I think it's a pretty deep point -- your memory and consciousness are shaped by the media environment (media being the thing that Klosterman thinks about obsessively)

And the media environment drastically changed after the 90's -- because of the Internet

I think his main point was about access to media in different eras, but it's worth reading directly if you feel like that

I have to admit I feel similarly conflicted after a few recent REAL WORLD experiences. The first is that Gemini and Claude taught me a whole bunch of things about beer that I feel like I should have known, or a friend should have told me in the last few decades :-) (I'm in my mid 40's)

As background, I am a simple man, and I like one type of beer (pale ale), and drink 1 beer a day. But it has to be fresh, crisp, bitter, and not make me want to fall asleep (lower alcohol)

So here is some color on what I learned:

- I worked in SF / the bay area for nearly 2 decades, and I'm now in Philly. I have been wondering for like FOUR YEARS why when I buy pale ale, it is somehow OFF. Well Claude told me that east coast pale ale is actually more like "English Ale" (e.g. Yards brewing), and west coast pale ale is its own thing (e.g. Lagunitas)

Apparently west coast is what I developed a taste for -- it is more bitter. And this totally tracks for me. So the same term is used for 2 slightly different things, and AI cleared that up for me

- I walked into a beer store today in Philly, and there are 5 IPAs for every Pale Ale, which annoys me because I prefer the latter. (That happened in San Francisco too, and I'm sure many people here have had that experience)

And Gemini taught me that "session IPA" is actually a marketing term for Pale Ale. I always choose things labeled "pale ale", but many people want something labeled "IPA". So they came up with the term "session IPA" -- an IPA with lower alcohol, aka Pale Ale.

I use Gemini anonymously, so it doesn't remember things about me. And after I described my beer preferences and the fact that I'm in Philly, it specifically brought up Tonewood Brewing in NJ, which is in fact the beer I've drank the most in the last 3 years!

---

So yeah I find it eerie that the LLMs are helping me with words versus reality.

- the same term "pale ale" referring to slightly different things on the east coast and west coast -- this was tripping me up for years

- the term "session IPA" being a marketing term for Pale Ale -- also something I didn't know for years

- based on simple verbal descriptions, they are able to recommend the beers that I actually drank and liked (even though I use Gemini anonymously, and that was my first time asking Claude about beer)

I've been drinking something slightly "wrong" quite often, because of a confusion over words on the label. I guess I simply don't know anyone who I can ask dumb questions about beer to.

So breathing better during the day can be trained, even at an old age, and it improves sleep

Not everyone breathes suboptimally of course, but I think more do than realize it. There’s a reason that breath work is in the traditions of many different cultures, and why it survived

But things like this aren’t necessarily profitable or worth a doctor’s time, so you have to do them yourself, or see therapists, etc

I had a good experience with a myofunctional therapist and posture therapist

https://zef-lang.dev/implementation

https://news.ycombinator.com/item?id=47843194

It has list of more than 20 optimizations for interpreters, with measured speedups. I'm pretty sure I was looking for something like this 3-5 years ago, but it didn't exist

The other thing is watching the videos in public with the tinny speakers blaring. Judging by reactions on the trains, this is socially acceptable to most people now ???

But I’m using the slower language because it still integrates with more things

For example, one reason AI is all in Python is because CUDA is basically part of the C ecosystem (ie build system)

Hm interesting

So they are making the distinction between regular "human brain" coding and AI-assisted coding?

Regular coding could be described as "not doing the thing yourself, but telling the computer what to do"

(FWIW I do think there is a huge difference; however I am not sure the general public has a very good idea of what "programming" is. I remember having some code up on my screen and my educated family was confused, even at the concept)

Then I found ncdu, and haven’t looked back since. So it saved me a lot of time

Thank you and RIP

In this case it was the Zed editor spawning the zsh shell:

How to Lose Control of your Shell - https://registerspill.thorstenball.com/p/how-to-lose-control...

zsh has a bug where it doesn't do the job control cleanup properly in some cases -- when fork-exec() optimizations happen.

This can mess up the calling process. For example, here you could no longer kill Zed by hitting Ctrl-C, even after zsh is done.

My comment: https://lobste.rs/s/hru0ib/how_lose_control_your_shell#c_dfl...

I have seen this misconception many times

In Oils, we have some pretty minor elaborations of the standard model, and it makes things a lot easier

How to Parse Shell Like a Programming Language - https://www.oilshell.org/blog/2019/02/07.html

Everything I wrote there still holds, although that post could use some minor updates (and OSH is the most bash-compatible shell, and more POSIX-compatible than /bin/sh on Debian - e.g. https://pages.oils.pub/spec-compat/2025-11-02/renamed-tmp/sp... )

---

To summarize that, I'd say that doing as much work as possible in the lexer, with regular languages and "lexer modes", drastically reduces the complexity of writing a shell parser

And it's not just one parser -- shell actually has 5 to 15 different parsers, depending on how you count

I often show this file to make that point: https://oils.pub/release/0.37.0/pub/src-tree.wwz/_gen/_tmp/m...

