at least a bit of built-in terminal empathy on average. But it feels like, in an economy and governance structure where humans and human-sympathizing AIs compete with non-human-sympathizing AIs, we're going have to solve this in a more robust and permanent way.

Seriously guys, polyethylene is such a fucking cool material.

You get very different properties depending on the length of the polymer chains:

• Short-ish chains: Low-Density Polyethylene (LDPE). This is what plastic grocery bags are made of. Easy to tear, easy to cut, lightweight.
• Highly branching chains: High-Density Polyethylene (HDPE). This is what Tyvek (the stuff from festival wristbands, construction projects, and some USPS envelopes) and gallon milk jugs are made of. Still easy to cut, but very hard to tear.
• Very long chains: Ultra-High Molecular Weight Polyethylene (UHMWPE, also called Dyneema or Spectra). This is one of the highest strength-to-weight ratio materials that you can buy in large quantities, competitive with the strongest carbon fiber (but also you can make it into flexible fabrics and cords, which you can't do with typical carbon fiber materials), twice as strong as spider silk, ten times stronger than steel by weight. You can't even cut the stuff with normal scissors. Some bulletproof jackets are made out of it. I just bought some cheap 1000lb-rated UHMWPE cord that is 2mm thick. 550lb paracord is twice as thick, and thus has four times the cross-sectional area, for only about half the strength.

All of these variants are extremely chemically stable, chemically simple, nontoxic, blah blah blah. When would you guess that this miracle material was discovered? The answer is... 1898 (though HDPE and UHMWPE weren't synthesized until... the 50s)

Around the beginning of my vacation I put some effort points into picking the "right" web browser for me as of mid-2025.

Desiderata:

• Open-source
• Non-kneecapped uBlock Origin
• Cross-device sync
• not on some weird slippery slope of trying to get monetized (though it's okay if it's downstream of a browser that is on that slope)

I had been using Brave for a few months since Chrome stopped supporting Manifest v2, but eventually Chromium is going to deprecate Manifest v2 as well, so Brave's days are numbered.

Conclusion:

• For desktop, LibreWolf: reasonably well-resourced Firefox rebrand without the monetization junk. To install it you have to tell macOS that it's okay that it hasn't been notarized by apple, e.g. using brew install --no-sandbox. There are some very annoying default settings that you'll want to change (e.g. I'd recommend turning off the stuff aimed at avoiding fingerprinting, which makes Google Maps buggy and prevents you from uploading images on many sites) but after I did that everything works great.
• for android, Fennec: similarly rebranded / un-monetized Firefox; available only from the F-Droid store. Didn't need to change any settings for it to be surprisingly nice; maybe even nicer than Chrome for android.

I don't use any iOS devices very often so I haven't figured out the "right" browser for me on iOS yet; probably there are firefox forks available there as well.

These can sync with each other using Firefox Sync, which you can even self-host if you want to badly (I don't).

I got a Claude Max subscription and have been playing around.

First big experiment was to see if we could reimplement a part of our pipeline from Julia into Rust. The quality of the code it produced was really good, and it did good analyses on what parts would be easier or harder to port. Eventually this convinced us that a port wouldn't have been worthwhile, because the Rust <> Julia interop is currently quite difficult. I think it might have taken us several days to come to that conclusion otherwise, so this was a big win.

Second significant experiment was asking Claude Code to optimize a pipeline. It pretty much failed abysmally – it successfully added a lot of profiling and came up with a lot of ideas, but not a single one of them made the code any faster. Later, I was able to make a significant improvement to that pipeline with about half an hour of work. So, a failure here.

Third was a moderately difficult refactor, replacing some dicts with structs, that then get serialized and read in another part of the codebase. Claude Code did really well here – it occasionally produced code that was overly verbose, but nothing was *wrong*, and it took corrections/feedback quite effectively. Another big win here.