"Time the First Call: Julia 1.13 and the Compiler’s Oldest Debt"

General Usage Performance
1

At Time the First Call: Julia 1.13 and the Compiler’s Oldest Debt

I thought it was a clear and informative article about the tradeoffs in Julia runtime, precompilation, and TTFx, and their histories.

EDIT: Yuck. This is LLM generated, confirmed by Pangram. I thought the style was weird, but the information was valuable. I should have paid attention to my misgivings and realized I was reading slop. I apologize for posting this here (tried to delete it but I’m not allowed to).

2

“by Bezansonian Julia Speed, Simulacrum” what a name lol

3

Yeah, I don’t know what’s going on there. It was almost enough to keep me from reading the article, but I’m glad I forged ahead.

4

I’m guessing the AI “author’s” name is deliberately strange to curb accidental identity theft.

5

wouldn’t you rather read the release notes directly (written by people — us) rather than a regurgitation re-written by claude …?

6

Although the LLM text generation shines through the article more than I’d want, I overall liked to read the article, because the question how related things changed across multiple releases is typically not answered in release notes (which are very worth reading, but for other reasons). These things take quite some effort to write a blog post about and it’s work to get the quotes and sources right.

So while I still had preferred to read the same article created/reworked by a good human writer, as long as no one volunteers their time to write it, I’ll rather take an article like linked above as the second best option instead of not having the article at all.

7

I broadly agree with the “AI summary better than no summary” sentiment, but I take issue with the portrayal of the discourse and consequences. Refreshingly, the LLM does correctly identify many sources; perhaps LeePhillips ran across this in the first place because his own blog was heavily cited. That actually proves people have been volunteering their time to write summaries or engage in conversations about the TTFX problem in their own time. It’s very possible that a Discourse topic or another blogpost would have or has already shown up within the next couple months, but the LLM website “decided” to summarize the first point in the recent 1.13 release notes and related links regardless. At the risk of rehashing several contentious conversations about AI usage, I do think there are many caveats to AI summaries, some that warrant some reservations about how we encourage it:

  • Collecting many sources across time and summarizing them a la review article may not be a good way to catch up on a topic, even a longstanding one. Sometimes, one or two particularly strong sources is enough, and diving deeper into them yields more useful information. Putting aside the otherwise decent summary for example, the scattered quoted statistics were probably less valuable than a deliberate consideration of the data in the Julia Ecosystem Benchmarks. LLMs don’t often do this and may even resist the suggestion because it is less generic and expensive.
  • More broadly, LLMs make mistakes in a very different way than humans, and fixing them in conversation is a lot less efficient. In many cases it’s better to let some agents figure something out in a much quicker “conversation” before getting personally involved.
  • In this particular case, we are rewarding an automated blog that evidently talks about anything it comes across and likely never again. Whether made by human or machine, this kind of outsider content has scaled to the point that people look to and reward the central community far less by comparison. Even before LLMs hit the mainstream, many online businesses completely rely on collecting or summarizing other creators’ content. At best, those are bigger creators who do not mind the extra traffic or are private individuals who are happy for any attention. At worst, the smaller creators had ambitions of their own and lost opportunities to fair use and an algorithm. I can’t speak for anyone, but LeePhillips evidently does not appreciate the citations in hindsight.
  • People might be less inclined to write summaries when demand shifts to faster AI summaries, ironically discouraging the deeper dives an LLM relies on for summaries. This isn’t a hypothetical; tutorials, documentation, and blogposts have broadly gotten a lot staler in the past few years, which may be fine for very established information but not so much for newer developments. It’s gotten to the point I no longer take any comparisons of fairly new libraries at face value.

To be clear, I’m not saying “AI summaries considered harmful”; it’d be extremely hypocritical. However, it’s really not as simple as “better than nothing.”

8

Informative read, although I sometimes read the blog posts for new releases, so I knew some stuff already

9

Some more background can be found in the host institution which seems to be running this substack:

But it’s not clear if there is more than one person behind it (About page: The Vision — Universitas Scholarium).

10

In some sense, the project website it very creative, as it mixes:

  • geometric, square flat colors, design which is 100% inspired from the abstract paintings of 1920s-1930s Piet Mondrian, along with bold sans serif heading
  • pictures of engraving mixed with Italic serif fonts and lots of Latin words, which sounds to me as a reminiscence of “something ancient”, anywhere in time between Ancient Greece and the Renaissance in Italy…
11

We can talk with the real @jeff.bezanson here (or he sometimes answers). I’m amazed by this project/university (though have not tried it yet). You can talk to “him” here (and seemingly mostly a lot of dead people, such as the greats, Albert Einstein Simulacrum — General Relativity (20th century) | Universitas Scholarium, McCarthyian Symbolic Intelligence Simulacrum — Coined “artificial intelligence” but also e.g. Carmackian Coder Simulacrum — Doom (20th–21st century) | Universitas Scholarium):

AI projects and YouTube channels that simulate and mislead people are unethical. I would really like to know what the real Bezanson thinks, or if he gave permission…

12

Man, that was awful.

Putting aside the general negative points regarding AI, the article is just badly structured and tries to create a coherent story from several disconnected subjects. Like, the paragraph about --trim misses the key details (they reached 1MB “Hello world” binary size but did not mention what the starting point was, leaving an impression that trimming was not yet applicable in a real-world scenario), and, most importantly, trimming has nothing to do with building system images. And half of the points in the final paragraph are not related to TTFX at all.

Also, regarding “AI summary is better than no summary”: no. Just no. It has zero value, since anyone can approach their AI agent of choice, ask a question and get the same summary. It’s like attaching three first Google results and saying “well, that’s better than no sources at all”.

Now, to not let myself be biased, I generally think LLMs are an incredible tool if you know how to use it (as any other tool!), although it does make programming much less fun. That said, this “academy” website is a huge pile of slop.

Better watch Jeff’s talks.