Hacker News: gmfawcett

New comment by gmfawcett in "Magnolias are so old that they're pollinated by beetles, not bees (2025)"

gmfawcett — Mon, 27 Jul 2026 18:30:27 +0000

> Restricting the word "insects" to hexapods is a modern convention that does not have any basis in etymology.

This is one of those very rare cases in which "entomology" or "etymology" might have been chosen without the intended meaning being clear. :)

New comment by gmfawcett in "Successful companies go blind"

gmfawcett — Mon, 20 Jul 2026 18:51:39 +0000

I wasn't criticizing the essay. I was criticizing your suggestion that a narrow concept like "competence blindness" can explain the behaviour of a system as broad and immensely complicated as "all of Western Culture."

New comment by gmfawcett in "Successful Companies Go Blind"

gmfawcett — Fri, 10 Jul 2026 16:31:57 +0000

Or perhaps just because it's not accurate. Reducing any complex system down to a trivial projection is very unlikely to explain anything well.

New comment by gmfawcett in "Steve Wozniak cheered after telling students they have AI – actual intelligence"

gmfawcett — Sat, 23 May 2026 11:45:12 +0000

Look up "perlis epigrams lisp" for the context here. :)

New comment by gmfawcett in "Package managers keep using Git as a database, it never works out"

gmfawcett — Fri, 26 Dec 2025 15:48:15 +0000

Ostrom's results didn't disprove ToC. She showed that common resources can be communally maintained, not that tragic outcomes could never happen.

New comment by gmfawcett in "Hacker News front page now, but the titles are honest"

gmfawcett — Fri, 19 Dec 2025 17:50:52 +0000

"Going Postal" was brilliant. GNU Terry Pratchett.

New comment by gmfawcett in "50 years of proof assistants"

gmfawcett — Sun, 14 Dec 2025 18:51:46 +0000

That's a fair question, but I didn't have another person read it. (If I'm being honest, I think my colleagues would have looked at me funny. We don't exactly have a culture of using TLA+ to describe sysadmin processes here, and it felt a bit like a sledgehammer-swatting-fly even by my own standards! They thought the visualizations were cool, though...)

Many of my rules seemed pretty basic to me -- a lot of things like, "there is no target-node candidate whose ordinal is higher than the current target's and also satisfies the Good_Match predicate." But if I had been writing it for a human reader, rather than just to document constraints, I would have put in more effort to explain why the constraints existed in the first place (what's a Good match? Are there Poor matches? Why do we care? etc.). I didn't skip this step altogether but I didn't invest much time into it.

I did take care to separate "physics" rules from "strategy" rules (i.e, explicitly separating core actions and limits from objectives and approaches). That seemed to help the AI, and I'm sure it would have helped people too.

New comment by gmfawcett in "Purdue University approves new AI requirement for all undergrads"

gmfawcett — Sun, 14 Dec 2025 03:01:46 +0000

I think this was more of a press release than an edict. The Purdue announcement says, "Built on recently launched AI majors, minors and certificates across colleges, and following the establishment of a working group last summer, with additional careful deliberation and advice from the University Senate through its Undergraduate Curriculum Council..."

New comment by gmfawcett in "Purdue University approves new AI requirement for all undergrads"

gmfawcett — Sat, 13 Dec 2025 22:31:48 +0000

Delegated to the provost and deans. Who else would you expect to hold accountable for developing a graduate attribute?

New comment by gmfawcett in "50 years of proof assistants"

gmfawcett — Sat, 13 Dec 2025 21:03:40 +0000

I've had the same intuition. I've had mixed results in this area, although I'm certainly no expert. Recently I wanted to formalize a model of a runbook for a tricky system migration, to help me reason through some alternatives. I ended up writing a TLA+ spec before generating some visualizations, and also some possible prototypes in MiniZinc. All three (spec, visualizations, CP models) were vibe-coded in different sessions, in about that order, though most of my personal effort went into the spec.

While the later AIs quickly understood many aspects of the spec, they struggled with certain constraints whose intuitive meaning was concealed behind too much math. Matters which I had assumed were completely settled, because a precise constraint existed in the spec, had to be re-explained to the AI after implementation errors were found. Eventually, I added more spec comments to explain the motivation for some of the constraints, which helped somewhat. (While it's an untested idea, my next step was going to be to capture traces of the TLA+ spec being tested against some toy models, and including those traces as inputs when producing the implementations, e.g. to construct unit tests. Reasoning about traces seemed to be a fairly strong suit for the AI helper.)

In hindsight, I feel I set my sights a little too high. A human reader would have had similar comprehension problems with my spec, and they probably would have taken longer to prime themselves than the AI did. Perhaps my takeaway is that TLA+ is a great way to model certain systems mathematically, because precision in meaning is a great quality; but you still have to show sympathy to your reader.

New comment by gmfawcett in "Pwning the Nix ecosystem"

gmfawcett — Wed, 15 Oct 2025 22:06:21 +0000

That's pretty impressive -- thanks for sharing the link.

New comment by gmfawcett in "How I influence tech company politics as a staff software engineer"

gmfawcett — Sat, 04 Oct 2025 18:17:41 +0000

Very well expressed. That's great early-career guidance, but also a good refresher for many senior staff.

New comment by gmfawcett in "How to Lead in a Room Full of Experts"

gmfawcett — Wed, 24 Sep 2025 16:50:54 +0000

I've often taken inspiration from RFC 2418, "IETF Working Group Guidelines and Procedures" [1], a rare RFC that defines a human protocol ("rough consensus") rather than a technical one.

[1] https://www.rfc-editor.org/rfc/rfc2418.html

New comment by gmfawcett in "Pony: An actor-model, capabilities-secure, high-performance programming language"

gmfawcett — Tue, 29 Jul 2025 16:50:49 +0000

(also, velociraptors)

New comment by gmfawcett in "Psilocybin shows promise as anti-aging therapy"

gmfawcett — Sat, 12 Jul 2025 13:12:51 +0000

Arguably the European churches were such an industry for centuries, and were highly effective at it.

New comment by gmfawcett in "Nearly all binary searches and mergesorts are broken (2006)"

gmfawcett — Sat, 11 Jan 2025 15:53:45 +0000

It gives a cue about how many times I've probably seen the article before. Quite useful, IMO. I read this particular article when it came out in 2006... it's convenient to know we're not discussing a novel finding on the same topic.

New comment by gmfawcett in "FTC takes action against Gravy Analytics, Venntel for selling location data"

gmfawcett — Tue, 03 Dec 2024 22:34:42 +0000

> we are free only as much as we don't have guns in our face telling us we're not.

Is this actually an old saying?

New comment by gmfawcett in "Adventures in Probability"

gmfawcett — Mon, 11 Nov 2024 21:27:23 +0000

Nice article!

New comment by gmfawcett in "Asking the wrong questions (2017)"

gmfawcett — Sun, 08 Sep 2024 11:51:45 +0000

> What very few SF writers understood

"Understood?" They were writing fiction, not instruction manuals.

New comment by gmfawcett in "Markdown is meant to be shown (2021)"

gmfawcett — Sat, 17 Aug 2024 01:10:36 +0000

He can say what he wants, but he literally inlined the entirety of HTML into the syntax. It's delusional to suggest that this design choice emphasizes readability above all else.