(linked from https://oils.pub/release/0.37.0/quality.html)

Fine-grained heterogenous algebraic data types also help. Shells in C tend to use a homogeneous command* and word* kind of representation

https://oils.pub/release/0.37.0/pub/src-tree.wwz/frontend/sy... (~700 lines of type definitions)

One thing we learned from implementing job control in https://oils.pub is that the differing pipeline semantics of bash and zsh makes a difference

In bash, the last part of the pipeline is forked (unless shopt -s lastpipe)

In zsh, it isn't

    $ bash -c 'echo hi | read x; echo $x'  # no output
          
    $ zsh -c 'echo hi | read x; echo $x'
    hi

    bash$ sleep 5 | read
    ^Z
    [1]+  Stopped                 sleep 5 | read


    zsh$ sleep 5 | read    # job control doesn't apply to this case in zsh
    ^Zzsh: job can't be suspended

I think there was almost a "double bug" -- the script relied on the `read` output being "lost", even though that was likely not the intended behavior

    - testing
    - reviewing, and reading/understanding/explaining
    - operations / SRE

Interesting results:

1. Gemini pointed me back at MY OWN comment, above, an hour after I wrote it. So Google is crawling the web FAST. It also pointed to: https://learning.acm.org/bytecast/ep78-russ-cox

This matches my recent experience -- Gemini is enhanced for many use cases by superior recall

2. Claude also knows this, pointing to pages like: https://usesthis.com/interviews/jeff.dean/ - https://goodlisten.co/clip/the-unlikely-friendship-that-shap... (never seen this)

3. ChatGPT did the worst. It said

... they have likely crossed paths professionally given their roles at Google and other tech circles. ...

While I can't confirm if they know each other personally or have worked directly together on projects, they both would have had substantial overlap in their careers at Google.

(edit: I should add I pay for Claude but not Gemini or ChatGPT; this was not a very scientific test)

And I actually think Google needs to pay more attention to AI ethics ... but it's a publically traded company and the incentives are all wrong -- i.e. it's going to do whatever it needs to do keep up with the competition, similar to what happened with Google+ (perceived competition from Facebook)

Another Jeff Dean fact should be "Russ Cox was Jeff Dean's intern"

This was either 2006 or 2007, whenever Russ started. I remember when Jeff and Sanjay wrote "gsearch", a distributed grep over google3 that ran on 40-80 machines [1].

There was a series of talks called "Nooglers and the PDB" I think, and I remember Jeff explained gsearch to maybe 20-40 of us in a small conference room in building 43.

It was a tiny and elegant piece of code -- something like ~2000 total lines of C++, with "indexer" (I think it just catted all the files, which were later mapped into memory), replicated server, client, and Borg config.

The auth for the indexer lived in Jeff's home dir, perhaps similar to the protobuf DB.

That was some of the first "real Google C++ distributed system" code I read, and it was eye opening.

---

After that talk, I submitted a small CL to that directory (which I think Sanjay balked at slightly, but Jeff accepted). And then I put a Perforce watch on it to see what other changes were being submitted.

I think the code was dormant for awhile, but later I saw someone named Russ Cox started submitting a ton of changes to it. That became the public Google Code Search product [2]. My memory is that Russ wrote something like 30K lines of google3 C++ in a single summer, and then went on to write RE2 (which I later used in Bigtable, etc.)

Much of that work is described here: https://swtch.com/~rsc/regexp/

I remember someone telling him on a mailing list something like "you can't just write your own regex engine; there are too many corner cases in PCRE"

And many people know that Russ Cox went on to be one of the main contributors to the Go language. After the Code Search internship, he worked on Go, which was open sourced in 2009.

---

[1] Actually I wonder if today if this could perform well enough a single machine with 64 or 128 cores. Back then I think the prod machines were something like 2, 4, or 8 cores.

[2] This was the trigram regex search over open source code on the web. Later, there was also the structured search with compiler front ends, led by Steve Yegge.

In particular communication between doctors and nurses is worse, because it’s all mediated by software

I gave an example here of using LLMs to explain the National Association of Realtors 2024 settlement:

https://news.ycombinator.com/item?id=46040967

Buyers agents often say "you don't pay; the seller pays"

And LLMs will repeat that. That idea is all over the training data

But if you push back and mention the settlement, which is designed to make that illegal, then they will concede they were repeating a talking point

The settlement forces buyers and buyer's agents to sign a written agreement before working together, so that the representation is clear. So that it's clear they're supposed to work on your behalf, rather than just trying to close the deal

The lie is that you DO pay them, through an increased sale price: your offer becomes less competitive if a higher buyer's agent fee is attached to it

They can flip-flop on any given issue, and it's of no consequence

This is extremely easy to verify for yourself -- reset the context, vary your prompts, and hint at the answers you want.

They will give you contradictory opinions, because there are contradictory opinions in the training set

---

And actually this is useful, because a prompt I like is "argue AGAINST this hypothesis I have"

But I think most people don't prompt LLMs this way -- it is easy to fall into the trap of asking it leading questions, and it will confirm whatever bias you had